UNIVERSIDADE DA BEIRA INTERIOR
Engenharia
Multi-sensor data fusion in mobile devices for the
identification of Activities of Daily Living
Ivan Miguel Serrano Pires
Tese para obtenção do Grau de Doutor em
Engenharia Informática
(3º ciclo de estudos)
Orientador: Prof. Dr. Nuno M. Garcia (Universidade da Beira Interior, Portugal)
Co-orientador: Prof. Dr. Francisco Flórez-Revuelta (Universidad de Alicante, Spain)
Covilhã, Novembro de 2018
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
ii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Thesis prepared at the Ambient Assisted Living Computing and Telecommunications Laboratory
(ALLab), Instituto de Telecomunicações, Universidade da Beira Interior, and submitted to the
Universidade da Beira Interior for defence in a public examination session.
Research partially financed by the Portuguese Fundação para a Ciência e a Tecnologia through
the project identified by UID/EEA/50008/2013 and partially supported by the COST Action
IC1303 – AAPELE – Architectures, Algorithms and Protocols for Enhanced Living Environments.
Work was also partially supported by Altran Portugal at different stages of the research.
iii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
iv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Dedicatory
“Do the difficult things while they are easy and do the great things while they are small.
A journey of a thousand miles must begin with a single step.”
Lao Tzu
To my parents, José e Helena, who often gave themselves and gave up their dreams, so that
I could accomplish mine. I want to say that this achievement is not only mine, but ours. All I
got was only possible thanks to the love, support and dedication you have always had for me.
I have always been taught to act with respect, simplicity, dignity, honesty and love for
others. With the union of all, the obstacles were overcome, the victories were achieved and
the happiness was shared.
Thank you for your patience and understanding during this long journey.
To the rest of my family, supervisors, professors and friends for their continuous
friendship, affection, companionship, unconditional support and encouragement.
Thanks you for cheering my achievements along this journey.
“The ultimate measure of a man is not where he stands in moments of comfort and
convenience, but where he stands at times of challenges and controversy”
Martin Luther King
v
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
vi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Acknowledgements
This thesis would not have been possible without the help of many people.
Firstly, I would like to thank to my parents for having been by my side in the difficult moments
and helped me with their knowledge, supporting me in all moments and encouraging me to
finish this PhD Thesis. I also would like to thank them for their unconditional love, because I
came from a humble family and I have this achievement. During all life, my parents had several
dreams that cannot realize, but one of their dreams was now achieved with my success.
Secondly, I would also like to thank my family for all support in the good and bad moments
during the last four years. They are very important to this achievement.
Friends are the support of our life, I would also like to thank them, specially those who have
always believed in my success. In our life, sometimes we find people that can help us with their
good viewpoints and friendship. However, sometimes we find false friends, that are the rocks
in our travel, but they always teach something to us.
I also want to thank my supervisor, Prof. Dr. Nuno M. Garcia, which is the person responsible
for my success. He had a lot of patience to give me the help, guidance, trust and motivation
needed, and to teach some important things for the future, but the adventure always continue
and we will together in a lot of things in the future. When I started the PhD research, I didn’t
know more than Portugal, and my English was not good. He found the best solutions and
encouraged me to go to different countries in order to do some work with people with different
cultures.
Related to this PhD research, I also need to thank my co-supervisor, Prof. Dr. Francisco FlórezRevuelta, for the availability and help in my travel to Kingston University, Kingston upon
Thames, United Kingdom, and for their support in the different activities of my PhD research.
During this PhD research, as I did some parts of the work at Universidad de Alicante, Alicante,
Spain, I need thank to all people for their support and availability in my visit, where it was very
useful for the success of this PhD research.
Other people are important to support my knowledge in different cultures, including Prof. Dr.
Susanna Spinsante, who helped me in my visit to the Marche Polytechnic University, Ancona,
Italy, with the support for my PhD research, new ideas for future research in cooperation with
her, and other support in the scientific publications of the last years.
vii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
During my Bachelor studies in Computer Science and Engineering at the Polytechnic Institute
of Castelo Branco, I was a student of Prof. Dr. Cristina Canavarro, who was very important in
the success of the statistical analysis of the results obtained with the framework for the
recognition of Activities of Daily Living (ADL) and environments developed during this PhD
research.
I would also like to thank other professors, especially Nuno Pombo, Eftim Zdravevski, Rossitza
Goleva and Natalia Diaz Rodriguez, for their help and support in the last activities and
publications of my PhD research.
For the financial support of all activities, I also would like to thank the Instituto de
Telecomunicações, that through the Portuguese Fundação para a Ciência e a Tecnologia through
the project identified by UID/EEA/50008/2013 supported several activities of this PhD research.
I would also like to acknowledge the contribution of the COST Action IC1303 – AAPELE –
Architectures, Algorithms and Protocols for Enhanced Living Environments, which helped to
know some people and supported financially some activities of my PhD research.
Other important thing during my PhD research was my work for Altran Portugal, that especially
with some people I have found the important trust for this achievement, including Pedro
Furtado, João Minhota, André Freitas and remaining colleagues and managers.
The Ambient Assisted Living Computing and Telecommunications Laboratory (ALLab), at
Universidade da Beira Interior, Covilhã, Portugal, was very important for the development and
success of this PhD research with the support of several colleagues, including Virginie Felizardo,
Henriques Zacarias, Rui Santos, Dmytro Vasyanovych, among others. Thank you for all
friendship and help during my PhD research.
In my travels to Kingston University, Kingston upon Thames, United Kingdom, and Marche
Polytechnic University, Ancona, Italy, I also need to thank the people who give me good
experiences and knowledge in the research about Ambient Assisted Living (AAL) systems.
One of the difficulties of my PhD research were focused in the finding of people that agreed to
perform the experiments needed for the acquisition of data from the sensors available in a
commodity smartphone. Thus, I would also to thank the people that agreed to help with their
efforts and availability. In addition, I would also like thank to the people that works in
physiotherapy that helped to know the physical signals acquired from the different sensors
available in the mobile devices.
During my PhD research, I had some health problems and I would also thank the healthcare
professionals of Centro Hospitalar Universitário de Coimbra, Coimbra, Portugal, who did all
efforts to improve my health status.
viii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
During our path, at the end but not the less important, we are always looking for opportunities,
but the best things magically appear out of nowhere, as new projects, new findings and people
that change our manner to view the life, including Tânia Valente, Diogo Marques, António
Santos, and other people that we can cooperate in the future, and I need to thank them for all
availability, encouragement and support at the different stages of my PhD research.
ix
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
x
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Foreword
This thesis describes the research work performed in the scope of the 4-year doctoral research
programme and presents its conclusions and contributions. The research activities were mainly
carried out at the Ambient Assisted Living Computing and Telecommunications Laboratory
(ALLab), Universidade da Beira Interior and also Instituto de Telecomunicações, Covilhã,
Portugal. Additionally, some research activities during the 4-year doctoral research programme
were performed in other universities in different European countries, including Kingston
University, Kingston upon Thames, United Kingdom, Universidad de Alicante, Alicante, Spain,
and Marche Polytechnic University, Ancona, Italy.
The research work was supervised by Dr. Nuno M. Garcia, from the Universidade da Beira
Interior, and co-supervised by Dr. Francisco Flórez-Revuelta, from Universidad de Alicante,
Spain. Dr. Francisco Flórez-Revuelta was at Kingston University at the time of the research
activities and a supervised the research work carried out there. During the time of the research
activities in the Marche Polytechnic University, the author was under supervision of Dr. Susanna
Spinsante. The work performed out of Portugal was mainly performed at the Universidad de
Alicante with supervision of Dr. Francisco Flórez-Revuelta, which contributed to the success of
this PhD research.
This work was financially supported in part by the Fundação para a Ciência e a Tecnologia,
Portugal, through the project identified by UID/EEA/50008/2013 and some activities were
supported by Short Term Scientific Meeting funds, granted by the COST Action IC1303 – AAPELE
– Architectures, Algorithms and Protocols for Enhanced Living Environments, funded by The
COST Association, through the European Science Foundation and the H2020 program.
The research resulted in the definition and development of a framework for the recognition of
Activities of Daily Living (ADL) and its environments based on the data acquired from the sensors
available in off-the-shelf mobile devices. The architecture of the framework relies on several
concepts, including data acquisition, data processing, data cleaning, data imputation, feature
extraction, data fusion and classification with machine learning methods. At Kingston
University, the review of the state of the art was carried out, resulting in a review article. The
research about the state of the art was completed with other three articles mainly related to
the validation techniques and classification of the data acquired from sensors, and a review
about audio fingerprinting techniques. These literature reviews were fundamental in the
definition of the architecture of the framework for the recognition of ADL and environments.
Therefore, during the development of the framework, it was verified that the step related to
the data imputation may be done in the future in order to increase the reliability of the
xi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
framework, but this research is out of the scope of this PhD thesis. The main development of
the framework is related to the data fusion and classification of the data acquired from offthe-shelf mobile devices in free living activities.
The research work developed during the doctoral programme and described in this Thesis is the
consequence of the activities performed in the different institutions visited, which allowed the
author to benefit from the different cultures and experiences of the researchers at these
different academies.
xii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
List of Publications
List of articles included in the thesis resulting from this
4-year doctoral research programme
1. From Data Acquisition to Data Fusion a Comprehensive Review and a Roadmap for
the Identification of Activities of Daily Living using Mobile Devices
Ivan Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, published in
Sensors, MDPI, Sensors 2016, 16(2), 184; doi:10.3390/s16020184, January 2016 (IF 2017:
2.677, Q1 Electrical and Electronic Engineering); 38 citations.
2. Validation Techniques for Sensor Data in Mobile Health Applications
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta and Natalia
Rodríguez, published in the special issue Integration of Sensors in Control and
Automation Systems, Journal of Sensors, Hindawi publishers, September 2016 (IF: 2017:
2.057, Q2 Engineering, Electrical & Electronic); 3 citations.
3. Recognition of Activities of Daily Living Based on Environmental Analyses Using
Audio Fingerprinting Techniques: A Systematic Review
Ivan Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Susanna Spinsante,
Rossitza Goleva, Eftim Zdravevski, published in Sensors, MDPI, Sensors 2018, 18(1), 160;
doi: 10.3390/s18010160, January 2018 (IF 2017: 2.677, Q1 Electrical and Electronic
Engineering); 0 citations.
4. Approach for the Development of a Framework for the Identification of Activities
of Daily Living Using Sensors in Mobile Devices
Ivan Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta and Susanna
Spinsante,
published
in
Sensors,
MDPI,
Sensors
2018,
18(2),
640;
doi:
10.3390/s18020640, February 2018 (IF 2017: 2.677, Q1 Electrical and Electronic
Engineering); 1 citations.
5. Identification of Activities of Daily Living through Data Fusion on Motion and
Magnetic Sensors embedded on Mobile Devices
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Susanna
Spinsante and Maria Canavarro Teixeira, published in Pervasive and Mobile Computing,
Elsevier, 47, pp. 78-93; doi: 10.1016/j.pmcj.2018.05.005, July 2018 (IF 2017: 2.349,
Q1 Computer Science (miscellaneous)); 0 citations.
xiii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
6. Android library for recognition of activities of daily living: implementation
considerations, challenges, and solutions
Ivan Miguel Pires, Maria Canavarro Teixeira, Nuno Pombo, Nuno M. Garcia, Francisco
Flórez-Revuelta, Susanna Spinsante, Rossitza Goleva and Eftim Zdravevski, published
in Open Bioinformatics Journal, Bentham Science Publishers B.V., 11, pp. 61-88; doi:
10.2174/1875036201811010061, May 2018 (CiteScore 2016: 4.86, D1 Computer Science
(miscellaneous)); 0 citations.
Other publications resulting from the doctoral research
programme not included in the thesis
1. Multi-sensor data fusion techniques for the identification of activities of daily living
using mobile devices
Ivan Miguel Pires, Nuno M. Garcia and Francisco Flórez-Revuelta, published in
ECMLPKDD 2015 Doctoral Consortium, European Conference on Machine Learning and
Principles and Practice of Knowledge Discovery in Databases, Porto, Portugal, 7-11
September 2015; 2 citations.
2. Identification of Activities of Daily Living using Sensors Available in Off-the-shelf
Mobile Devices: Research and Hypothesis
Ivan Pires, Nuno M. Garcia, Nuno Pombo and Francisco Flórez-Revuelta, published in
7th International Conference on Ambient Intelligence, Seville, Spain, 1-3 June 2016; 2
citations.
3. Limitations of energy expenditure calculation based on a mobile phone
accelerometer
Ivan Pires, Virginie Felizardo, Nuno Pombo, Nuno M. Garcia, published in the
International Conference on High Performance Computing & Simulation (HPCS 2017),
Genova, Italy, 17-21 July 2017; 0 citations.
4. Limitations of the Use of Mobile Devices and Smart Environments for the Monitoring
of Ageing People
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo and Francisco Flórez-Revuelta,
published in ICT4AWE 2018 4th International Conference on Information and
Communication Technologies for Ageing Well and e-Health, Madeira, Portugal, 22-23
March 2018; 0 citations.
xiv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
5. User Environment Detection with Acoustic Sensors Embedded on Mobile Devices for
the Recognition of Activities of Daily Living
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Susanna
Spinsante, Maria Canavarro Teixeira and Eftim Zdravevski, Accepted for publication in
Statistic, Optimization & Information Computing, iapress, 2018; 0 citations.
6. A Multiple Source Framework for the Identification of Activities of Daily Living Based
on Mobile Device Data
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Susanna
Spinsante, Maria Canavarro Teixeira and Eftim Zdravevski, submitted for publication in
an international journal, March 2018.
7. Measurement of the Reaction Time in the 30-S Chair Stand Test using the
Accelerometer Sensor Available in off-the-Shelf Mobile Devices
Ivan Miguel Pires, Diogo Marques, Nuno Pombo, Nuno M. Garcia, Mário C. Marques and
Francisco Flórez-Revuelta, published in ICT4AWE 2018 4th International Conference on
Information and Communication Technologies for Ageing Well and e-Health, Madeira,
Portugal, 22-23 March 2018; 0 citations.
8. A review on the artificial intelligence algorithms for the recognition of Activities of
Daily Living using sensors in mobile devices
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Eftim
Zdravevski and Susanna Spinsante, submitted for publication in an international
journal, March 2018.
9. Pattern Recognition Techniques for the Identification of Activities of Daily Living
using Mobile Device Accelerometer
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Susanna
Spinsante, Maria Canavarro Teixeira and Eftim Zdravevski, submitted for publication in
an international journal, April 2018.
10. Mobile Platform for the Recognition of Activities of Daily Living and their
Environments based on Artificial Neural Networks
Ivan Miguel Pires, Nuno Pombo, Nuno M. Garcia and Francisco Flórez-Revuelta,
accepted for presentation in IJCAI-ECAI 2018, the 27th International Joint Conference
on Artificial Intelligence and the 23rd European Conference on Artificial Intelligence,
Stockholm, Sweden, 9-19 July 2018; 0 citations.
xv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
11. Framework for the Recognition of Activities of Daily Living and their Environments
in the Development of a Personal Digital Life Coach
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo and Francisco Flórez-Revuelta,
accepted for presentation in DATA 2018, the 7th International Conference on Data
Science, Technologies and Applications, Porto, Portugal, 26-28 July 2018; 0 citations.
xvi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Resumo
Após os recentes avanços tecnológicos e o crescente uso dos dispositivos móveis, como por
exemplo os smartphones, várias soluções podem ser desenvolvidas para melhorar a qualidade
de vida dos utilizadores no contexto de Ambientes de Vida Assistida (AVA) ou Ambient Assisted
Living (AAL). Os dispositivos móveis integram vários sensores, tais como acelerómetro,
giroscópio, magnetómetro, microfone e recetor de Sistema de Posicionamento Global (GPS),
que permitem a aquisição de vários parâmetros físicos e fisiológicos para o reconhecimento de
diferentes Atividades da Vida Diária (AVD) e os seus ambientes. A definição de AVD inclui um
conjunto bem conhecido de tarefas que são tarefas básicas de autocuidado, baseadas nos tipos
de habilidades que as pessoas geralmente aprendem na infância. Essas tarefas incluem
alimentar-se, tomar banho, vestir-se, fazer os cuidados pessoais, caminhar, correr, pular, subir
escadas, dormir, ver televisão, trabalhar, ouvir música, cozinhar, comer, entre outras. No
contexto de AVA, alguns indivíduos (comumente chamados de utilizadores) precisam de
assistência particular, seja porque o utilizador tem algum tipo de deficiência, seja porque é
idoso, ou simplesmente porque o utilizador precisa/quer monitorizar e treinar o seu estilo de
vida. A investigação e desenvolvimento de sistemas que fornecem algum tipo de assistência
particular está em crescente em muitas áreas de aplicação. Em particular, no futuro, o
reconhecimento das AVD é uma parte importante para o desenvolvimento de um assistente
pessoal digital, fornecendo uma assistência pessoal de baixo custo aos diferentes tipos de
pessoas. pessoas. Para ajudar no reconhecimento das AVD, os ambientes em que estas se
desenrolam devem ser reconhecidos para aumentar a fiabilidade destes sistemas.
O foco principal desta Tese é o desenvolvimento de métodos para a fusão e classificação dos
dados adquiridos a partir dos sensores disponíveis nos dispositivos móveis, para o
reconhecimento quase em tempo real das AVD, tendo em consideração a grande diversidade
das características dos dispositivos móveis disponíveis no mercado. Para atingir este objetivo,
esta Tese iniciou-se com a revisão dos métodos e tecnologias existentes para definir a
arquitetura e os módulos do novo método de identificação das AVD. Com esta revisão da
literatura e com base no conhecimento adquirido sobre os sensores disponíveis nos dispositivos
móveis disponíveis no mercado, um conjunto de tarefas que podem ser identificadas foi
definido para as pesquisas e desenvolvimentos desta Tese. Esta revisão também identifica os
principais conceitos para o desenvolvimento do novo método de identificação das AVD,
utilizando os sensores, são eles: aquisição de dados, processamento de dados, correção de
dados, imputação de dados, extração de características, fusão de dados e extração de
resultados recorrendo a métodos de inteligência artificial. Um dos desafios está relacionado
aos diferentes tipos de dados adquiridos pelos diferentes sensores, mas outros desafios foram
encontrados, sendo os mais relevantes o ruído ambiental, o posicionamento do dispositivo
xvii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
durante a realização das atividades diárias, as capacidades limitadas dos dispositivos móveis.
As diferentes características das pessoas podem igualmente influenciar a criação dos métodos,
escolhendo pessoas com diferentes estilos de vida e características físicas para a aquisição e
identificação dos dados adquiridos a partir de sensores. Com base nos dados adquiridos,
realizou-se o processamento dos dados, implementando-se métodos de correção dos dados e a
extração de características, para iniciar a criação do novo método para o reconhecimento das
AVD. Os métodos de imputação de dados foram excluídos da implementação, pois não iriam
influenciar os resultados da identificação das AVD e dos ambientes, na medida em que são
utilizadas as características extraídas de um conjunto de dados adquiridos durante um intervalo
de tempo definido.
A seleção dos sensores utilizáveis, bem como das AVD identificáveis, permitirá o
desenvolvimento de um método que, considerando o uso de tecnologias para a fusão de dados
adquiridos com múltiplos sensores em coordenação com outras informações relativas ao
contexto do utilizador, tais como a agenda do utilizador, permitindo estabelecer um perfil de
tarefas que o utilizador realiza diariamente. Com base nos resultados obtidos na literatura, o
método escolhido para o reconhecimento das AVD são as diferentes variantes das Redes
Neuronais Artificiais (RNA), incluindo Multilayer Perceptron (MLP), Feedforward Neural
Networks (FNN) with Backpropagation and Deep Neural Networks (DNN). No final, após a
criação dos métodos para cada fase do método para o reconhecimento das AVD e ambientes, a
implementação sequencial dos diferentes métodos foi realizada num dispositivo móvel para
testes adicionais.
Após a definição da estrutura do método para o reconhecimento de AVD e ambientes usando
dispositivos móveis, verificou-se que a aquisição de dados pode ser realizada com os métodos
comuns. Após a aquisição de dados, os mesmos devem ser processados no módulo de
processamento de dados, que inclui os métodos de correção de dados e de extração de
características. O método de correção de dados utilizado para sensores de movimento e
magnéticos é o filtro passa-baixo de modo a reduzir o ruído, mas para os dados acústicos, a
Transformada Rápida de Fourier (FFT) foi aplicada para extrair as diferentes frequências.
Após a correção dos dados, as diferentes características foram extraídas com base nos tipos de
sensores usados, sendo a média, desvio padrão, variância, valor máximo, valor mínimo e
mediana de dados adquiridos pelos sensores magnéticos e de movimento, a média, desvio
padrão, variância e mediana dos picos máximos calculados com base nos dados adquiridos pelos
sensores magnéticos e de movimento, as cinco maiores distâncias entre os picos máximos
calculados com os dados adquiridos dos sensores de movimento e magnéticos, a média, desvio
padrão, variância e 26 Mel-Frequency Cepstral Coefficients (MFCC) das frequências obtidas
com FFT com base nos dados obtidos a partir do microfone, e a distância calculada com os
dados adquiridos pelo recetor de GPS.
xviii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Após a extração das características, as mesmas são agrupadas em diferentes conjuntos de dados
para a aplicação dos métodos de RNA de modo a descobrir o método e o conjunto de
características que reporta melhores resultados. O módulo de classificação de dados foi
incrementalmente desenvolvido, começando com a identificação das AVD comuns com sensores
magnéticos e de movimento, i.e., andar, correr, subir escadas, descer escadas e parado. Em
seguida, os ambientes são identificados com dados de sensores acústicos, i.e., quarto, bar, sala
de aula, ginásio, cozinha, sala de estar, hall, rua e biblioteca. Com base nos ambientes
reconhecidos e os restantes sensores disponíveis nos dispositivos móveis, os dados adquiridos
dos sensores magnéticos e de movimento foram combinados com o ambiente reconhecido para
diferenciar algumas atividades sem movimento (i.e., dormir e ver televisão), onde o número
de atividades reconhecidas nesta fase aumenta com a fusão da distância percorrida, extraída
a partir dos dados do recetor GPS, permitindo também reconhecer a atividade de conduzir.
Após a implementação dos três métodos de classificação com diferentes números de iterações,
conjuntos de dados e configurações numa máquina com alta capacidade de processamento, os
resultados relatados provaram que o melhor método para o reconhecimento das atividades
comuns de AVD e atividades sem movimento é o método DNN, mas o melhor método para o
reconhecimento de ambientes é o método FNN with Backpropagation. Dependendo do número
de sensores utilizados, esta implementação reporta uma exatidão média entre 85,89% e 89,51%
para o reconhecimento das AVD comuns, igual a 86,50% para o reconhecimento de ambientes,
e igual a 100% para o reconhecimento de atividades sem movimento, reportando uma exatidão
global entre 85,89% e 92,00%.
A última etapa desta Tese foi a implementação do método nos dispositivos móveis, verificando
que o método FNN requer um alto poder de processamento para o reconhecimento de
ambientes e os resultados reportados com estes dispositivos são inferiores aos resultados
reportados com a máquina com alta capacidade de processamento utilizada no
desenvolvimento do método. Assim, o método DNN foi igualmente implementado para o
reconhecimento dos ambientes com os dispositivos móveis. Finalmente, os resultados relatados
com os dispositivos móveis reportam uma exatidão entre 86,39% e 89,15% para o
reconhecimento das AVD comuns, igual a 45,68% para o reconhecimento de ambientes, e igual
a 100% para o reconhecimento de atividades sem movimento, reportando uma exatidão geral
entre 58,02% e 89,15%.
Com base nos resultados relatados na literatura, os resultados do método desenvolvido mostram
uma melhoria residual, mas os resultados desta Tese identificam mais AVD que os demais
estudos disponíveis na literatura. A melhoria no reconhecimento das AVD com base na média
das exatidões é igual a 2,93%, mas o número máximo de AVD e ambientes reconhecidos pelos
estudos disponíveis na literatura é 13, enquanto o número de AVD e ambientes reconhecidos
com o método implementado é 16. Assim, o método desenvolvido tem uma melhoria de 2,93%
na exatidão do reconhecimento num maior número de AVD e ambientes.
xix
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Como trabalho futuro, os resultados reportados nesta Tese podem ser considerados um ponto
de partida para o desenvolvimento de um assistente digital pessoal, mas o número de ADL e
ambientes reconhecidos pelo método deve ser aumentado e as experiências devem ser
repetidas com diferentes tipos de dispositivos móveis (i.e., smartphones e smartwatches), e os
métodos de imputação e outros métodos de classificação de dados devem ser explorados de
modo a tentar aumentar a confiabilidade do método para o reconhecimento das AVD e
ambientes.
Palavras-Chave
Fusão de dados de sensores, aquisição de dados, acelerómetro, giroscópio, magnetómetro,
microfone, recetor de Sistema de Posicionamento Global (GPS), processamento de sinais,
sensores, atividades da vida diária, dispositivos móveis, reconhecimento de padrões,
processamento de dados, correção de dados, imputação de dados, extração de características,
inteligência artificial, redes neuronais artificiais.
xx
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Resumo Alargado
Introdução
Este capítulo resume, de forma alargada e em Língua Portuguesa, o trabalho de investigação
descrito na Tese de doutoramento intitulada de “Multi-sensor data fusion in mobile devices for
the identification of Activities of Daily Living”. Inicialmente, este capítulo descreve o
enquadramento da Tese, o problema abordado e os objetivos desta Tese de doutoramento,
bem como o enquadramento da mesma e as principais contribuições. Seguidamente, será
apresentado um resumo de cada um dos capítulos desta Tese, que correspondem às principais
contribuições desta Tese. O capítulo termina com a apresentação das principais conclusões
desta Tese, bem como a apresentação de algumas linhas de investigação para o futuro.
Enquadramento da Tese
Com base na literatura, o reconhecimento das Atividades de Vida Diária (AVD) concentra-se na
identificação de um conjunto bem conhecido de tarefas que são tarefas básicas de autocuidado,
baseadas nos tipos de habilidades que as pessoas geralmente aprendem na infância. Essas
tarefas incluem alimentar-se, tomar banho, vestir-se, fazer os cuidados pessoais, caminhar,
correr, pular, subir escadas, dormir, ver televisão, trabalhar, ouvir música, cozinhar, comer,
entre outras. No contexto dos Ambientes de Vida Assistida (AVA) [1, 2], alguns indivíduos
(comumente chamados de utilizadores) precisam de assistência particular, seja porque o
utilizador tem algum tipo de deficiência, seja porque é idoso, ou simplesmente porque o
utilizador precisa/quer monitorizar e treinar o seu estilo de vida. No entanto, as AVD incluem
um conjunto mais amplo de tarefas, não limitado a tarefas relacionadas com cuidados pessoais
e higiene, incluindo também as tarefas de socialização e profissionais. O desenvolvimento de
métodos para o reconhecimento das AVD é importante para apoiar a autonomia de pessoas
idosas e/ou com doenças crónicas e utilizadores que tenham algum problema de saúde [3, 4].
No entanto, pode ser útil para todos, incluindo atletas e jovens, já que essas soluções podem
ser integradas num método de monitorização e treino de estilos de vida [5].
xxi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Atualmente, o uso de sensores para o reconhecimento das AVD e ambientes é um tema muito
estudado [6-14]. No entanto, não existem estudos que definam um método estruturado para o
reconhecimento das AVD e ambientes, usando um grande número de sensores disponíveis em
dispositivos móveis e para o reconhecimento de um grande número de AVD. O foco desta Tese
inclui o desenho de um novo método através da seleção e desenvolvimento de métodos
adequados para cada módulo do método proposto para o reconhecimento de AVD e ambientes,
sendo estes: aquisição, fusão, processamento e filtragem dos dados, e extração de
características e classificação dos dados; com um especial destaque para os métodos
relacionados com a fusão e classificação dos dados, descobrindo-se quais são os aspetos mais
importantes a serem abordados para obter resultados com maior exatidão.
A definição de um conjunto de tarefas que podem ser corretamente identificadas inclui a
deteção e a análise de um grande número de tarefas, bem como a identificação das tarefas
que podem ser identificadas pelos sensores disponíveis nos dispositivos móveis. A seleção do
conjunto de sensores utilizáveis e de tarefas identificáveis permitem o desenvolvimento de um
método que, considerando tecnologias de fusão de dados de múltiplos sensores e tendo em
conta o contexto, em coordenação com outras informações do utilizador, como agenda e a hora
do dia, permite estabelecer um perfil das tarefas que o utilizador realiza diariamente.
Existem vários desafios envolvidos no desenvolvimento de um novo método para o
reconhecimento das AVD. Por exemplo, com a aquisição de dados, os desafios incluem o
posicionamento incorreto do dispositivo móvel, a taxa de amostragem de dados, a
indisponibilidade dos sensores necessários para reconhecer uma determinada AVD e as
diferentes condições ambientais [15]. Existem vários métodos que podem ajudar a minimizar
esses problemas, incluindo os referidos no artigo [16] apresentado no capítulo 2. Em relação ao
processamento de dados, os desafios estão ligados às diferentes arquiteturas e plataformas nas
quais a solução pode ser implementada, como, por exemplo, utilização de processamento
remoto ou localmente. Para fornecer um resultado aceitável em vários ambientes de aquisição
de dados, o método deve funcionar sem uma ligação de rede e ser adaptada aos baixos recursos
de processamento, memória e armazenamento de alguns dispositivos móveis, sendo que a
implementação dos métodos é objeto de estudo desta Tese [15, 17-19].
O âmbito desta Tese consiste no uso de métodos de fusão de dados e Redes Neurais Artificiais
(RNA) para o reconhecimento das AVD e ambientes com base nos dados adquiridos pelos
sensores disponíveis num dispositivo móvel, como um smartphone. As limitações dos
smartphones também foram exploradas, criando um método que adquire 5 segundos de dados
dos sensores a cada 5 minutos, permitindo o mapeamento dos estilos de vida e o
desenvolvimento de novas soluções para AVA. A pesquisa relacionada aos métodos de
classificação desta Tese está relacionada com a comparação de diferentes implementações de
métodos de RNA adaptados aos recursos restritos de um dispositivo móvel, utilizando técnicas
xxii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
de processamento local, de modo a fornecer o feedback ao utilizador relacionado com as suas
AVD e os seus ambientes quase em tempo real.
Descrição do Problema e Objetivos da Investigação
O problema abordado nesta Tese consiste em segmentar e identificar corretamente um
conjunto significativo de AVD que os utilizadores executam nas suas rotinas diárias ou semanais,
usando os sensores e as informações do utilizador disponíveis num smartphone. A identificação
de AVD é um desafio complexo, devido à existência de um grande conjunto de AVD e à
complexidade da identificação do método correto para identificar cada uma delas. Para detetar
e identificar corretamente uma AVD, existem vários fatores que influenciam na aquisição dos
dados e, consequentemente, afetam a identificação de vários tipos de atividades, que devem
ser considerados, como as restrições ambientais, os movimentos involuntários do utilizador, a
posição do dispositivo móvel, os movimentos do dispositivo móvel durante a aquisição de dados
e a frequência da aquisição de dados dos sensores.
O reconhecimento de AVD e ambientes é, até ao momento, um tópico importante para a criação
de sistemas para diversas finalidades, como assistentes pessoais, monitorização das condições
de saúde, monitorização do estilo de vida e outros. Atualmente, esse reconhecimento é
comumente realizado por sistemas complexos e caros, com alto poder de processamento e
capacidade de memória [1, 20-24]. No entanto, essa tarefa pode ser executada com recurso
aos dispositivos móveis utilizados diariamente utilizando técnicas de processamento local para
um rápido feedback.
Hoje em dia, e com base na literatura, existem vários estudos usando diferentes subconjuntos
dos sensores disponíveis nos dispositivos móveis. No entanto, os estudos analisados reportam a
inexistência de métodos que utilizem um grande número de sensores disponíveis em dispositivos
móveis para o reconhecimento das AVD e ambientes. Atualmente, a literatura existente reporta
que as Redes Neurais Artificiais (RNA) e as suas variantes são os métodos mais utilizados,
apresentando uma exatidão relevante para a identificação de algumas AVD, conforme
apresentado em [16].
O principal objetivo desta Tese é a criação de um novo método para o reconhecimento de AVD
e ambientes com a fusão e classificação dos dados adquiridos por um grande conjunto de
sensores disponíveis nos dispositivos móveis. O método desenvolvido tem em conta o diferente
número de sensores disponíveis nos diferentes dispositivos móveis e as características limitadas
dos mesmos, de modo a permitir a sua execução localmente [25]. O método desenvolvido inclui
xxiii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
várias etapas, tais como aquisição de dados, processamento de dados, filtragem de dados,
imputação de dados, extração de características e reconhecimento de padrões. As AVD
identificadas durante esta Tese são andar, correr, subir escadas, estar parado/ver televisão,
descer escadas, conduzir e dormir, e os ambientes que foram identificados durante esta Tese
foram quarto, bar, sala de aula, ginásio, cozinha, sala de estar, hall, rua e biblioteca.
Os seguintes objetivos intermédios foram definidos a fim de dividir e organizar o trabalho de
pesquisa necessário para atingir o objetivo principal desta Tese:
1. De modo a compreender as soluções atualmente existentes e possíveis métodos, esta
Tese começou com a revisão da literatura relacionada com os diferentes conceitos
incluídos nesta Tese. Além disso, devido aos tipos de dados utilizados neste trabalho,
esta revisão também inclui o estudo relativo a métodos de validação de dados e
métodos de impressão digital de áudio. Finalmente, foram descritos os sensores, as
AVD e ambientes que podem ser identificados com os dados adquiridos a partir dos
dispositivos móveis.
2. Como segundo objetivo intermédio, apresentou-se a definição da arquitetura proposta
para o método de reconhecimento de AVD e ambientes, incluindo os diferentes
conceitos disponíveis na literatura.
3. Como terceiro objetivo intermédio, foi realizado o desenvolvimento de uma aplicação
móvel para aquisição dos dados dos sensores disponíveis nos dispositivos móveis. Alguns
utilizadores foram selecionados para executar a AVD e etiquetá-las. Durante a aquisição
de dados, o dispositivo móvel foi colocado no bolso frontal das calças do utilizador.
Esta aplicação móvel executa o processo de aquisição de dados.
4. Após a aquisição dos dados, o quarto objetivo intermédio consiste na implementação
de métodos de processamento de dados, incluindo filtragem de dados e extração de
características. Os métodos de imputação de dados foram evitados, pois sua
implementação não tem influência nos resultados da identificação das AVD e dos
ambientes, pois as características são extraídas com um conjunto de dados adquiridos
durante um intervalo de tempo definido.
5. Após o processamento dos dados, o quinto objetivo intermédio está relacionado com a
implementação e comparação de três configurações diferentes de RNA, de modo a
descobrir a melhor configuração para o reconhecimento das AVD e ambientes. Além
disso, esta análise incluiu a fusão das diferentes combinações de características
extraídas dos dados dos sensores para a implementação de diferentes RNA.
6. O sexto objetivo intermédio consiste na implementação de métodos que abordam os
diferentes módulos do método proposto para o reconhecimento de AVD e ambientes,
implementando-a numa aplicação móvel para a realização de testes adicionais.
O reconhecimento das ADL e ambientes implementados como resultado desta Tese é
importante para projetar o desenvolvimento de um assistente pessoal digital [5]. O tema desta
xxiv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Tese é também importante para apoiar a autonomia de idosos, pessoas com doenças crónicas
e utilizadores que apresentem algum tipo de deficiência [3, 4]. No entanto, pode ser útil para
todos, incluindo atletas e jovens, já que o método proposto pode ser utilizado para
monitorização e treino de estilos de vida [5].
Argumento da Tese
Esta Tese propõe um novo método para o reconhecimento de AVD e ambientes, tendo em conta
a fusão dos dados adquiridos com os sensores disponíveis nos dispositivos móveis.
Especificamente, o argumento da Tese é:
A identificação automática das Atividades de Vida Diária (AVD) e ambientes é um desafio
complexo amplamente estudado na literatura, mas não há métodos estruturados para o
reconhecimento de AVD. Existem diferentes tipos de dados que podem ser adquiridos com
dispositivos móveis, de onde podem ser extraídas diversas características, sendo aplicadas
técnicas de fusão de dados a confiabilidade do reconhecimento de AVD e ambientes. A fusão e
classificação de dados são os passos mais importantes no desenvolvimento de um método para
o reconhecimento de AVD e ambientes. O posicionamento do dispositivo móvel e a frequência
da aquisição de dados também são fatores importantes que devem ser tidos em conta.
Para suportar o argumento desta Tese, foi realizada a revisão bibliográfica sobre os diferentes
conceitos envolvidos neste tema, incluindo aquisição, processamento, validação, filtragem,
fusão e classificação de dados, bem como extração de características e métodos de impressão
digital de áudio, propondo uma arquitetura para um método para o reconhecimento de AVD e
ambientes. Depois da revisão da literatura, a aquisição de dados foi realizada para as AVD e
ambientes selecionados. Após a aquisição de dados, a implementação dos diferentes métodos
para cada módulo do método proposto foi realizada, comparando três implementações
diferentes de RNA para a criação dos diferentes métodos, tendo em conta diferentes
subconjuntos de sensores. No final, o método foi implementado numa aplicação móvel para
testes adicionais.
xxv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Principais Contribuições
Esta secção descreve resumidamente as principais contribuições científicas resultantes do
trabalho de pesquisa apresentado nesta Tese.
A primeira contribuição está relacionada com a proposta de uma estrutura para o método de
reconhecimento de AVD, descrevendo também os métodos existentes na literatura referentes
a cada conceito relacionado com esta Tese, incluindo métodos de classificação, métodos de
aquisição, processamento, imputação e fusão de dados adquiridos com os sensores de
dispositivos móveis, bem como a aplicabilidade destes métodos no reconhecimento de AVD
[16].
A segunda contribuição desta Tese é a classificação e aplicabilidade dos métodos existentes
para a validação dos dados adquiridos pelos sensores, tendo em conta os diferentes tipos de
dados em falta e/ou inválidos [26].
A terceira contribuição desta Tese é a comparação da exatidão das características e métodos
utilizados para a definição de uma impressão digital de áudio e já reportados na literatura,
utilizados aqui para o reconhecimento dos ambientes com base nos diferentes tipos de sinais
acústicos [27].
A quarta contribuição desta Tese é a definição da arquitetura do método proposto para o
reconhecimento de AVD e seus ambientes, propondo possíveis métodos para cada módulo [28].
A quinta contribuição desta Tese consiste na apresentação e análise dos resultados da
investigação relacionada com o reconhecimento das atividades de andar, correr, subir e descer
escadas, e estar parado com base na utilização dos dados adquiridos pelo acelerómetro,
magnetómetro
e
giroscópio
disponíveis
num
dispositivo
móvel,
comparando
três
implementações diferentes de RNA com dados normalizados e não normalizados, de modo a
descobrir a melhor implementação para o reconhecimento destas AVD, concluindo que o melhor
método é o DNN com dados normalizados [29].
A sexta contribuição desta Tese consiste na construção, implementação e testes de uma
biblioteca para dispositivos com o sistema operativo Android com o método para o
reconhecimento de AVD e ambientes com base nos resultados obtidos em estudos anteriores,
apresentando uma validação dos resultados obtidos pelo método proposto [30].
xxvi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Estado da arte
Inicialmente, esta Tese iniciou com a pesquisa do estado da arte relativa aos diversos conceitos,
tais como aquisição, processamento, validação, imputação, filtragem, fusão e classificação de
dados, criação de impressões digitais de áudio e outros conceitos relacionados com dispositivos
móveis e sensores.
1. Da aquisição até à fusão de dados: Uma revisão de literatura
para a identificação de atividades de vida diária usando
dispositivos móveis
A identificação das AVD é possível com o uso dos sensores disponíveis nos dispositivos móveis
utilizados diariamente, tais como smartphones, smartwatches ou outros dispositivos, pois estes
incluem diversos sensores, e.g., acelerómetro, magnetómetro, giroscópio, microfone e recetor
de Sistema de Posicionamento Global (GPS) [31, 32]. Para tal, começou por se estudar um
determinado conjunto de conceitos, tais como sensores de dispositivos móveis, aquisição,
processamento, imputação e fusão de dados, de modo a, baseado na literatura, aferir quais são
os melhores métodos e sensores que se adequam à identificação das AVD.
Os sensores disponíveis nos dispositivos móveis são de diversos tipos, tendo sido proposta a
classificação que inclui sensores magnéticos, e.g., magnetómetro e compasso, sensores de
ambiente, e.g., temperatura ambiente, pressão atmosférica e humidade, sensores de
localização, e.g., recetor de GPS e serviços de localização por Wi-Fi, sensores de movimento,
e.g., acelerómetro, sensor de gravidade, giroscópio e sensor de orientação, sensores de
imagem, e.g., câmara fotográfica e/ou de vídeo, sensores acústicos, e.g., microfone, sensores
óticos, e.g., sensor infravermelhos e sensor de Radio Frequency Identifier (RFID), e sensores
de pressão, e.g., sensor de impressão digital [16].
Relativamente à aquisição de dados, na literatura tem vindo a ser criadas várias metodologias
diferentes, de entre as quais uma das mais utilizadas é a Acquisition Cost-Aware QUery
Adaptation (ACQUA) [33], embora no decorrer desta Tese se tivesse optado por utilizar os
métodos comuns para a aquisição de dados.
xxvii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Relativamente ao processamento de dados, as metodologias apresentadas na literatura
dividem-se em duas vertentes, são elas [34]: processamento local e processamento num
servidor remoto. Os métodos utilizados em incluem a posteriori o uso de Redes Neuronais
Artificial (RNA), Support Vector Machine (SVM), k-Nearest Neighbour (k-NN), entre outros [3542].
O tema seguinte foi o da imputação de dados, tendo sido concluído que os métodos utilizados
variam de acordo com o tipo de dados em falta, podendo os mesmos serem classificados como
dados em falta completamente aleatórios, dados em falta distribuídos por diferentes conjuntos
de dados e dados em falta não aleatórios [43, 44].
Relativamente à fusão de dados, os métodos utilizados dividem-se em três grandes grupos, são
eles: métodos probabilísticos, métodos estatísticos, métodos baseados na teoria do
conhecimento e métodos baseados nas evidências de raciocínio [45]. Os métodos probabilísticos
são métodos que permitem uma aprendizagem não supervisionada, estimando modelos para as
diversas tarefas, contudo necessitam ter um conhecimento prévio do sistema. Os métodos
estatísticos têm melhores resultados com a redução do erro, mas exigem uma grande
capacidade computacional. Os métodos baseados na teoria do conhecimento permitem a
inclusão da incerteza, são fáceis de implementar e robustos para dados com ruido, embora
exija intervenção humana para a aprendizagem. Os métodos baseados nas evidências de
raciocínio requerem um determinado grau de incerteza para cada uma das entrados dos
métodos. Contudo, a utilização destes métodos depende dos sensores utilizados, sendo que o
método mais usado para a fusão de dados é o filtro de Kalman [46], que é um algoritmo que
utiliza os mínimos quadrados recursivos ponderados dinamicamente, e as RNA [35-42], que são
também utilizados para a classificação dos dados.
Os conceitos até aqui apresentados nesta Tese podem ser utilizados para a identificação das
AVD e os seus ambientes, servindo de prova de conceito para o desenvolvimento de um
assistente pessoal digital. Nesta secção foram somente apresentados alguns exemplos, sendo
que serão apresentados os métodos de aquisição, processamento, imputação e fusão de dados
com mais detalhe no capítulo 2.
2. Técnicas de validação de dados de sensores utilizados em
aplicações móveis para a saúde
Aquando da captura dos dados dos diversos sensores disponíveis nos dipositivos móveis são
vários os problemas que podem acontecer, causando falhas na aquisição dos dados que pode
xxviii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
incluir a aquisição de dados inválidos ou a inexistência dos mesmos em alguns instantes durante
a aquisição dos dados.
A validação dos dados permite-nos reduzir os dados em falta ou incorretos, aplicando
mecanismos de filtragem e/ou imputação de dados. Assim, durante o processamento devem
ser aplicados filtros de novo a reduzir os valores inválidos presentes no sinal capturado, sendo
posteriormente aplicados mecanismos de imputação de dado no caso da existência de valores
em falta.
Os mecanismos de filtragem dos dados dependem do tipo de dados a serem adquiridos, isto é,
para os dados de sensores de movimento ou magnéticos, os mecanismos de filtragem são
normalmente constituídos por filtro de passa-baixo [47], enquanto para dados acústicos é
comummente aplicada a transformada de Fourier para extrair as frequências mais relevantes
[48].
A validação de dados, bem como a correção dos mesmos, permite a obtenção de resultados
com maior exatidão e em diferentes condições de aquisição de dados. Nesta secção foram
somente apresentados alguns exemplos, sendo que serão apresentados os métodos de
validação, filtragem e imputação de dados com mais detalhe no capítulo 2.
3. Técnicas de criação de impressões digitais de áudio para a
deteção do ambiente das atividades de vida diária
A impressão digital de áudio é aplicada para processar dados de áudio e permitir a identificação
das AVD. De acordo com os autores de [49], a implementação de um sistema que gera
impressões digitais de áudio inclui a definição de uma plataforma ou método, das
representações de impressões digitais, das estruturas de dados, das técnicas para a medição
de similaridades e dos métodos estatísticos para o processamento dos mesmo. Diversos autores
apresentaram diferentes métodos de criação d impressões digitais de áudio para diferentes
tipos de utilização [49-54]. Os resultados obtidos com as impressões digitais de áudio dependem
da granularidade dos dados armazenados na base de dados, bem como dos dados a serem
capturados para a comparação.
Com base nos vários estudos, a Google, Inc. implementou um método chamado Musicg [55],
que é uma API (Application Programming Interface) Java para análise de áudio, compatível
com o sistema operativo Android. Os recursos implementados incluem a deteção da entrada de
áudio que corresponde a um aplauso e um assobio, a leitura dos cabeçalhos WAVE do PCM
xxix
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
(Pulse-code modulation) e a leitura e seleção dos dados de áudio. Todas as diferentes
características consideradas para a impressão digital de áudio são processadas pelo teorema de
Nyquist para definir o número de amostras, a definição dos valores mínimos e máximos das
amplitudes da frequência, intensidade, desvio padrão, número de zeros e robustez do sinal de
áudio, bem como a definição de filtros passa-alta e passa-baixo para maximizar os recursos
importantes do sinal de áudio.
Nesta secção foram somente apresentados alguns exemplos, sendo que serão apresentados os
métodos de impressão digital de áudio com mais detalhe no capítulo 2.
Proposta de um método para identificação das atividades
de vida diária
Após a revisão de literatura relativa aos diversos conceitos, a arquitetura do método proposto
nesta Tese para a identificação das AVD e ambientes foi definida e apresentada, incluindo os
diversos conceitos abordados na revisão de literatura.
1. Abordagem para o desenvolvimento de um método para a
identificação de atividades de vida diária utilizando os
sensores de dispositivos móveis
O método proposto tem por objetivo a utilização de vários sensores disponíveis nos dispositivos
móveis, tais como acelerómetro, magnetómetro, giroscópio, microfone e recetor de GPS, sendo
que o método adaptar ao máximo número de sensores disponível em cada dispositivo móvel no
mercado. Assim, o método deve implementar quatro grandes módulos, são eles: aquisição de
dados, processamento de dados, fusão de dados e classificação de dados.
Relativamente ao módulo de aquisição de dados, o mesmo será implementado com os métodos
comuns de aquisição de dados, verificando-se desnecessária a utilização de métodos complexos
dados a sua simplicidade. Assim, serão adquiridos 5 segundos de dados dos diversos sensores
disponíveis para a identificação da AVD que está a ser realizada a cada 5 minutos, bem como o
ambiente a ser frequentado.
xxx
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
No caso do módulo de processamento de dados, o mesmo será composto por filtragem de dados,
imputação de dados e extração de características. A filtragem de dados depende dos sensores
utilizados, sendo que para o dados capturados com o acelerómetro, o magnetómetro e o
giroscópio devem ser implementados filtros de passa-baixo [47], para dados capturados com o
microfone deve ser aplicada a Transformada Rápida de Fourier (FFT) para extrair as frequências
mais relevantes [48], e, por fim, para os dados capturados com recurso ao recetor de GPS
verificou-se que não seria necessária a aplicação de qualquer método de filtragem. A imputação
de dados verificou-se desnecessária devido às características selecionadas para a classificação
de dados, sento sido implementada de seguida os métodos de extração de características. Como
explicado com mais detalhe no capitulo 2, para os dados capturados com o acelerómetro, o
magnetómetro e o giroscópio devem ser extraídas as cinco maiores distâncias entre os picos, a
médio, desvio padrão, variância e mediana dos picos máximos, e a média, desvio padrão,
máximo, mínimo, variância e mediana do sinal em bruto, para os dados adquiridos com o
microfone devem ser extraídos 26 coeficientes de MFCC, e a média, desvio padrão, máximo,
mínimo, variância e mediana do sinal em bruto, e para os dados adquiridos com o recetor de
GPS deve ser extraída a distância percorrida.
Relativamente aos módulos de fusão e classificação de dados são normalmente executados em
conjunto, tendo sido escolhido o uso de RNA, pois são os métodos que reportam melhores
resultados na literatura e são dos mais utilizados. Assim, no módulo de classificação começouse por identificar as AVD mais comuns como andar, correr, estar parado, subir escadas e descer
escadas, com o recurso aos dados adquiridos com o acelerómetro, magnetómetro e/ou
giroscópio. Numa fase posterior foram identificados os ambientes frequentados, como sala de
aula, ginásio, rua, hall, cozinha, bar, biblioteca, sala de estar e quarto, com recurso aos dados
adquiridos com o microfone. Finalmente, tendo em conta o ambiente identificado e os dados
extraídos com o recetor de GPS tornou-se possível a diferenciação de atividades que são
consideradas sem movimento com dados adquiridos com o acelerómetro, magnetómetro e
giroscópio, optando-se por diferenciar entre dormir, ver televisão e conduzir. Para a
classificação, propôs-se a comparação entre três diferentes bibliotecas, implementando
diversas configurações que serão apresentadas com mais detalhe no capitulo 4, incluindo
Multilayer Perceptron (MLP) with Backpropagation [56], Feedforward Neural Networks (FNN)
with Backpropagation [57] e Deep Neural Networks (DNN) [58].
Em conclusão, o método proposto inclui a classificação de 9 ambientes e 7 AVD que poderão
ser posteriormente incrementadas com novos desenvolvimentos.
xxxi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Fusão de dados para a identificação de atividades de vida
diária
Por fim, os detalhes da implementação do método para identificação das AVD e ambientes
foram apresentados, utilizando os diversos sensores disponíveis nos dispositivos móveis e com
foco nos métodos de fusão e classificação de dados.
1. Identificação de atividades de vida diária através a fusão dos
dados dos sensores magnéticos e de movimento dos
dispositivos móveis
Tendo em conta os dados previamente adquiridos e categorizados nas diferentes AVD (i.e.,
andar, correr, subir escadas, descer escadas e estar parado), disponíveis em [59], e a estrutura
do método apresentado no capítulo 3, os dados foram processados aplicando o filtro de passabaixo e foram extraídas as características dos dados adquiridos com os diferentes sensores, tais
como acelerómetro, giroscópio e magnetómetro, sendo exploradas as diferentes combinações
dos mesmos, isto é, foram explorados os resultados obtidos com o acelerómetro, com a
combinação do acelerómetro com o magnetómetro e, por fim, com a combinação entre
acelerómetro, magnetómetro e giroscópio. Antes da aplicação dos métodos de classificação, os
dados foram normalizados, sendo que para as configurações de Multilayer Perceptron (MLP)
with Backpropagation e Feedforward Neural Networks (FNN) with Backpropagation os dados
foram normalizados com recurso aos máximos e os mínimos e a normalização dos dados para o
método Deep Neural Networks (DNN) foi realizada com recurso à média e desvio padrão.
Assim, no decorrer deste estudo foram testados diferentes conjuntos de características
extraídas dos diferentes sensores, são elas:
•
Cinco maiores distâncias entre os picos máximos combinados com a média, desvio
padrão, variância e mediana dos picos máximos, e desvio padrão, média, máximo,
mínimo, variância e mediana do sinal em bruto;
•
Média, desvio padrão, variância e mediana dos picos máximos, combinados com o
desvio padrão, média, máximo, mínimo, variância e mediana do sinal em bruto;
•
xxxii
Desvio padrão, média, máximo, mínimo, variância e mediana do sinal em bruto;
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
•
Desvio padrão, média, variância e mediana do sinal em bruto;
•
Desvio padrão e média do sinal em bruto.
Os diferentes conjuntos de características foram utilizados para as diferentes combinações de
sensores, obtendo-se melhores resultados com a utilização de todas as características extraídas
nas diferentes combinações e com as diferentes configurações de RNA apresentadas no capítulo
4. Assim, obtiveram-se os seguintes resultados com as diferentes configurações de redes
neuronais implementadas com dados normalizados e não normalizados:
•
Acelerómetro:
o
Dados não normalizados:
§
Neste caso o método MLP reportou uma exatidão de 34.75%, enquanto
o método FNN reportou uma exatidão de 74.45% e o método DNN
reportou uma exatidão de 80.35%;
o
Dados normalizados:
§
Neste caso o método MLP reportou uma exatidão de 24.03%, enquanto
o método FNN reportou uma exatidão de 37.07% e o método DNN
reportou uma exatidão de 85.89%.
•
Acelerómetro e magnetómetro:
o
Dados não normalizados:
§
Neste caso o método MLP reportou uma exatidão de 35.15%, enquanto
o método FNN reportou uma exatidão de 42.75% e o método DNN
reportou uma exatidão de 70.43%;
o
Dados normalizados:
§
Neste caso o método MLP reportou uma exatidão de 24.93%, enquanto
o método FNN reportou uma exatidão de 64.94% e o método DNN
reportou uma exatidão de 86.49%.
•
Acelerómetro, magnetómetro e giroscópio:
o
Dados não normalizados:
§
Neste caso o método MLP reportou uma exatidão de 38.32%, enquanto
o método FNN reportou uma exatidão de 76.13% e o método DNN
reportou uma exatidão de 74.47%;
o
Dados normalizados:
§
Neste caso o método MLP reportou uma exatidão de 37.13%, enquanto
o método FNN reportou uma exatidão de 29.54% e o método DNN
reportou uma exatidão de 89.51%.
Em conclusão verificou-se que com os resultados obtidos é possível identificar um determinado
conjunto de AVD, sendo que o método que reporta melhores resultados é o método DNN com
dados normalizados em qualquer uma das combinações de sensores.
xxxiii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
2. Desenvolvimento de uma biblioteca para o reconhecimento
de
atividades
de
vida
diária:
considerações
de
implementação, desafios e soluções
Com base nos resultados obtidos com os estudos anteriores relacionados com a identificação
das AVD comuns, identificação de ambientes e identificação de atividades sem movimento, foi
implementada a estrutura do método anteriormente apresentado numa biblioteca desenvolvida
para dispositivos móveis com o sistema operativo Android. Durante a implementação
verificaram-se alguns desafios relacionados com a exatidão apresentada pelos métodos
anteriormente criados.
A aplicação móvel desenvolvida implementa métodos de aquisição de dados, implementa os
métodos de processamento e os métodos de fusão e classificação de dados anteriormente
apresentados. Contudo, especialmente na implementação do método para identificação de
ambientes que anteriormente tinha sido verificado que os melhores resultados eram obtidos
com o método de Feedforward Neural Networks (FNN) with Backpropagation com dados não
normalizados, na aplicação móvel foi notado um decréscimo da exatidão e do desempenho no
processamento de dados de áudio com este método, tendo sido testada a implementação deste
método em combinação com o método Deep Neural Networks (DNN) com dados normalizados,
verificando-se que os resultados melhoravam, mas os melhores resultados foram obtidos
somente com o utilização do método DNN com dados normalizados, melhorando o desempenho
e a exatidão dos resultados. Para as restantes fases, foi implementado o método DNN com
dados normalizados, tendo-se verificado os resultados esperados. Assim, os resultados obtidos
pela aplicação móvel desenvolvida são os seguintes:
•
Identificação de AVD comuns:
o
Acelerómetro:
§
o
Acelerómetro e magnetómetro:
§
o
Microfone
§
Os resultados reportados têm uma exatidão de 45,68%
Identificação de atividades sem movimento
o
Inclusão do recetor de GPS e/ou ambiente reconhecido:
§
xxxiv
Os resultados reportados têm uma exatidão de 89,15%.
Identificação de ambientes
o
•
Os resultados reportados têm uma exatidão de 86,49%;
Acelerómetro, magnetómetro e giroscópio:
§
•
Os resultados reportados têm uma exatidão de 86,39%;
Os resultados reportados têm uma exatidão de 100%;
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Em conclusão, os resultados reportam uma exatidão global entre 58,02% e 89.15%, que se
verificou por comparação com os resultados disponíveis na literatura que se obteve uma
melhoria de 2.93% com base nos resultados médios, tendo-se obtido resultados similares em
termos máximos (sem melhoria), embora o número máximo de eventos detetado pelos estudos
anteriores seja, no máximo, de 13, enquanto os resultados apresentados nesta Tese realizam a
identificação de 16 eventos.
Principais Conclusões
O foco desta Tese está relacionado ao desenvolvimento de um novo método para o
reconhecimento de AVD e seus ambientes, incluindo a seleção dos melhores métodos para os
diferentes módulos, como métodos de aquisição, fusão, processamento e filtragem de dados,
extração de características e classificação dos dados, com base nos sensores disponíveis num
smartphone. Os módulos mais importantes são os que incluem os métodos de fusão e
classificação de dados que influenciam diretamente os resultados obtidos pelo método
resultante da investigação desta Tese.
Os
dispositivos
móveis
integram
vários
sensores,
e.g.,
acelerómetro,
giroscópio,
magnetómetro, microfone e recetor de GPS, que permitem a aquisição de vários tipos de
parâmetros físicos e fisiológicos. Os dispositivos móveis podem ser considerados dispositivos
com múltiplos sensores, que consiste na primeira parte do título desta Tese. Existe uma grande
variedade de dispositivos móveis, mas o âmbito desta Tese foca-se apenas no uso de
smartphones. A aquisição dos diferentes parâmetros permite o reconhecimento das AVD, onde
o trabalho desta Tese é implementar métodos de fusão e classificação de dados para o seu
reconhecimento e reconhecimento de ambientes. Existem várias AVD que podem ser
reconhecidas, mas esta Tese foca-se no reconhecimento de sete AVD (i.e., andar, correr, subir,
descer, estar parado/assistir TV, dormir e conduzir) e nove ambientes (i.e., quarto, bar, sala
de aula, academia, cozinha, sala de estar, hall, rua e biblioteca) com recurso aos dados
recolhidos com os sensores disponíveis nos dispositivos móveis. No entanto, esses dispositivos
têm várias limitações ao nível da capacidade da bateria, de processamento e de memória,
limitações estas cujo impacto foi minimizado com a implementação de diferentes métodos. No
final, o método desenvolvido pode ser utilizado para o mapeamento de estilos de vida tendo
em conta o reconhecimento de AVD e seus ambientes.
O principal objetivo do trabalho desta Tese é a criação de um novo método para o
reconhecimento de AVD e ambientes com a fusão e classificação dos dados adquiridos dos
sensores disponíveis nos dispositivos móveis disponíveis. O método desenvolvido é deve ter em
xxxv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
conta a variedade e número de sensores disponíveis nos diferentes dispositivos móveis. Deve
ser implementado com algoritmos adaptados às características destes dispositivos, para
permitir a sua execução localmente. O objetivo principal foi dividido em seis objetivos
secundários, que começaram com a revisão da literatura, verificando que não existem estudos
que utilizem um grande número de sensores disponíveis nos dispositivos móveis e apliquem
técnicas de fusão de dados adquiridos pelos mesmos, mas é um tópico que está crescendo
amplamente com diferentes trabalhos de investigação. A revisão da literatura também revela
que as Redes Neurais Artificiais (RNA) são amplamente utilizadas e reportam melhores
resultados que outros métodos. Por fim, a revisão da literatura ajudou na definição das AVD e
ambientes a serem reconhecidos pelo método desenvolvido, bem como a definição da sua
arquitetura que inclui aquisição, fusão, processamento e filtragem dos dados, e extração de
características e classificação dos mesmos. A aquisição de dados dos sensores dos dispositivos
móveis é realizada por métodos básicos. Os métodos de processamento e filtragem de dados,
bem como de extração de características, dependem dos tipos de sensores utilizados, conforme
apresentado no capítulo 3, que é focado na definição da arquitetura do método para o
reconhecimento de AVD e ambientes. Seguidamente, a aquisição de dados foi realizada com
recurso a uma aplicação móvel que permite a captura e etiquetagem das AVD e dos ambientes
para análise e criação dos métodos. Após a aquisição dos dados, os mesmos os foram analisados
e os diferentes métodos foram implementados, incluindo filtros passa-baixo, Transformada
Rápida de Fourier (FFT), entre outros. Após a extração das características foram analisadas três
diferentes implementações de métodos de RNA, incluindo Multilayer Perceptron (MLP) with
Backpropagation, Feedforward Neural Networks (FNN) with Backpropagation e Deep Neural
Networks (DNN), apresentando mais detalhes no capítulo 4. As diferentes implementações
reportaram que o método que obtém os melhores resultados é o método DNN com dados
normalizados.
A primeira parte desta Tese, apresentada no capítulo 2, consiste na revisão da literatura sobre
os diferentes conceitos relacionados à criação de um novo método para o reconhecimento das
AVD e ambientes. Inicialmente, os conceitos estudados consistem na aquisição, processamento,
imputação e fusão de dados foram explorados, analisando os vários métodos anteriormente
utilizados para cada conceito, bem como a sua aplicabilidade. No entanto, as soluções
encontradas são difíceis de ser adaptadas aos dispositivos móveis, porque o desenvolvimento
de soluções para esses dispositivos deve ter em conta as diferentes restrições de hardware que
dependem dos fabricantes. Além disso, durante a primeira parte da revisão da literatura,
verificou-se que a imputação de dados faz parte do processamento de dados, no entanto esta
etapa foi excluída da investigação desta Tese, pois os métodos de classificação utilizam as
características extraídas dos dados dos sensores, não sendo relevante a inexistência de dados
em alguns instantes. A segunda parte da revisão de literatura consiste na pesquisa sobre os
métodos de validação de dados, incluindo métodos de filtragem e imputação de dados no
módulo de processamento de dados. Os métodos de filtragem de dados incluem os filtros de
xxxvi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
passa-baixo, filtros de passa-alto e outros; e os métodos de imputação de dados analisados
consistem principalmente no K-Nearest Neighbor (KNN) e as suas variantes, verificando-se que
um dos métodos mais utilizados com melhores resultados são as RNA. Como os dados acústicos
também foram utilizados para o reconhecimento dos ambientes, os métodos de impressão
digital de áudio também foram analisados, verificando-se que a FFT deve ser implementada
para a extração das frequências, concluindo-se que os métodos de classificação mais utilizados
com melhores resultados são os métodos estatísticos.
A segunda parte desta Tese, apresentada no capítulo 3, consiste na definição da arquitetura do
método proposto para o reconhecimento das AVD, mencionando-se que a arquitetura começa
com a aquisição de dados. Após a aquisição dos dados, os dados são processados, incluindo
filtragem, imputação e extração de características. A filtragem dos dados tem em conta os
tipos de sensores utilizados, implementando o método de filtro de passa-baixo para os sensores
de movimento e magnéticos e a FFT para os sensores acústicos. Primeiramente, para sensores
magnéticos e de movimento, as características extraídas são as cinco maiores distâncias entre
os picos máximos combinados com a média, desvio padrão, variância e mediana dos picos
máximos, e desvio padrão, média, máximo, mínimo, variância e mediana do sinal em bruto.
Seguidamente, para os sensores acústicos, as características extraídas são os 26 Mel-Frequency
Cepstral Coefficients (MFCC) combinados com o desvio padrão, média, máximo, mínimo,
variância e mediana do sinal. Finalmente, para os sensores de localização, a característica
extraída é a distância percorrida. Após a extração de características e com base nas RNA, o
método de classificação foi implementado. Os resultados da classificação devem ser
combinados com a agenda dos utilizadores para desenvolvimentos futuros.
A terceira parte desta Tese, apresentada no capítulo 4, consiste na apresentação dos resultados
principais desta Tese, que consistem na implementação dos diferentes métodos para o
reconhecimento das AVD e ambientes. A implementação apresentada nesta Tese foi
incremental, começando com a implementação com apenas um sensor, i.e., acelerómetro,
seguindo-se a implementação com sensores mais comuns nos smartphones, i.e., acelerómetro,
magnetómetro, giroscópio, microfone e recetor GPS. As implementações são realizadas com
três diferentes configurações de RNA, tais como o método MLP with Backpropagation, método
FNN with Backpropagation e método DNN com dados normalizados e não normalizados. Todas
as implementações são baseadas em vários grupos de características extraídas a partir dos
dados dos diferentes sensores. A normalização dos dados para os métodos MLP e FNN foi
realizada com recurso à normalização com os máximos e os mínimos e a normalização dos dados
para o método DNN foi realizada com recurso à média e desvio padrão. Os detalhes sobre as
configurações dos métodos de RNA estão disponíveis no capítulo 4. O método proposto realiza
o reconhecimento de ADL e ambientes em 3 fases, tais como o reconhecimento de AVD comuns
com sensores magnéticos e de movimento, o reconhecimento de ambientes com sensores
acústicos, e o reconhecimento de atividades sem movimento com sensores de movimento,
xxxvii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
magnéticos e de localização combinados com o ambiente reconhecido. Para os resultados
obtidos com os dados do acelerómetro, os melhores resultados foram obtidos com o método
DNN, reportando uma exatidão de 85,89% com dados normalizados e 80,35% com dados não
normalizados. Para os resultados obtidos com a fusão dos dados do acelerómetro e do
magnetómetro, os melhores resultados também foram obtidos com o método DNN, reportando
uma exatidão de 86,49% com dados normalizados e 70,43% com dados não normalizados. Para
os resultados obtidos com a fusão dos dados do acelerómetro, do magnetómetro e do
giroscópio, os melhores resultados foram também obtidos com o método DNN, reportando uma
exatidão de 89,51% com dados normalizados e 74,47% com dados não normalizados. Para os
resultados obtidos com os dados acústicos, os melhores resultados foram obtidos com o método
FNN with Backpropagation, reportando uma exatidão de 82,75% com dados normalizados e
86,50% com dados não normalizados. Para os resultados obtidos com a fusão dos sensores
utilizados anteriormente com os sensores de localização e/ou ambiente reconhecido, os
melhores resultados foram obtidos com o método DNN, relatando uma precisão de 100% com as
diferentes o método proposto foi implementado e validado num smartphone, reportando uma
exatidão geral entre 58,02% e 89,15%.
O objetivo principal desta Tese foi atingido com o desenvolvimento de um novo método para o
reconhecimento de AVD e seus ambientes, reconhecendo um maior número de AVD/ambientes
que os trabalhos anteriores que foram analisados, relatando uma exatidão geral superior à
média dos resultados, e exatidão semelhante aos melhores resultados reportados pelos
trabalhos existentes na literatura.
Direções para trabalho futuro
O reconhecimento das AVD e ambientes pode ser realizado com outros métodos, e os resultados
obtidos pelo método desenvolvido podem ser melhorados com pesquisa e implementação de
métodos de Imputação de dados a serem incorporados no módulo de processamento de dados.
O módulo de processamento de dados também pode ser melhorado com outras abordagens
relacionadas aos métodos de validação e filtragem de dados.
No entanto, o módulo mais importante do método para o reconhecimento de ADL e ambientes
é o módulo de classificação, no qual outros tipos de métodos de inteligência artificial podem
ser implementados e testados com os conjuntos de dados adquiridos durante esta Tese,
incluindo Support Vector Machines (SVM), Árvores de Decisão e outros métodos de
aprendizagem automática (e.g., Adaboost) para verificar se os resultados são melhores que
com outros métodos.
xxxviii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Tendo em conta os recursos limitados de bateria, processamento e memória dos smartphones,
o desafio mais importante no desenvolvimento do método é descobrir uma forma de adquirir e
processar os dados continuamente, sem efeitos no desempenho dos smartphones, a fim de
poder reconhecer situações de risco, e.g., detecção de quedas.
Atualmente, o método desenvolvido é preparado para smartphones. A implementação do
método para execução noutros tipos de dispositivos (e.g., smartwatches) exigiria mais
experimentação, adquirindo novos dados e desenvolvendo novos métodos de classificação
adaptados a esses dispositivos. Consequentemente, o método deve ser adaptado às
características específicas do dispositivo móvel em uso. Uma das diferenças mais importantes
na aquisição de dados entre o smartphone e o smartwatch é a posição deles, porque o
smartphone pode ser colocado no bolso frontal das calças do utilizador, mas o smartwatch é
comumente colocado no pulso.
Como trabalho futuro, o número de AVD e ambientes reconhecidos pelo método podem ser
incrementados, mas isso também requereria mais experiências. Como objetivo futuro, este
método pode ser usado no desenvolvimento de um assistente pessoal digital.
Referências
[1]
C. Dobre, C. x. Mavromoustakis, N. Garcia, R. I. Goleva, and G. Mastorakis, Ambient
Assisted Living and Enhanced Living Environments: Principles, Technologies and
Control: Butterworth-Heinemann, 2016
[2]
N. M. Garcia and J. J. P. Rodrigues, Ambient assisted living: CRC Press, 2015
[3]
M. Kara, O. Lamouchi, and A. Ramdane-Cherif, "A Quality Model for the Evaluation
AAL Systems," Procedia Computer Science, vol. 113, pp. 392-399, 2017/01/01/ 2017.
doi: https://doi.org/10.1016/j.procs.2017.08.354
[4]
S. Spinsante, E. Gambi, L. Raffaeli, L. Montanini, L. Paciello, R. Bevilacqua, et al.,
"Technology-based assistance of people with dementia: state of the art, open
challenges, and future developments," Human Monitoring, Smart Health and Assisted
Living: Techniques and Technologies, vol. 9, p. 55, 2017
[5]
N. M. Garcia, "A Roadmap to the Design of a Personal Digital Life Coach," in ICT
Innovations 2015, ed: Springer, 2016.
xxxix
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[6]
A. D’Ambrosio, M. Aria, and R. Siciliano, "Accurate Tree-based Missing Data
Imputation and Data Fusion within the Statistical Learning Paradigm," Journal of
Classification, vol. 29, pp. 227-258, 2012. doi: 10.1007/s00357-012-9108-1
[7]
J. Dong, D. Zhuang, Y. Huang, and J. Fu, "Advances in multi-sensor data fusion:
algorithms and applications," Sensors (Basel), vol. 9, pp. 7771-84, 2009. doi:
10.3390/s91007771
[8]
R. C. King, E. Villeneuve, R. J. White, R. S. Sherratt, W. Holderbaum, and W. S.
Harwin, "Application of data fusion techniques and technologies for wearable health
monitoring," Med Eng Phys, vol. 42, pp. 1-12, Apr 2017. doi:
10.1016/j.medengphy.2016.12.011
[9]
O. Banos, M. Damas, H. Pomares, and I. Rojas, "On the use of sensor fusion to reduce
the impact of rotational and additive noise in human activity recognition," Sensors
(Basel), vol. 12, pp. 8039-54, 2012. doi: 10.3390/s120608039
[10]
M. A. A. Akhoundi and E. Valavi, "Multi-Sensor Fuzzy Data Fusion Using Sensors with
Different Characteristics," arXiv preprint arXiv:1010.6096, 2010
[11]
P. Paul and T. George, "An Effective Approach for Human Activity Recognition on
Smartphone," 2015 Ieee International Conference on Engineering and Technology
(Icetech), pp. 45-47, 2015. doi: 10.1109/icetech.2015.7275024
[12]
Y.-W. Hsu, K.-H. Chen, J.-J. Yang, and F.-S. Jaw, "Smartphone-based fall detection
algorithm using feature extraction," in 2016 9th International Congress on Image and
Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Datong,
China, 2016, pp. 1535-1540.
[13]
S. Dernbach, B. Das, N. C. Krishnan, B. L. Thomas, and D. J. Cook, "Simple and
Complex Activity Recognition through Smart Phones," in 2012 8th International
Conference on Intelligent Environments (IE), Guanajuato, Mexico, 2012, pp. 214-221.
[14]
C. Shen, Y. F. Chen, and G. S. Yang, "On Motion-Sensor Behavior Analysis for HumanActivity Recognition via Smartphones," in 2016 Ieee International Conference on
Identity, Security and Behavior Analysis (Isba), Sendai, Japan, 2016, pp. 1-6.
[15]
S. D. Bersch, D. Azzi, R. Khusainov, I. E. Achumba, and J. Ries, "Sensor data
acquisition and processing parameters for human activity classification," Sensors
(Basel), vol. 14, pp. 4239-70, 2014. doi: 10.3390/s140304239
xl
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[16]
I. Pires, N. Garcia, N. Pombo, and F. Flórez-Revuelta, "From Data Acquisition to Data
Fusion: A Comprehensive Review and a Roadmap for the Identification of Activities of
Daily Living Using Mobile Devices," Sensors, vol. 16, p. 184, 2016
[17]
V. Pejovic and M. Musolesi, "Anticipatory Mobile Computing," ACM Computing Surveys,
vol. 47, pp. 1-29, 2015. doi: 10.1145/2693843
[18]
F. X. Lin, A. Rahmati, and L. Zhong, "Dandelion: a framework for transparently
programming phone-centered wireless body sensor applications for health," in
Wireless Health 2010, 2010, pp. 74-83.
[19]
O. Postolache, P. S. Girao, M. Ribeiro, M. Guerra, J. Pincho, F. Santiago, et al.,
"Enabling telecare assessment with pervasive sensing and Android OS smartphone," in
2011 IEEE International Workshop on Medical Measurements and Applications
Proceedings (MeMeA), 2011, pp. 288-293.
[20]
C. Siegel, A. Hochgatterer, and T. E. Dorner, "Contributions of ambient assisted living
for health and quality of life in the elderly and care services--a qualitative analysis
from the experts' perspective of care service professionals," BMC Geriatr, vol. 14, p.
112, 2014. doi: 10.1186/1471-2318-14-112
[21]
J. A. Botia, A. Villa, and J. Palma, "Ambient Assisted Living system for in-home
monitoring of healthy independent elders," Expert Systems with Applications, vol. 39,
pp. 8136-8148, 2012. doi: 10.1016/j.eswa.2012.01.153
[22]
N. M. Garcia, J. J. P. C. Rodrigues, D. C. Elias, and M. S. Dias, Ambient Assisted
Living: Taylor & Francis, 2014
[23]
M. Huch, A. Kameas, J. Maitland, P. J. McCullagh, J. Roberts, A. Sixsmith, et al.,
Handbook of Ambient Assisted Living: Technology for Healthcare, Rehabilitation and
Well-being - Volume 11 of Ambient Intelligence and Smart Environments: IOS Press,
2012
[24]
R. I. Goleva, N. M. Garcia, C. X. Mavromoustakis, C. Dobre, G. Mastorakis, R. Stainov,
et al., "AAL and ELE Platform Architecture," 2017
[25]
I. M. Pires, N. M. Garcia, N. Pombo, and F. Flórez-Revuelta, "Limitations of the Use of
Mobile Devices and Smart Environments for the Monitoring of Ageing People," in
ICT4AWE 2018 4th International Conference on Information and Communication
Technologies for Ageing Well and e-Health, Madeira, Portugal, 2018.
xli
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[26]
I. M. Pires, N. M. Garcia, N. Pombo, F. Flórez-Revuelta, and N. D. Rodríguez,
"Validation Techniques for Sensor Data in Mobile Health Applications," Journal of
Sensors, vol. 2016, pp. 1687-725, 2016
[27]
I. M. Pires, R. Santos, N. Pombo, N. M. Garcia, F. Florez-Revuelta, S. Spinsante, et
al., "Recognition of Activities of Daily Living Based on Environmental Analyses Using
Audio Fingerprinting Techniques: A Systematic Review," Sensors (Basel), vol. 18, Jan 9
2018. doi: 10.3390/s18010160
[28]
I. M. Pires, N. M. Garcia, and F. Flórez-Revuelta, "Multi-sensor data fusion techniques
for the identification of activities of daily living using mobile devices," in Proceedings
of the ECMLPKDD 2015 Doctoral Consortium, European Conference on Machine
Learning and Principles and Practice of Knowledge Discovery in Databases, Porto,
Portugal, 2015.
[29]
I. M. Pires, N. M. Garcia, N. Pombo, F. Flórez-Revuelta, and S. Spinsante,
"Identification of Activities of Daily Living through Data Fusion on Motion and
Magnetic Sensors embedded on Mobile Devices," in Pervasive and Mobile Computing,
vol. 47, pp. 78-93, 2018. doi: 10.1016/j.pmcj.2018.05.005
[30]
I. M. Pires, M. C. Teixeira, N. Pombo, N. M. Garcia, F. Flórez-Revuelta, S. Spinsante,
et al., "Android library for recognition of activities of daily living: implementation
considerations, challenges, and solutions," in Open Bioinformatics Journal, vol. 11,
pp. 61-88, Bentham Science Publishers B.V.. doi: 10.2174/1875036201811010061
[31]
L. H. A. Salazar, T. Lacerda, J. V. Nunes, and C. Gresse von Wangenheim, "A
Systematic Literature Review on Usability Heuristics for Mobile Phones," International
Journal of Mobile Human Computer Interaction, vol. 5, pp. 50-61, 2013. doi:
10.4018/jmhci.2013040103
[32]
D. Foti and J. S. Koketsu, "Activities of daily living," Pedretti’s Occupational Therapy:
Practical Skills for Physical Dysfunction, vol. 7, pp. 157-232, 2013
[33]
L. Lim, A. Misra, and T. Mo, "Adaptive data acquisition strategies for energy-efficient,
smartphone-based, continuous processing of sensor streams," Distributed and Parallel
Databases, vol. 31, pp. 321-351, 2012. doi: 10.1007/s10619-012-7093-3
[34]
K. Dolui, S. Mukherjee, and S. K. Datta, "Smart Device Sensing Architectures and
Applications," 2013 International Computer Science and Engineering Conference
(Icsec), pp. 91-96, 2013. doi: 10.1109/icsec.2013.6694759
xlii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[35]
A. Bujari, B. Licar, and C. E. Palazzi, "Movement pattern recognition through
smartphone's accelerometer," in Consumer Communications and Networking
Conference (CCNC), 2012 IEEE, Las Vegas, NV, 2012, pp. 502-506.
[36]
B. Chikhaoui, S. Wang, and H. Pigot, "A Frequent Pattern Mining Approach for ADLs
Recognition in Smart Environments," in Advanced Information Networking and
Applications (AINA), 2011 IEEE International Conference on, Biopolis, 2011, pp. 248255.
[37]
I. Kouris and D. Koutsouris, "A comparative study of pattern recognition classifiers to
predict physical activities using smartphones and wearable body sensors," Technol
Health Care, vol. 20, pp. 263-75, 2012. doi: 10.3233/THC-2012-0674
[38]
C. Zhu and W. Sheng, "Realtime recognition of complex human daily activities using
human motion and location data," IEEE Trans Biomed Eng, vol. 59, pp. 2422-30, Sep
2012. doi: 10.1109/TBME.2012.2190602
[39]
K. Farrahi and D. Gatica-Perez, "Daily Routine Classification from Mobile Phone Data,"
in Machine Learning for Multimodal Interaction. vol. 5237, ed: Springer Berlin
Heidelberg, 2008, pp. 173-184.
[40]
J.-H. Hong, J. Ramos, C. Shin, and A. K. Dey, "An Activity Recognition System for
Ambient Assisted Living Environments," in Evaluating AAL Systems Through
Competitive Benchmarking. vol. 362, ed: Springer Berlin Heidelberg, 2013, pp. 148158.
[41]
S. Phithakkitnukoon, T. Horanont, G. Lorenzo, R. Shibasaki, and C. Ratti, "ActivityAware Map: Identifying Human Daily Activity Pattern Using Mobile Phone Data," in
Human Behavior Understanding. vol. 6219, ed: Springer Berlin Heidelberg, 2010, pp.
14-25.
[42]
Y. Nam, S. Rho, and C. Lee, "Physical activity recognition using multiple sensors
embedded in a wearable device," ACM Transactions on Embedded Computing
Systems, vol. 12, pp. 1-14, 2013. doi: 10.1145/2423636.2423644
[43]
P. Vateekul and K. Sarinnapakorn, "Tree-Based Approach to Missing Data Imputation,"
presented at the Data Mining Workshops, 2009. ICDMW '09. IEEE International
Conference on Miami, FL, 2009.
[44]
W. Ling and F. Dong-Mei, "Estimation of Missing Values Using a Weighted K-Nearest
Neighbors Algorithm," presented at the Environmental Science and Information
Application Technology, 2009. ESIAT 2009. International Conference on, Wuhan, 2009.
xliii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[45]
N. Pombo, K. Bousson, P. Araujo, and J. Viana, "Medical decision-making inspired
from aerospace multisensor data fusion concepts," Inform Health Soc Care, vol. 40,
pp. 185-97, 2015. doi: 10.3109/17538157.2013.872113
[46]
F. Tanveer, O. T. Waheed, and Atiq-ur-Rehman, "Design and Development of a Sensor
Fusion based Low Cost Attitude Estimator," Journal of Space Technology,, vol. 1, pp.
45-50, Jun 2011
[47]
V. Graizer, "Effect of low-pass filtering and re-sampling on spectral and peak ground
acceleration in strong-motion records," in Proc. 15th World Conference of Earthquake
Engineering, Lisbon, Portugal, 2012, pp. 24-28.
[48]
C. Rader and N. Brenner, "A new principle for fast Fourier transformation," IEEE
Transactions on Acoustics, Speech, and Signal Processing, vol. 24, pp. 264-266, 1976.
doi: 10.1109/tassp.1976.1162805
[49]
P. J. O. Doets, Modeling Audio Fingerprints: Structure, Distortion, Capacity: TU
Delft, Delft University of Technology, 2010
[50]
P. Cano, E. Batle, T. Kalker, and J. Haitsma, "A review of algorithms for audio
fingerprinting," presented at the 2002 IEEE Workshop on Multimedia Signal Processing,
2002.
[51]
P. Cano, E. Batlle, E. Gómez, L. de C.T.Gomes, and M. Bonnet, "Audio Fingerprinting:
Concepts And Applications," vol. 2, pp. 233-245, 2005. doi: 10.1007/10966518_17
[52]
P. Cano, E. Batlle, T. Kalker, and J. Haitsma, "A Review of Audio Fingerprinting,"
Journal of VLSI signal processing systems for signal, image and video technology, vol.
41, pp. 271-284, 2005. doi: 10.1007/s11265-005-4151-3
[53]
P. Doets, M. M. Gisbert, and R. Lagendijk, "On the comparison of audio fingerprints
for extracting quality parameters of compressed audio," in Electronic Imaging 2006,
2006, pp. 60720L-60720L-12.
[54]
H. Kekre, N. Bhandari, N. Nair, P. Padmanabhan, and S. Bhandari, "A Review of Audio
Fingerprinting and Comparison of Algorithms," International Journal of Computer
Applications, vol. 70, 2013
[55]
Google. (2012, 2nd February). Google Code Archive - Long-term storage for Google
Code Project Hosting. Available: https://code.google.com/archive/p/musicg/
[56]
Neuroph. (2017, 2 Sep. 2017). Java Neural Network Framework Neuroph. Available:
http://neuroph.sourceforge.net/
xliv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[57]
H. Research. (2017, 2 Sep. 2017). Encog Machine Learning Framework. Available:
http://www.heatonresearch.com/encog/
[58]
A. Chris Nicholson. (2017, 2 Sep. 2017). Deeplearning4j: Open-source, Distributed
Deep Learning for the JVM. Available: https://deeplearning4j.org/
[59]
ALLab. (2017, September 2nd). August 2017- Multi-sensor data fusion in mobile
devices for the identification of activities of daily living - ALLab Signals. Available:
https://allab.di.ubi.pt/mediawiki/index.php/August_2017-_Multisensor_data_fusion_in_mobile_devices_for_the_identification_of_activities_of_daily_l
iving
xlv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
xlvi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Abstract
Following the recent advances in technology and the growing use of mobile devices such as
smartphones, several solutions may be developed to improve the quality of life of users in the
context of Ambient Assisted Living (AAL). Mobile devices have different available sensors, e.g.,
accelerometer, gyroscope, magnetometer, microphone and Global Positioning System (GPS)
receiver, which allow the acquisition of physical and physiological parameters for the
recognition of different Activities of Daily Living (ADL) and the environments in which they are
performed. The definition of ADL includes a well-known set of tasks, which include basic selfcare tasks, based on the types of skills that people usually learn in early childhood, including
feeding, bathing, dressing, grooming, walking, running, jumping, climbing stairs, sleeping,
watching TV, working, listening to music, cooking, eating and others. On the context of AAL,
some individuals (henceforth called user or users) need particular assistance, either because
the user has some sort of impairment, or because the user is old, or simply because users
need/want to monitor their lifestyle. The research and development of systems that provide a
particular assistance to people is increasing in many areas of application. In particular, in the
future, the recognition of ADL will be an important element for the development of a personal
digital life coach, providing assistance to different types of users. To support the recognition
of ADL, the surrounding environments should be also recognized to increase the reliability of
these systems.
The main focus of this Thesis is the research on methods for the fusion and classification of the
data acquired by the sensors available in off-the-shelf mobile devices in order to recognize ADL
in almost real-time, taking into account the large diversity of the capabilities and
characteristics of the mobile devices available in the market. In order to achieve this objective,
this Thesis started with the review of the existing methods and technologies to define the
architecture and modules of the method for the identification of ADL. With this review and
based on the knowledge acquired about the sensors available in off-the-shelf mobile devices,
a set of tasks that may be reliably identified was defined as a basis for the remaining research
and development to be carried out in this Thesis. This review also identified the main stages
for the development of a new method for the identification of the ADL using the sensors
available in off-the-shelf mobile devices; these stages are data acquisition, data processing,
data cleaning, data imputation, feature extraction, data fusion and artificial intelligence. One
of the challenges is related to the different types of data acquired from the different sensors,
but other challenges were found, including the presence of environmental noise, the positioning
of the mobile device during the daily activities, the limited capabilities of the mobile devices
and others. Based on the acquired data, the processing was performed, implementing data
cleaning and feature extraction methods, in order to define a new framework for the
xlvii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
recognition of ADL. The data imputation methods were not applied, because at this stage of
the research their implementation does not have influence in the results of the identification
of the ADL and environments, as the features are extracted from a set of data acquired during
a defined time interval and there are no missing values during this stage. The joint selection of
the set of usable sensors and the identifiable set of tasks will then allow the development of a
framework that, considering multi-sensor data fusion technologies and context awareness, in
coordination with other information available from the user context, such as his/her agenda
and the time of the day, will allow to establish a profile of the tasks that the user performs in
a regular activity day. The classification method and the algorithm for the fusion of the features
for the recognition of ADL and its environments needs to be deployed in a machine with some
computational power, while the mobile device that will use the created framework, can
perform the identification of the ADL using a much less computational power. Based on the
results reported in the literature, the method chosen for the recognition of the ADL is composed
by three variants of Artificial Neural Networks (ANN), including simple Multilayer Perceptron
(MLP) networks, Feedforward Neural Networks (FNN) with Backpropagation, and Deep Neural
Networks (DNN).
Data acquisition can be performed with standard methods. After the acquisition, the data must
be processed at the data processing stage, which includes data cleaning and feature extraction
methods. The data cleaning method used for motion and magnetic sensors is the low pass filter,
in order to reduce the noise acquired; but for the acoustic data, the Fast Fourier Transform
(FFT) was applied to extract the different frequencies. When the data is clean, several features
are then extracted based on the types of sensors used, including the mean, standard deviation,
variance, maximum value, minimum value and median of raw data acquired from the motion
and magnetic sensors; the mean, standard deviation, variance and median of the maximum
peaks calculated with the raw data acquired from the motion and magnetic sensors; the five
greatest distances between the maximum peaks calculated with the raw data acquired from
the motion and magnetic sensors; the mean, standard deviation, variance, median and 26 MelFrequency Cepstral Coefficients (MFCC) of the frequencies obtained with FFT based on the raw
data acquired from the microphone data; and the distance travelled calculated with the data
acquired from the GPS receiver. After the extraction of the features, these will be grouped in
different datasets for the application of the ANN methods and to discover the method and
dataset that reports better results. The classification stage was incrementally developed,
starting with the identification of the most common ADL (i.e., walking, running, going upstairs,
going downstairs and standing activities) with motion and magnetic sensors. Next, the
environments were identified with acoustic data, i.e., bedroom, bar, classroom, gym, kitchen,
living room, hall, street and library. After the environments are recognized, and based on the
different sets of sensors commonly available in the mobile devices, the data acquired from the
motion and magnetic sensors were combined with the recognized environment in order to
differentiate some activities without motion, i.e., sleeping and watching TV. The number of
xlviii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
recognized activities in this stage was increased with the use of the distance travelled,
extracted from the GPS receiver data, allowing also to recognize the driving activity.
After the implementation of the three classification methods with different numbers of
iterations, datasets and remaining configurations in a machine with high processing
capabilities, the reported results proved that the best method for the recognition of the most
common ADL and activities without motion is the DNN method, but the best method for the
recognition of environments is the FNN method with Backpropagation. Depending on the
number of sensors used, this implementation reports a mean accuracy between 85.89% and
89.51% for the recognition of the most common ADL, equals to 86.50% for the recognition of
environments, and equals to 100% for the recognition of activities without motion, reporting
an overall accuracy between 85.89% and 92.00%.
The last stage of this research work was the implementation of the structured framework for
the mobile devices, verifying that the FNN method requires a high processing power for the
recognition of environments and the results reported with the mobile application are lower
than the results reported with the machine with high processing capabilities used. Thus, the
DNN method was also implemented for the recognition of the environments with the mobile
devices. Finally, the results reported with the mobile devices show an accuracy between 86.39%
and 89.15% for the recognition of the most common ADL, equal to 45.68% for the recognition
of environments, and equal to 100% for the recognition of activities without motion, reporting
an overall accuracy between 58.02% and 89.15%.
Compared with the literature, the results returned by the implemented framework show only
a residual improvement. However, the results reported in this research work comprehend the
identification of more ADL than the ones described in other studies. The improvement in the
recognition of ADL based on the mean of the accuracies is equal to 2.93%, but the maximum
number of ADL and environments previously recognized was 13, while the number of ADL and
environments recognized with the framework resulting from this research is 16. In conclusion,
the framework developed has a mean improvement of 2.93% in the accuracy of the recognition
for a larger number of ADL and environments than previously reported.
In the future, the achievements reported by this PhD research may be considered as a start
point of the development of a personal digital life coach, but the number of ADL and
environments recognized by the framework should be increased and the experiments should be
performed with different types of devices (i.e., smartphones and smartwatches), and the data
imputation and other machine learning methods should be explored in order to attempt to
increase the reliability of the framework for the recognition of ADL and its environments.
xlix
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Keywords
Sensor data fusion, data collection, accelerometer, gyroscope, magnetometer, microphone,
Global Positioning System (GPS) receiver, signal processing, sensors signal, activities of daily
living, mobile devices, pattern recognition, data processing, data cleaning, data imputation,
feature extraction, artificial intelligence, machine learning, artificial neural networks.
l
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Contents
Dedicatory .................................................................................................... v
Acknowledgements ....................................................................................... vii
Foreword .................................................................................................... xi
List of Publications ...................................................................................... xiii
List of articles included in the thesis resulting from this 4-year doctoral research
programme ............................................................................................. xiii
Other publications resulting from the doctoral research programme not included in the
thesis ..................................................................................................... xiv
Resumo .................................................................................................... xvii
Resumo Alargado ......................................................................................... xxi
Abstract ................................................................................................... xlvii
Contents ..................................................................................................... li
List of Figures .............................................................................................. lv
List of Tables.............................................................................................. lvii
Acronyms ................................................................................................... lxi
Chapter 1
Introduction ................................................................................................. 1
1.
Thesis Focus and Scope ........................................................................... 1
2.
Problem Definition and Research Objectives .................................................. 3
3.
Thesis Statement ................................................................................... 5
4.
Main Contributions ................................................................................. 5
5.
Thesis Organization ................................................................................ 6
References ................................................................................................. 7
Chapter 2
State-of-the-art ........................................................................................... 11
1.
From Data Acquisition to Data Fusion: A Comprehensive Review and a Roadmap for
the Identification of Activities of Daily Living Using Mobile Devices ........................... 11
Abstract ............................................................................................... 13
li
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
1.
Introduction ................................................................................... 13
2.
Sensors ......................................................................................... 15
3.
Sensor Data Fusion in Mobile Devices ...................................................... 23
4.
Application in Ambient Assisted Living .................................................... 30
5.
Conclusions .................................................................................... 30
References ........................................................................................... 32
2.
Validation Techniques for Sensor Data in Mobile Health Applications ................... 41
Abstract ............................................................................................... 43
1.
Introduction ................................................................................... 43
2.
Data Validation Methods ..................................................................... 44
3.
Classification of Data Validation Methods ................................................ 46
4.
Applicability of the Sensor Data Validation Methods .................................... 47
5.
Conclusion ..................................................................................... 49
References ........................................................................................... 49
3.
Audio Fingerprinting Techniques for detection of the environment of Activities of
Daily Living (ADL): A Systematic Review ............................................................ 53
Abstract ............................................................................................... 55
1.
Introduction ................................................................................... 56
2.
Methodology ................................................................................... 56
3.
Results .......................................................................................... 63
4.
Discussion ...................................................................................... 70
5.
Conclusions .................................................................................... 74
References ........................................................................................... 75
Chapter 3
Framework for the Identification of Activities of Daily Living ................................... 79
1.
Approach for the Development of a Framework for the Identification of Activities of
Daily Living Using Sensors in Mobile Devices ....................................................... 79
Abstract ............................................................................................... 81
1.
Introduction ................................................................................... 81
2.
Related Work .................................................................................. 83
3.
Methods and Expected Results.............................................................. 92
4.
Discussion and Conclusions .................................................................. 95
References ........................................................................................... 96
Chapter 4
Data fusion for the Identification of Activities of Daily Living ................................. 103
1.
Identification of Activities of Daily Living through Data Fusion on Motion and Magnetic
Sensors embedded on Mobile Devices .............................................................. 103
lii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Abstract .............................................................................................. 105
1.
Introduction .................................................................................. 105
2.
Related Work ................................................................................. 106
3.
Methods ....................................................................................... 108
4.
Results ......................................................................................... 112
5.
Discussion ..................................................................................... 115
6.
Conclusions ................................................................................... 118
References .......................................................................................... 118
2.
Android library for recognition of activities of daily living: implementation
considerations, challenges, and solutions ......................................................... 121
Abstract .............................................................................................. 123
1.
Introduction .................................................................................. 124
2.
Methods ....................................................................................... 130
3.
Results ......................................................................................... 133
4.
Discussion ..................................................................................... 137
5.
Conclusions ................................................................................... 145
References .......................................................................................... 146
Chapter 5
Conclusions and Future Work ........................................................................ 151
1.
Final Conclusions ................................................................................ 151
2.
Future Work ...................................................................................... 154
liii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
liv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
List of Figures
Chapter 2
State-of-the-art
1. From Data Acquisition to Data Fusion: A Comprehensive Review and a Roadmap for
the Identification of Activities of Daily Living Using Mobile Devices
Figure 1. Schema of a multi-sensor mobile system to recognize activities of daily
living .............................................................................. 15
2. Validation Techniques for Sensor Data in Mobile Health Applications
Figure 1. Sequence of activities performed during the data validation process . 45
Figure 2. Different categories of data validation methods .......................... 48
3. Audio Fingerprinting Techniques for detection of the environment of Activities of
Daily Living (ADL): A Systematic Review
Figure 1. Flow diagram of the identification and inclusion papers ................. 64
Chapter 3
Framework for the Identification of Activities of Daily Living
1.
Approach for the Development of a Framework for the Identification of Activities of
Daily Living Using Sensors in Mobile Devices
Figure 1. Schema for the framework for the recognition of Activities of Daily
Living (ADL) ...................................................................... 90
Figure 2. Sensors used for the recognition of Activities of Daily Living (ADL) and
environments for each phase of development ............................. 94
Chapter 4
Data fusion of the Identification of Activities of Daily Living
1.
Identification of Activities of Daily Living through Data Fusion on Motion and
Magnetic Sensors embedded on Mobile Devices
Figure 1. Simplified diagram for the framework for the identification of ADL .. 111
Figure 2. Schema of the architecture of the different implementations of ANN
used in this study .............................................................. 113
Figure 3. Results obtained with Neuroph framework for the different datasets of
accelerometer and magnetometer sensors ................................ 114
Figure 4. Results obtained with Encog framework for the different datasets of
accelerometer and magnetometer sensors ................................ 114
Figure 5. Results obtained with DeepLearning4j framework for the different
datasets of accelerometer and magnetometer sensors. ................. 114
Figure 6. Results obtained with Neuroph framework for the different datasets of
accelerometer, magnetometer and gyroscope sensors .................. 115
lv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Figure 7. Results obtained with Encog framework for the different datasets of
accelerometer, magnetometer and gyroscope sensors .................. 115
Figure 8. Results obtained with DeepLearning4j framework for the different
datasets of accelerometer, magnetometer and gyroscope sensors .... 115
2.
Android library for recognition of activities of daily living: implementation
considerations, challenges, and solutions
Figure 1. Schema of the classification stage of the framework for the recognition
of ADL and their environments .............................................. 132
Figure 2. Design of the interface of the Android application developed ......... 133
Figure 3. Comparison between the minimum, maximum and average of the
accuracies obtained in the literature review with the common ADL and
environments recognized by the Android library developed. ........... 142
lvi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
List of Tables
Chapter 2
State-of-the-art
1. From Data Acquisition to Data Fusion: A Comprehensive Review and a Roadmap for
the Identification of Activities of Daily Living Using Mobile Devices
Table 1. Services classified by categories .............................................. 17
Table 2. Examples of data acquisition methods. ...................................... 19
Table 3. Data processing: architectures and methods................................ 21
Table 4. Examples of data imputation methods. ...................................... 22
Table 5. Advantages and disadvantages of the sensor data fusion methods. ..... 24
Table 6. Examples of sensor data fusion methods. ................................... 29
2. Validation Techniques for Sensor Data in Mobile Health Applications
Table 1. Classification of the data validation methods by functionality .......... 47
3. Audio Fingerprinting Techniques for detection of the environment of Activities of
Daily Living (ADL): A Systematic Review
Table 1. Study analysis .................................................................... 57
Table 2. Study summaries ................................................................. 61
Table 3. Distribution of the features extracted in the studies ...................... 70
Table 4. Distribution of the methods implemented in the studies ................. 71
Table 5. Potential accuracies for the top most accurate methods vs top most
mean accurate features ....................................................... 73
Chapter 3
Framework for the Identification of Activities of Daily Living
1. Approach for the Development of a Framework for the Identification of Activities of
Daily Living Using Sensors in Mobile Devices
Table 1. List of sensors available in mobile devices .................................. 83
Table 2. Summary of data acquisition methods ....................................... 84
Table 3. Relation between the types of sensors and the data cleaning techniques
allowed. .......................................................................... 86
Table 4. Relation between the sensors and the extracted features ............... 88
Table 5. Relation between the different types of sensor and some data fusion
methods .......................................................................... 89
Table 6. Relation between the different types of sensors and some pattern
recognition methods ........................................................... 91
Table 7. Sensors, Activities of Daily Living (ADL) and environments for recognition
with the framework proposed ................................................ 94
lvii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 4
Data fusion of the Identification of Activities of Daily Living
1. Identification of Activities of Daily Living through Data Fusion on Motion and
Magnetic Sensors embedded on Mobile Devices
Table 1. Distribution of the ADL extracted in the studies analysed. .............. 108
Table 2. Distribution of the features extracted in the studies analysed. ........ 109
Table 3. Distribution of the classification methods used in the studies analysed.
................................................................................... 109
Table 4. Configurations of the ANN methods implemented. ....................... 112
Table 5. Best accuracies obtained with the different frameworks, datasets and
number of iterations. ......................................................... 116
Table 6. Best accuracies obtained with the different frameworks, datasets and
number of iterations .......................................................... 116
Table 7. Best accuracies achieved by the method using only the accelerometer
sensor ............................................................................ 116
Table 8. Comparison between the best results achieved only using the
accelerometer sensors, and using the accelerometer and magnetometer
sensors ........................................................................... 117
Table 9. Comparison between the best results achieved only using the
accelerometer sensors, and using the accelerometer, magnetometer
and gyroscope sensors ........................................................ 117
Table 10. Comparison between the best results achieved only using the
accelerometer and magnetometer sensors, and using the accelerometer
and magnetometer sensors ................................................... 117
2. Android library for recognition of activities of daily living: implementation
considerations, challenges, and solutions
Table 1. Summary of the literature studies related to ADL recognition based on
the ANN method. .............................................................. 126
Table 2. Summary of the literature studies related to ADL recognition based on
the ANN method (cont.) ...................................................... 127
Table 3. Summary of the literature studies related to environment recognition
based on the ANN method .................................................... 127
Table 4. Summary of the literature studies related to environment recognition
based on the ANN method (cont.)........................................... 128
Table 5. Summary of the literature studies related to environment recognition
based on the ANN method (cont.)........................................... 128
Table 6. Average accuracy of the ADL recognized ................................... 128
Table 7. Average accuracy of the environments recognized ....................... 129
Table 8. Configurations of the classification methods .............................. 132
Table 9. Analysis of the results of the stage 1 with accelerometer sensor ...... 134
lviii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Table 10. Analysis of the results of the stage 1 with accelerometer and
magnetometer sensors ........................................................ 134
Table 11. Analysis of the results of the stage 1 with accelerometer,
magnetometer and gyroscope sensors ...................................... 135
Table 12. Analysis of the results of the stage 2 using FNN method with
Backpropagation ............................................................... 135
Table 13. Analysis of the results of the stage 2 using FNN method with
Backpropagation and DNN method .......................................... 135
Table 14. Analysis of the results of the stage 2 using DNN method ............... 136
Table 15. Analysis of the overall recognition of ADL with accelerometer and
environment recognized ...................................................... 136
Table 16. Analysis of the overall recognition of ADL with accelerometer,
magnetometer and environment recognized .............................. 136
Table 17. Analysis of the overall recognition of ADL with accelerometer,
magnetometer, gyroscope and environment recognized ................ 136
Table 18. Analysis of the overall recognition of ADL with accelerometer, GPS
receiver and environment recognized ...................................... 137
Table 19. Analysis of the overall recognition of ADL with accelerometer,
magnetometer, GPS receiver and environment recognized ............. 137
Table 20. Analysis of the overall recognition of ADL with accelerometer,
magnetometer, gyroscope, GPS receiver and environment recognized
................................................................................... 137
Table 21. Analysis of the accuracies reported in studies [65-68] using the
accelerometer sensor ......................................................... 137
Table 22. Analysis of the accuracies reported in studies [65-68] using the
accelerometer and the magnetometer sensors ........................... 138
Table 23. Analysis of the accuracies reported in studies [65-68] using the
accelerometer, the magnetometer and the gyroscope sensors ........ 138
Table 24. Analysis of the accuracies reported by the Android library using the
accelerometer sensor ......................................................... 138
Table 25. Analysis of the accuracies reported by the Android library using the
accelerometer and the magnetometer sensors ........................... 138
Table 26. Analysis of the accuracies reported by the Android library using the
accelerometer, the magnetometer and the gyroscope sensors ........ 138
Table 27. Recognition of each ADL in stage 1 ........................................ 139
Table 28. Recognition of each environment in stage 2 .............................. 139
Table 29. Statistical analysis of the recognition of each ADL in the stage 3 ..... 140
Table 30. Statistical analysis of the recognition of each ADL in the stage 3 (cont.)
................................................................................... 140
Table 31. Statistical analysis of the recognition of each ADL in the framework 140
lix
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Table 32. Statistical analysis of the recognition of each ADL in the framework
(cont.) ........................................................................... 141
Table 33. Comparison between the minimum, maximum and average of the
accuracies obtained in the literature review and the accuracy of our
results in the recognition of the selected ADL and environments. .... 142
Table 34. Normality tests ................................................................ 143
Table 35. Values of the Student's t-test for a sample mean for comparing our
results and average accuracy ................................................ 143
Table 36. Normality tests ................................................................ 143
Table 37. Values of the Student's t-test for a sample mean for comparing our
results with minimum, and our results with maximum accuracy....... 144
Table 38. Number of ADL/Environments recognized by the studies, where the
minimum, maximum and average of the accuracies obtained in the
literature review were verified .............................................. 144
Table 39. Normality tests ................................................................ 144
Table 40. Values of the Student's t-test for a sample mean for comparing our
number of ADL/Environments with the average number of
ADL/Environments recognized in the literature .......................... 144
lx
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Acronyms
AAL
Ambient Assisted Living
AANN
Auto-Associative Neural Network
ACQUA
Acquisition Cost-Aware Query Adaptation
ADL
Activities Of Daily Living
AGO
Accumulated Generating Operation
ALHMM
Adaptive Learning Hidden Markov Model
ALLab
Assisted Living Computing And Telecommunications Laboratory
ANN
Artificial Neural Networks
AR
Autoregressive
ARMA
Autoregressive Moving Averages
AS
Audio Signature
ASF
Audio Spectrum Flatness
ASRS
Automated Storage And Retrieval System
ASV
Algorithmic Sensor Validation
AVI
Automatic Vehicle Identification
BBQ
Barbie-Q
BMW
Balanced Multiwavelets
BP
Back Propagate
BSW
Binary Spray-And-Wait
BSWF
Binary Spray-And-Wait With Fusion
C-SPINE
Collaborative-Signal Processing In Node Environment
CAALYX
Complete Ambient Assisting Living Experiment
CAN
Controller Area Network
CBSN
Collaborative Body Sensor Networks
CHG
Continuous Hand Gestures
CkWAS
Correlation Based K-Weighted Angular Similarity
CNT
Carbon Nanotube
COUPON
Cooperative Framework For Building Sensing Maps In Mobile Opportunistic
Networks
CQT
Constant Q Transform
CRF
Conditional Random Field
CS
Compressive Sampling
D2FAS
Decentralized Data Fusion And Active Sensing
DAL
Divide-And-Locate
DBN
Dynamic Bayesian Network
lxi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
DCT
Discrete Cosine Transform
DD
Diverse Density
DFF
Depth From Focus
DFT
Discrete Fourier Transform
DNN
Deep Neural Networks
DNRF
Drift And Noise Removal Filter
DR
Dead-Reckoning
DSC
Distributed Sparsity Classifier
DTW
Dynamic Time Warping
ECG
Electrocardiography
EDA
Electrodermal Activity
EEG
Electroencephalography
EKF
Extended Kalman Filter
EM
Expectation-Maximization
EMD
Empirical Mode Decomposition
EMG
Electromiography
EOG
Electrooculography
EQ
Estimation Quantization
ER
Epidemic Routing
ERF
Epidemic Routing With Fusion
FCMimpute
Fuzzy C-Means Clustering Imputation
FDI
Fault Detection And Isolation
FFT
Fast Fourier Transform
FFTW
Fastest Fourier Transform In The West
FKL
Fisher Kernel Learning
FLk-NN
Fourier And Lagged K-NN Combined System
FNN
Feedforward Neural Networks
FR
Frequency Response
GBM
Grey Bootstrap Method
GDA
Gaussian Discriminant Analysis
GMM
Gaussian Mixture Model
GPS
Global Positioning System
GPU
Graphic Processing Unit
GTD
Generalized Trend Diffusion
HAS
Human Auditory System
HHMM
Two-Level Hierarchical Hidden Markov Model
HMM
Hidden Markov Models
HNR
Harmonic-To-Noise Ratio
HPF
High Pass Filter
HSV
Heuristic Sensor Validation
lxii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
IB
Information Bottleneck
IBL
Instance-Based Learning
IBV
Index Bit Vectors
IDFT
Inverse Discrete Fourier Transform
IEPE
Integrated Electronic Piezoelectric
IPTV
Internet Protocol Television
IR
Infrared
IRTR
Improved Real-Time TV-Channel Recognition
IT
Instituto de Telecomunicações
ITree
Imputation Tree
JDL
Joint Directors Of Laboratories
KD
K-Dimensional
KMI
K-Means-Based Imputation
KNN
K-Nearest Neighbor
KNNimpute
K-Nearest Neighbour Imputation
KPCA
Kernel Principal Component Analysis
LDA
Linear Description Analysis
LDB
Local Discriminant Bases
LEC
Local Energy Centroid
LMCE
Local Maximum Chroma Energy
LMCLT
Logarithmic Module Complex Lapped Transform
LMT
Logistic Model Trees
LPC
Linear Predictive Coding
LPCC
Linear Prediction Coefficient Derived Cepstral Coefficients
LPF
Low-Pass Filter
LSM
Least Squares Method
LSTM
Long Short Term Memory
MA
Moving Average
MAR
Missing At Random
MASK
Masked Audio Spectral Keypoints
MCAR
Missing Completely At Random
MDCT
Modified Discrete Cosine Transform
MEI
Mean Imputation
MEMS
Micro-Electro-Mechanical Systems
MFCC
Mel-Frequency Cepstral Coefficients
MICE
Multivariate Imputation By Chained Equations
MKNNimpute
K-Nearest Neighbour Imputation Method Based On Mahalanobis Distance
MLP
Multilayer Perceptron
MMF
Multi-Matrices Factorization Model
MNAR
Missing Not At Random
lxiii
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
MP
Matching Pursuit
MPT
Missing Pattern Tree
MSDF
Mobile-Sensor Data Fusion
MSE
Mean Squared Error
MV
Magnitude Of Vector
NFC
New Field Communication
NTP
Network Time Protocol
OSGi
Open Services Gateway Initiative
PARs
Personal Audio Recordings
PAT
Phone Acceleration Threshold
PCA
Principal Component Analysis
PCM
Pulse Code Modulation
PDF
Probability Density Function
PDLC
Personal Digital Life Coach
PIR
Passive Infra-Red
PNN
Probabilistic Neural Networks
PPR
Phone Pattern Recognition
PRH
Philips Robust Hash
PSN
Personal Navigation System
QDA
Quadratic Discriminant Analysis
QIFFT
Quadratically Interpolated Fast Fourier Transform
QoS
Quality Of Service
QUC
Query Context
RBA
Rule-Based Algorithms
RBF
Radial Basis Function
RBFNN
Radial Basis Function Neural Network
RBUKF
Rao-Blackwellization Unscented Kalman Filter
RFID
Radio-Frequency Identifier
RFS
Rough And Fuzzy Sets
RMS
Root-Mean-Square
RNN
Recurrent Neural Networks
RRP
Random Recursive Partitioning
RSE
Relative Spectral Entropy
RTLS
Real-Time Location Systems
RVM
Relevance Vector Machine
SAD
Speech Activity Detection
SAF
Streaming Audio Fingerprinting
SEVA
Self-Validating
SKNNimpute
Sequential K-Nearest Neighbour Method-Based Imputation
SLA
Statistical Learning Algorithms
lxiv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
SMO
Sequential Minimal Optimization
SMV
Signal Magnitude Vector
SNR
Signal-To-Noise Ratio
SOM
Self-Organizing Maps
SPB
Symmetric Pairwise Boosting
SPR
Surface Plasmon Resonance
SPRT
Sequential Probability Ratio Test
SRR
Search By Range Reduction
SSC
Spectral Subband Centroid
STFT
Short-Time Fourier Transform
SVD
Singular Value Decomposition
SVDD
Support Vector Data Description
SVM
Support Vector Machines
SWNC
Sensor Weighted Network Classifier
TBA
Threshold Based Algorithm
TO-Combo-
Threshold Optimized Combo Speech Activity Detection
SAD
TSM
Time-Scale Modification
TSS
Two-Stage Search
UBI
Universidade da Beira Interior
UKF
Unscented Kalman Filter
VLSI
Very Large-Scale Integration
VRFV
Validated Random Fuzzy Variable
VRS
Video Retrieval System
VU
Validated Uncertainty
WIP
Walking-In-Place
WMA
Weighted Moving Average
WSN
Wireless Sensor Network
WTPR
Watch Threshold And Pattern Recognition
ZCR
Zero-Crossing-Rate
lxv
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
lxvi
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 1
Introduction
This Thesis addresses the problem of the automatic recognition of Activities of Daily Living
(ADL) [1] and the environments in which they are performed based on the fusion of the data
acquired by the sensors available in off-the-shelf mobile devices, such as the accelerometer,
the gyroscope, the magnetometer, the microphone and the Global Positioning System (GPS)
receiver [2]. As a result, this Thesis proposes an architecture for a framework to perform the
recognition of ADL and its environments. The focus, scope and research objectives of the Thesis
are described in this chapter, followed by the Thesis statement, the main contributions and
the Thesis organization.
1. Thesis Focus and Scope
Following a recurrent definition of the literature, the identification of ADL focuses on the
identification of a well-known set of tasks that are basic self-care tasks, based on the kinds of
skills that people usually learn in early childhood. These activities include feeding, bathing,
dressing, grooming, moving without danger, and other simple tasks related to personal care
and hygiene. On the context of Ambient Assisted Living (AAL) [3, 4], some individuals
(henceforth called user or users) need particular assistance, either because the user has some
sort of disability, or because the user is an older adult, or simply because the user needs or
wants to monitor and train his/her lifestyle. In this later context, ADL include a wider range of
tasks, not only related to personal care and hygiene, but also extended to social and
professional tasks. The development of methods for the recognition of ADL is important to
support the autonomy of older people, patients with chronic diseases, and users that may have
some type of disability [5, 6]. However, it may be useful for everyone, including athletes and
young users, as these solutions can be integrated in a tool for the monitoring and training of
lifestyles [7].
To date, the use of sensors for the recognition of ADL and environments is a well studied topic
[8-16]. However, there are no studies that define a structured architecture for the recognition
of ADL and environments using a large set of sensors available in off-the-shelf mobile devices
and for the recognition of a large set of ADL. The focus of this Thesis includes the design of a
novel framework through the selection and development of the most adequate methods for
each stage of the recognition of ADL and its environments, being these: Data Acquisition, Data
1
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Fusion, Data Processing, Data Cleaning, Feature Extraction and Classification; with a special
focus in the methods related to Data Fusion and Classification, concluding on what are the most
important aspects to address in order to obtain results with reliable accuracy.
The definition of a set of tasks that may be reliably identified, includes sensing and measuring
a large set of tasks and identifying those tasks that are reliably identified with the sensors
available in the mobile devices. The joint selection of the set of usable sensors and the
identifiable set of tasks allow the development of a tool that, considering multi-sensor data
fusion technologies and context awareness, in coordination with other information available
from the user context, such as his/her agenda and the time of the day, will further allow to
establish a profile of the tasks that the user performs in a regular activity day.
There are several challenges involved in the development of a novel framework for the
recognition of ADL. For instance, related to Data Acquisition, the challenges include the
incorrect positioning of the mobile device, the uncontrolled data sampling rate, the
unavailability of all sensors needed to recognize a specific ADL, and the different environmental
conditions [17]. There are several frameworks that may help to minimize these problems,
including the ones referred in the article presented in chapter 2 [18]. In relation to Data
Processing, the challenges are linked to the different architectures and platforms in which the
solution may be implemented, such as for example, relying on server-side processing or on local
processing. In order to provide a reliable result in several data collection environments where
a network connection may not be available and to be adapted to the low processing, storage
and memory resources of some mobile devices, the implementation of lightweight methods is
also object of study of this Thesis [17, 19-21].
The scope of this Thesis consists in the use of data fusion and Artificial Neural Networks (ANN)
methods for the recognition of ADL and its environments based on the data collected from
sensors available in an off-the-shelf mobile device, such as a smartphone. The limitations of
smartphones were also explored, creating a method that acquires 5 seconds of sensors’ data
every 5 minutes, allowing the mapping of the lifestyles and handling the development of new
solutions for AAL. The research part related to the classification methods of this Thesis is
related to the comparison of different implementations of ANN methods adapted to the
constrained resources of a mobile device, using local processing techniques, in order to provide
feedback to the user related to the ADL performed and its environment in an almost real-time
manner.
2
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
2. Problem Definition and Research Objectives
The problem addressed in this Thesis is to segment and identify correctly a significant set of
ADL that users perform as their daily or weekly routines, using the sensors and user information
available on a commodity smartphone. The identification of ADL is a complex challenge, due
to the existence of a large set of ADL and the complexity of the identification of the correct
method to identify each one of them. To correctly detect and identify an ADL, various factors
that influence the acquisition of the data and, consequently, affect the identification of several
types of activities, have to be taken into consideration, such as the environmental problems,
the involuntary movements that the user performs, the position of the mobile device, the
movements of the mobile device during the data acquisition, and the frequency of the
acquisition of data from the sensors.
The recognition of ADL and the environment where the subject is at a given moment is
important for the creation of systems aimed at several purposes, such as personal coaching,
health monitoring, lifestyle monitoring and others. Nowadays, such recognition is commonly
performed by complex and often expensive systems with high processing power and memory
capabilities [3, 22-26]. However, to a reasonable extent, this task can be performed with offthe-shelf mobile devices, supported by local processing techniques.
To date, and based on the available literature, there are several studies using different subsets
of sensors available in off-the-shelf mobile devices. However, studies reporting methods that
rely on a larger set of sensors available in mobile devices for the recognition of ADL and its
environments were not found. Currently, existing literature reports the use of Artificial Neural
Networks (ANN) and their variants as the most used methods, claiming relevant accuracies for
the identification of some ADL, as presented in [18].
The main objective of this Thesis is the creation of a new method and framework for the
recognition of ADL and its environments with the fusion and classification of the data acquired
from a large set of sensors available in the off-the-shelf mobile devices. The developed method
is a function of the number of sensors available on the different mobile devices that should be
executed with lightweight algorithms in order to allow its execution locally in the mobile
devices [27]. The developed method includes several stages, including data acquisition, data
processing, data cleaning, feature extraction and pattern recognition. The ADL that were
identified during this Thesis are walking, running, going upstairs, standing/watching TV, going
downstairs, driving and sleeping, and environments that were identified during this Thesis were
bedroom, bar, classroom, gym, kitchen, living room, hall, street and library.
The following intermediate objectives were defined in order to divide and organize the research
work required to accomplish the main objective of this thesis:
3
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
1. In order to understand the existing solutions and possible methods, the Thesis started
with the review of the state of the art related to the different concepts included in
this Thesis. Additionally, due to the types of data used in this work, this review also
included the study of data validation and audio fingerprinting methods. Finally, the
useful sensors, and the ADL and their environments that can be identified with the data
acquired from off-the-shelf mobile devices were described.
2. As second intermediate objective, the proposed architecture of the framework for the
recognition of ADL and environments was defined, including the different concepts
available in the literature.
3. As third intermediate objective, the development of a mobile application that acquires
the data from the sensors available in mobile devices has been performed. Some users
were selected to perform ADL and label them. For this data set the mobile device was
worn in the front pocket of the user’s trousers. This mobile application performs the
data acquisition process.
4. After
data
acquisition,
the
fourth
intermediate
objective
consists
in
the
implementation of data processing methods, including data cleaning and feature
extraction. Data imputation methods were avoided, because their implementation does
not have influence in the results of the identification of the ADL and environments,
because the features are measured with a data set acquired during a defined time
interval.
5. After the processing of the data, the fifth intermediate objective is related to the
implementation and comparison of three different configurations of ANN in order to
discover the best configuration for the recognition of ADL and environments.
Additionally, this analysis included the fusion of the different combinations of sensors’
data for the implementation of different methodologies.
6. The sixth intermediate objective consists in the implementation of methods addressing
the different stages of the framework for the recognition of ADL and environments,
implementing the framework in a mobile application for further testing of its
performance.
The recognition of ADL and environments implemented as results of this Thesis is important to
design a personal digital life coach [7]. It is also important to support the autonomy of older
users, patients with chronic diseases and users that may have some type of disability [5, 6].
Therefore, it may be useful for everyone, including athletes and young users, as the proposed
framework can be integrated in a tool for the monitoring and training of lifestyles [7].
4
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
3. Thesis Statement
This Thesis proposed a new framework for the recognition of ADL and environments, fusing the
data acquired with the sensors available in off-the-shelf mobile devices. Specifically, the Thesis
statement is:
The automatic identification of the Activities of Daily Living (ADL) and environments is a
complex challenge widely studied in the literature, but there are no structured frameworks
for the recognition of ADL. There are different types of data that can be acquired with mobile
devices, where the features should be defined and fused for the increasing reliability of the
recognition of ADL and environments. Data fusion and classification are the most important
steps in the development of a method for the recognition of ADL and environments. The
positioning of the mobile device and the timeframe of the data acquisition are also important
factors that should be taken in account.
To support this Thesis statement, the literature review about the different concepts involved
in this research was done, including data acquisition, data processing, data validation, data
cleaning, data imputation, feature extraction, audio fingerprinting, data fusion and
classification methods, proposing an architecture for a framework to recognize the ADL and
environments. After that, the data acquisition have been performed for the ADL and
environments selected. When all data sets were available, the implementation of the different
methods for each stage of the framework was performed, comparing three different types of
ANN for the creation of the different methods with the different subsets of sensors. At the end,
the framework was implemented in a mobile application for further testing.
4. Main Contributions
This section briefly describes the main scientific contributions resulting from the research work
presented in this Thesis.
The first contribution is the proposal of a framework for the recognition of ADL and it also
describes the existing methods related to each concept related to this Thesis, including sensor
classification, data acquisition, data processing, data imputation and sensor data fusion in
mobile devices, and the applicability of the recognition of ADL [18].
The second contribution of this Thesis presents the existing methods for the validation of the
data acquired from sensors, their classification and their applicability based on the different
types of missing and/or invalid data [28].
5
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
The third contribution of this Thesis presents the existing audio fingerprinting methods used
for the recognition of the environments based on the different acoustic signals, discussing the
accuracy of the features and methods implemented [29].
The fourth contribution of this Thesis details the architecture of the proposed framework for
the recognition of ADL and their environments, proposing possible methods for each module of
the framework [30].
The fifth contribution of this Thesis presents the research about the recognition of walking,
standing, running, going upstairs and going downstairs using the data acquired from the
accelerometer,
magnetometer
and
gyroscope
sensors,
comparing
three
different
implementations of ANN with normalized and non-normalized data in order to discover the best
implementation for the recognition of these ADL. The conclusion is that the best method is the
DNN method with normalized data [31].
The sixth contribution of this Thesis presents the results of the implementation of an Android
library with the framework for the recognition of ADL and environments based on the results
reported by previous studies, verifying the reliability of the framework proposed [32].
5. Thesis Organization
This Thesis is organized as follows:
Chapter 1: Introduction
A brief introduction to the Thesis is presented including the focus and scope, Thesis objectives,
Thesis statement, and main contributions of the work carried out.
Chapter 2: State-of-the-art
The main concepts involved in this PhD Thesis are presented and discussed, including data
acquisition, data processing, data validation, data cleaning, data imputation, feature
extraction, data fusion and classification methods. Additionally, the methods for the processing
of the acoustic data were also presented and discussed. These concepts are presented in three
papers.
Chapter 3: Framework for the Identification of Activities of Daily Living
This chapter presents the proposed architecture and possible methods for the development of
the framework for the recognition of Activities of Daily Living. The proposed architecture of
the framework is presented in one paper.
6
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 4: Data fusion of the Identification of Activities of Daily Living
This chapter presents and discusses the use of Artificial Neural Networks (ANN) and the
comparison between three different implementations for the recognition of ADL and its
environments. This chapter also presents the implementation of the different stages of the
framework for these recognitions. The implementation of the framework is presented in two
papers, as a small overview of the implementation of the framework for the recognition of ADL
and its environments.
Chapter 5: Conclusion and Future Work
The results presented throughout the Thesis are discussed and the main achievements are
summarized, pointing directions for the future.
References
[1]
D. Foti and J. S. Koketsu, "Activities of daily living," Pedretti’s Occupational Therapy:
Practical Skills for Physical Dysfunction, vol. 7, pp. 157-232, 2013
[2]
L. H. A. Salazar, T. Lacerda, J. V. Nunes, and C. Gresse von Wangenheim, "A
Systematic Literature Review on Usability Heuristics for Mobile Phones," International
Journal of Mobile Human Computer Interaction, vol. 5, pp. 50-61, 2013. doi:
10.4018/jmhci.2013040103
[3]
C. Dobre, C. x. Mavromoustakis, N. Garcia, R. I. Goleva, and G. Mastorakis, Ambient
Assisted Living and Enhanced Living Environments: Principles, Technologies and
Control: Butterworth-Heinemann, 2016
[4]
N. M. Garcia and J. J. P. Rodrigues, Ambient assisted living: CRC Press, 2015
[5]
M. Kara, O. Lamouchi, and A. Ramdane-Cherif, "A Quality Model for the Evaluation
AAL Systems," Procedia Computer Science, vol. 113, pp. 392-399, 2017/01/01/ 2017.
doi: https://doi.org/10.1016/j.procs.2017.08.354
[6]
S. Spinsante, E. Gambi, L. Raffaeli, L. Montanini, L. Paciello, R. Bevilacqua, et al.,
"Technology-based assistance of people with dementia: state of the art, open
challenges, and future developments," Human Monitoring, Smart Health and Assisted
Living: Techniques and Technologies, vol. 9, p. 55, 2017
[7]
N. M. Garcia, "A Roadmap to the Design of a Personal Digital Life Coach," in ICT
Innovations 2015, pp. 21-27, Springer, 2016.
7
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[8]
A. D’Ambrosio, M. Aria, and R. Siciliano, "Accurate Tree-based Missing Data
Imputation and Data Fusion within the Statistical Learning Paradigm," Journal of
Classification, vol. 29, pp. 227-258, 2012. doi: 10.1007/s00357-012-9108-1
[9]
J. Dong, D. Zhuang, Y. Huang, and J. Fu, "Advances in multi-sensor data fusion:
algorithms and applications," Sensors (Basel), vol. 9, pp. 7771-84, 2009. doi:
10.3390/s91007771
[10]
R. C. King, E. Villeneuve, R. J. White, R. S. Sherratt, W. Holderbaum, and W. S.
Harwin, "Application of data fusion techniques and technologies for wearable health
monitoring," Med Eng Phys, vol. 42, pp. 1-12, Apr 2017. doi:
10.1016/j.medengphy.2016.12.011
[11]
O. Banos, M. Damas, H. Pomares, and I. Rojas, "On the use of sensor fusion to reduce
the impact of rotational and additive noise in human activity recognition," Sensors
(Basel), vol. 12, pp. 8039-54, 2012. doi: 10.3390/s120608039
[12]
M. A. A. Akhoundi and E. Valavi, "Multi-Sensor Fuzzy Data Fusion Using Sensors with
Different Characteristics," arXiv preprint arXiv:1010.6096, 2010
[13]
P. Paul and T. George, "An Effective Approach for Human Activity Recognition on
Smartphone," 2015 Ieee International Conference on Engineering and Technology
(Icetech), pp. 45-47, 2015. doi: 10.1109/icetech.2015.7275024
[14]
Y.-W. Hsu, K.-H. Chen, J.-J. Yang, and F.-S. Jaw, "Smartphone-based fall detection
algorithm using feature extraction," in 2016 9th International Congress on Image and
Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), Datong, China,
2016, pp. 1535-1540.
[15]
S. Dernbach, B. Das, N. C. Krishnan, B. L. Thomas, and D. J. Cook, "Simple and
Complex Activity Recognition through Smart Phones," in 2012 8th International
Conference on Intelligent Environments (IE), Guanajuato, Mexico, 2012, pp. 214-221.
[16]
C. Shen, Y. F. Chen, and G. S. Yang, "On Motion-Sensor Behavior Analysis for HumanActivity Recognition via Smartphones," in 2016 Ieee International Conference on
Identity, Security and Behavior Analysis (Isba), Sendai, Japan, 2016, pp. 1-6.
[17]
S. D. Bersch, D. Azzi, R. Khusainov, I. E. Achumba, and J. Ries, "Sensor data
acquisition and processing parameters for human activity classification," Sensors
(Basel), vol. 14, pp. 4239-70, 2014. doi: 10.3390/s140304239
8
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[18]
I. Pires, N. Garcia, N. Pombo, and F. Flórez-Revuelta, "From Data Acquisition to Data
Fusion: A Comprehensive Review and a Roadmap for the Identification of Activities of
Daily Living Using Mobile Devices," Sensors, vol. 16, p. 184, 2016
[19]
V. Pejovic and M. Musolesi, "Anticipatory Mobile Computing," ACM Computing Surveys,
vol. 47, pp. 1-29, 2015. doi: 10.1145/2693843
[20]
F. X. Lin, A. Rahmati, and L. Zhong, "Dandelion: a framework for transparently
programming phone-centered wireless body sensor applications for health," in
Wireless Health 2010, 2010, pp. 74-83.
[21]
O. Postolache, P. S. Girao, M. Ribeiro, M. Guerra, J. Pincho, F. Santiago, et al.,
"Enabling telecare assessment with pervasive sensing and Android OS smartphone," in
2011 IEEE International Workshop on Medical Measurements and Applications
Proceedings (MeMeA), 2011, pp. 288-293.
[22]
C. Siegel, A. Hochgatterer, and T. E. Dorner, "Contributions of ambient assisted living
for health and quality of life in the elderly and care services--a qualitative analysis
from the experts' perspective of care service professionals," BMC Geriatr, vol. 14, p.
112, 2014. doi: 10.1186/1471-2318-14-112
[23]
J. A. Botia, A. Villa, and J. Palma, "Ambient Assisted Living system for in-home
monitoring of healthy independent elders," Expert Systems with Applications, vol. 39,
pp. 8136-8148, 2012. doi: 10.1016/j.eswa.2012.01.153
[24]
N. M. Garcia, J. J. P. C. Rodrigues, D. C. Elias, and M. S. Dias, Ambient Assisted
Living: Taylor & Francis, 2014
[25]
M. Huch, A. Kameas, J. Maitland, P. J. McCullagh, J. Roberts, A. Sixsmith, et al.,
Handbook of Ambient Assisted Living: Technology for Healthcare, Rehabilitation and
Well-being - Volume 11 of Ambient Intelligence and Smart Environments: IOS Press,
2012
[26]
R. I. Goleva, N. M. Garcia, C. X. Mavromoustakis, C. Dobre, G. Mastorakis, R. Stainov,
et al., "AAL and ELE Platform Architecture," Ambient Assisted Living and Enhanced
Living Environments, pp. 171-209, Elsevier, 2017
[27]
I. M. Pires, N. M. Garcia, N. Pombo, and F. Flórez-Revuelta, "Limitations of the Use of
Mobile Devices and Smart Environments for the Monitoring of Ageing People," in
ICT4AWE 2018 4th International Conference on Information and Communication
Technologies for Ageing Well and e-Health, Madeira, Portugal, 2018.
9
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
[28]
I. M. Pires, N. M. Garcia, N. Pombo, F. Flórez-Revuelta, and N. D. Rodríguez,
"Validation Techniques for Sensor Data in Mobile Health Applications," Journal of
Sensors, vol. 2016, pp. 1687-725, 2016
[29]
I. M. Pires, R. Santos, N. Pombo, N. M. Garcia, F. Florez-Revuelta, S. Spinsante, et
al., "Recognition of Activities of Daily Living Based on Environmental Analyses Using
Audio Fingerprinting Techniques: A Systematic Review," Sensors (Basel), vol. 18, Jan 9
2018. doi: 10.3390/s18010160
[30]
I. M. Pires, N. M. Garcia, N. Pombo, F. Florez-Revuelta, and S. Spinsante, "Approach
for the Development of a Framework for the Identification of Activities of Daily Living
Using Sensors in Mobile Devices," Sensors (Basel), vol. 18, Feb 21 2018. doi:
10.3390/s18020640
[31]
I. M. Pires, N. M. Garcia, N. Pombo, F. Flórez-Revuelta, and S. Spinsante,
"Identification of Activities of Daily Living through Data Fusion on Motion and Magnetic
Sensors embedded on Mobile Devices," in Pervasive and Mobile Computing, vol. 47,
pp. 78-93, 2018. doi: 10.1016/j.pmcj.2018.05.005
[32]
I. M. Pires, M. C. Teixeira, N. Pombo, N. M. Garcia, F. Flórez-Revuelta, S. Spinsante,
et al., "Android library for recognition of activities of daily living: implementation
considerations, challenges, and solutions," in Open Bioinformatics Journal, vol. 11,
pp. 61-88, Bentham Science Publishers B.V.. doi: 10.2174/1875036201811010061
10
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 2
State-of-the-art
This chapter is related to the state-of-the-art and it is composed by three articles, each
presented in its section. These three articles provide a comprehensive review of the different
areas addressed in this research.
1. From
Data
Acquisition
to
Data
Fusion:
A
Comprehensive Review and a Roadmap for the
Identification of Activities of Daily Living Using
Mobile Devices
The following article is the first part of chapter 2.
From Data Acquisition to Data Fusion A Comprehensive Review and a Roadmap for the
Identification of Activities of Daily Living Using Mobile Devices
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo and Francisco Flórez-Revuelta
Sensors (MDPI), published, 2016.
According to 2016 Journal Citation Reports published by Thomson Reuters in 2017, this journal
has the following performance metrics:
ISI Impact Factor (2016): 2.677
ISI Article Influence Score (2016): 0.6
Journal Ranking (2016): 104/642 (Electrical and Electronic Engineering)
11
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
12
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
sensors
Article
From Data Acquisition to Data Fusion: A
Comprehensive Review and a Roadmap for the
Identification of Activities of Daily Living Using
Mobile Devices
Ivan Miguel Pires 1,2,3, *, Nuno M. Garcia 1,3,4 , Nuno Pombo 1,3 and Francisco Flórez-Revuelta 5
1
2
3
4
5
*
Instituto de Telecomunicações, University of Beira Interior, 6201-001 Covilhã, Portugal; ngarcia@it.ubi.pt or
ngarcia@di.ubi.pt (N.M.G.); ngpombo@di.ubi.pt or ngpombo@ubi.pt (N.P.)
Altranportugal, 1990-096 Lisbon, Portugal
ALLab - Assisted Living Computing and Telecommunications Laboratory, Department of Informatics,
University of Beira Interior, 6201-001 Covilhã, Portugal
ECATI, Universidade Lusófona de Humanidades e Tecnologias, 1749-024 Lisbon, Portugal
Faculty of Science, Engineering and Computing, Kingston University, Kingston upon Thames KT1 2EE, UK;
F.Florez@kingston.ac.uk
Correspondence: impires@it.ubi.pt or ivan.pires@altran.com or impires@ubi.pt; Tel.: +351-966-379-785
Academic Editor: Vittorio M. N. Passaro
Received: 21 November 2015; Accepted: 26 January 2016; Published: 2 February 2016
Abstract: This paper focuses on the research on the state of the art for sensor fusion techniques,
applied to the sensors embedded in mobile devices, as a means to help identify the mobile device
user’s daily activities. Sensor data fusion techniques are used to consolidate the data collected from
several sensors, increasing the reliability of the algorithms for the identification of the different
activities. However, mobile devices have several constraints, e.g., low memory, low battery life and
low processing power, and some data fusion techniques are not suited to this scenario. The main
purpose of this paper is to present an overview of the state of the art to identify examples of sensor
data fusion techniques that can be applied to the sensors available in mobile devices aiming to identify
activities of daily living (ADLs).
Keywords: sensor data fusion; accelerometer; data collection; signal processing; sensors signal;
activities of daily living
1. Introduction
Sensors are present in much of the equipment used in everyday life by everyone, including mobile
devices, which are currently applied in ambient assisted living (AAL) systems, such as smartphones,
smartwatches, smart wristbands, tablets and medical sensors. In these devices, sensors are commonly
used to improve and support peoples’ activities or experiences. There are a variety of sensors that
allow the acquisition of the various types of data, which can then be used for different types of tasks.
While sensors may be classified according to the type of data they manage and their application
purposes, the data acquisition is a task that is highly dependent on the user’s environment and
application purpose.
The main objective of this paper is to present a comprehensive review of sensor data fusion
techniques that may be employed with off-the-shelf mobile devices for the recognition of ADLs.
We present a classification of the sensors available in mobile devices and review multi-sensor devices
and data fusion techniques.
Sensors 2016, 16, 184; doi:10.3390/s16020184
www.mdpi.com/journal/sensors
13
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
2 of 27
The identification of ADLs using sensors available in off-the-shelf mobile devices is one of the
most interesting goals for AAL solutions, as this can be used for the monitoring and learning of a
user’s lifestyle. Focusing on off-the-shelf mobile devices, these solutions may improve the user’s
quality of life and health, achieving behavioural changes, such as to reduce smoking or control other
addictive habits. This paper does not comprehend the identification of ADLs in personal health or
well-being, as this application ecosystem is far wider and deserves a more focused research, therefore
being addressed in future work.
AAL has been an important area for research and development due to population ageing and
to the need to solve societal and economic problems that arise with an ageing society. Among other
areas, AAL systems employ technologies for supporting personal health and social care solutions.
These systems mainly focus on elderly people and persons with some type of disability to improve
their quality of life and manage the degree of independent living [1,2]. The pervasive use of mobile
devices that incorporate different sensors, allowing the acquisition of data related to physiological
processes, makes these devices a common choice as AAL systems, not only because the mobile devices
can combine data captured with their sensors with personal information, such as, e.g., the user’s
texting habits or browsing history, but also with other information, such as the user’s location and
environment. These data may be processed either in the device or sent to a server using communication
technologies for later processing [3], requiring a high level of quality of service (QoS) to be needed
to achieve interoperability, usability, accuracy and security [2]. The concept of AAL also includes the
use of sophisticated intelligent sensor networks combined with ubiquitous computing applications
with new concepts, products and services for P4-medicine (preventive, participatory, predictive and
personalized). Holzinger et al. [4] present a new approach using big data to work from smart health
towards the smart hospital concept, with the goal of providing support to health assistants to facilitate
a healthier life, wellness and wellbeing for the overall population.
Sensors are classified into several categories, taking into account different criteria, which include
the environmental analysis and the type of data acquired. The number and type of sensors available
in an off-the-shelf mobile device are limited due to a number of factors, which include the reduced
processing capacity and battery life, size and form design and placement of the mobile device during
the data acquisition. The number and type of available sensors depend on the selected mobile platform,
with variants imposed by the manufacturer, operating system and model. Furthermore, the number
and type of sensors available are different between Android platforms [5] and iOS platforms [6].
Off-the-shelf mobile devices may include an accelerometer, a magnetometer, an ambient air
temperature sensor, a pressure sensor, a light sensor (e.g., photometer), a humidity sensor, an air
pressure sensor (e.g., hygrometer and barometer), a Global Positioning System (GPS) receiver, a gravity
sensor, a gyroscope, a fingerprint sensor, a rotational sensor, an orientation sensor, a microphone, a
digital camera and a proximity sensor. These sensors may be organized into different categories, which
we present in Section 2, defining a new classification of these sensors. Jointly with this classification, the
suitability of the use of these sensors in mobile systems for the recognition of ADLs is also evaluated.
The recognition of ADLs includes a number of different stages, namely data acquisition, data
processing, data imputation, sensor data fusion and data mining, which consist of the application of
machine learning or pattern recognition techniques (Figure 1).
As shown in Figure 1, the process for recognizing ADLs is executed at different stages, which
starts with the data acquisition by the sensors. Afterwards, the data processing is executed, which
includes the validation and/or correction of the acquired data. When data acquisition fails, data
correction procedures should be performed. The correction of the data consists of the estimation of
the missing or inconsistent values with sensor data imputation techniques. Valid data may be sent to
the data fusion module, which consolidates the data collected from several of the available sensors.
After the data fusion, ADLs can be identified using several methods, such as data mining, pattern
recognition or machine learning techniques. Eventually, the system may require the user’s validation
14
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
3 of 27
or feedback, either at an initial stage, or randomly in time, and this validation may be used to improve,
to train and to fine-tune the algorithms that handle the consolidated data.
Figure 1. Schema of a multi-sensor mobile system to recognize activities of daily living.
Due to the processing, memory and battery constraints of mobile devices, the selection of the
algorithms has to be done in a particular way, or otherwise, the inappropriate use of resource-greedy
algorithms will render the solution unusable and non-adoptable by the users.
The remaining sections of this paper are organized as follows: Section 2 proposes a new
classification for the sensors embedded in off-the-shelf mobile devices and identifies different
techniques related to data acquisition, data processing and data imputation; Section 3 is focused on the
review of data fusion techniques for off-the-shelf mobile devices; some applications of mobile sensor
fusion techniques are presented in Section 4; in Section 5, the conclusions of this review are presented.
2. Sensors
Sensors are hardware components that have the capability to capture different types of signals.
Sensors are widely available in equipment used daily, including in smartphones, smartwatches, tablets
and specific devices, including medical and industrial devices, and may be used to collect data in a
plethora of situations. Examples of the use of sensors’ data include measuring some property of the
sensor’s environment, such as chemical sensing, the measurement of motion, touch and proximity
data or acoustic or imaging detection.
2.1. Sensor Classification
In AAL systems, sensors are mainly used to measure data from the user and his/her environment.
For example, the sensors available in most mobile devices allow the detection of movements and
activities of the user, which may lead to identifying the ADLs, with a high accuracy.
The research about sensor data fusion techniques should include the identification of the different
types of sensors available for use. The analysis of the different types of sensors is commonly named
sensor classification, consisting of the definition of different classes to which sensors may be assigned.
Examples of features that define the different sensor classes include the purpose of the sensor, its
working environment and the type of data they acquired [7,8]. The environment itself can be classified
as controlled, uncontrolled, static, dynamic, uncertain and undefined [7,8]. The classification of sensors
presented in this paper has a special focus on the identification of the ADLs.
According to [9], the sensors used for sleep detection include the measurement of different
activities and physical phenomena, such as electro-encephalography (EEG), electro-cardiography
(ECG), blood pressure, photoplethysmography, non-cardiac electropotentials, oxygen saturation,
respiration, body movement, arterial tonometry and temperature. These measurements include the
s
15
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
4 of 27
use of different sensors, such as accelerometer, gyroscope, heart band, microphones and video cameras,
to assess the body’s acceleration and its vital signs while recording audio and video.
For the detection of physical activities, the authors in [10] propose the use of inertial sensors, i.e.,
accelerometers and gyroscopes, complemented with orientation measurement using magnetic sensors,
e.g., a compass and a magnetometer, and location measurement using location sensors, e.g., a GPS
receiver. However, location measurement can also be achieved using inertial sensors and acoustic
sensors, e.g., a microphone [11].
Sensors can be used to perform the measurement of physical activities in controlled or
uncontrolled environments. When used in controlled environments, these sensors, e.g., the video
sensors, may be distributed in different places to measure the movements in a defined and often a
static spatial area [12]. Gentili and Mirchandani [13] make use of counting sensors, image sensors
and automatic vehicle identification (AVI) readers to identify the origin and the location of an object.
The use of static or dynamic images relies heavily on image processing algorithms to extract the
relevant features. Due to the large variety of data that can be extracted from still images or videos,
it is an autonomous research topic. On the other hand, Lee et al. [14] make use of a wearable
electrogoniometer composed of a knee angular and a three-axis accelerometer sensor to detect the
ADLs in uncontrolled environments, presenting also four different machine learning techniques for
detecting occurrences of walking. Other studies were performed in uncontrolled environments using
different mobile technologies. The accuracy of the results is influenced by several constraints also
presented in this paper.
Sensors may be grouped into different categories related to the identification of the movements
and ADLs. One of the usual categories is related to acoustic sensors, which can be used in almost all
environmental scenarios. This category includes simple microphones, silicon microphones [15] and
other acoustic wave devices [16], which are usually used for physical sound wave sensing.
Another proposed category integrates chemical sensors, especially useful for detecting the
presence of specific molecules in the air and/or the environment. In [17], the classification of these
sensors is presented as metal-oxide, semiconductive polymers, conductive electroactive polymers,
optical, surface acoustic wave and electrochemical gas sensors. These types of sensors are also used for
medical purposes, e.g., for the detection of the correct dosage in a patient’s treatment [18].
Another class of sensors includes mechanical sensors, such as mass sensors, strain sensors [19],
pressure sensors, contact sensors, mechanical switches and others. These sensors may be used to detect
movements of the user’s body.
Magnetic sensors may be used to complement the sensing of the different movements.
These include search-coil magnetometers, fluxgate magnetometers, superconductor magnetometers,
hall effect sensors, magnetoresistive magnetometers, spin-valve transistors, giant magnetoimpedance
magnetic sensors, magnetodiodes, magnetotransistors, magnetostrictive magnetometers,
magneto-optical sensors and micro-electro-mechanical systems (MEMS)-based magnetometers [20].
Optical sensors may also be used to measure the parameters of ADLs [21], which include
photoplethysmography, fiber optic sensors, infrared sensors and radio frequency sensors. These sensors
are able to work with adverse temperatures, corrosion/erosion surroundings, high-vibration, voltage
and pressure environments [22], but these are not usual on smartphones.
The detection of physical activities can also be achieved by sensors used in medical applications.
The majority of these sensors can be included in the categories previously mentioned, such as inertial
sensors, e.g., accelerometers, gyroscopes, inclinometers or pedometers, mechanical sensors, e.g.,
ball/tilt/foot switches, sole switches and force-sensitive resistors, acoustic sensors, e.g., microphones,
optical sensors, e.g., infrared sensors, radio frequency sensors, e.g., radio frequency identifiers
(RFID), New Field Communication (NFC) and Ubisense Real-Time Location Systems (RTLS),
atmospheric sensors, e.g., barometers and hygrometers, electrical sensors, e.g., electromyogram (EMG),
ECG, electrodermal activity (EDA) and electrooculography (EOG) sensors, magnetic sensors, e.g.,
magnetometers, photoelectric sensors, e.g., oximeters, and chemical sensors, e.g., actinometers [23].
16
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
5 of 27
Another category of sensors includes the location sensors, such as GPS receiver or WiFi location
methods. Zhang et al. [24] presented a comparison between the accuracy of the use of WiFi location
and the use of a GPS receiver, in which the GPS signal is limited in indoor environments, and in this
case, the use of a WiFi location method is advised.
In [25], wearable sensors are considered. These include the sensors embedded in mobile devices,
such as the GPS receiver, accelerometer, proximity sensor, camera, microphone, gyroscope, light sensor,
gravity sensor and other sensors connected by, for example, Bluetooth (e.g., heart rate sensors, chemical
sensors and others) [26]. Wearable devices of this type are mainly used in the medical area to capture
data that will be used for diagnosis [27,28], e.g., EEG, optical sensors, thermal sensors, acoustic sensors,
magnetic sensors or mechanical sensors.
A variety of sensors is commonly available in off-the-shelf mobile devices, improving the ability
to detect some physiological parameters anywhere and at anytime. Taking previous research into
consideration and having as a goal the selection of sensors that may be used in off-the-shelf devices
to allow the identification of ADLs, a new classification of sensors is proposed with the following
categories: magnetic/mechanical sensors, environmental sensors, motion sensors, imaging/video
sensors, proximity sensors, acoustic sensors, medical sensors, chemical sensors and force sensors.
Table 1 presents a non-exhaustive list for the sensors embedded in mobile devices and external
sensors for each category. In addition, other sensors may be connected to the off-the-shelf mobile
devices using over-the-air technologies, improving the capabilities of the data collection, but as these
are not integrated into the mobile device itself, they are not considered in this review.
Table 1. Sensors classified by categories.
Category
External Sensors
Mobile Sensors
Magnetic/Mechanical Compass; Magnetometer; Strain sensors; Search-coil magnetometer;
sensors
Fluxgate magnetometer; Superconductor magnetometers; Hall effect
sensor; Magnetoresistive magnetometers; Spin-valve transistors; Giant
magnetoimpedance magnetic sensors; Magnetodiode; Magnetotransistor;
Magnetostrictive magnetometers; Magneto-optical sensor; MEMS Based
Magnetometers; Ball/tilt/foot switch; Sole pressure switch; Pressure
sensors; Contact sensors; Mechanical switches
Magnetometer; Compass
Environmental
sensors
Barometer; Humidity; Light sensor; Thermal sensor
Ambient air temperature and pressure; Light
Sensor; Humidity Barometer; Photometer;
Thermal sensor
Location sensors
GPS receiver; Automatic Vehicle Identification (AVI) readers; Ubisense
Real-Time Location Systems (RTLS); Wi-Fi Location-Based Services
GPS receiver; Wi-Fi Location-Based Services
Motion sensors
Accelerometer; Gyroscope; Pressure sensor; Gravity sensor; Inclinometer;
Pedometer; Rotational sensor
Accelerometer; Gravity sensor; Gyroscope;
Rotational vector sensor; Orientation sensor
Imaging/Video
sensors
Digital camera; 3D camera; Optical sensor; Infrared sensor
Digital camera; Infrared sensor
Proximity sensors
Proximity sensor; Touch sensor; RFID; Tactile sensor; NFC
Proximity sensor; Touch sensor; RFID; NFC
Acoustic sensors
Microphone; Silicon microphones; Acoustic wave devices; Surface
acoustic wave
Microphone
Medical sensors
EEG; ECG; EMG; EOG; EDA; Photoplethysmogram; Blood pressure and
arterial tonometry; Respiration; Dosage control/detection; Stress sensors;
Heart rate sensors; electrooculography; electrodermal activity sensors
None
Chemical sensors
Oxygen saturation; Aroma sensors; Metal-oxide; Semi conductive
polymers; Conductive electro active polymers; Electrochemical gas
sensors; Actinometer
None
Optical sensors
Photoplethysmography sensors; Fiber optic sensors; Infrared sensors;
Radio frequency sensors
Infrared sensors; Radio frequency sensors
Force sensors
Force sensitive resistor; Mass sensor; Fingerprint sensor
Fingerprint sensor
Photoelectric
sensors
Oximeter
None
2.2. Data Acquisition
Data acquisition depends on the particular characteristics of the sensors selected to perform the
recognition of ADLs, on the environments where data are captured and also on the architecture of each
17
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
6 of 27
system. The data acquisition process is accomplished by a module embedded in the mobile device and
consists of the measurement and conversion of the electrical signals received by each sensor into a
readable format [29].
Several challenges are associated with the data acquisition process when recognizing ADLs,
including the positioning of the mobile device, the actual data sampling rate and the number of sensors
to be managed [30]. The problems associated with data acquisition in an off-the-shelf mobile device
influence the correct extraction of meaningful features of the ADLs. As the sensors are embedded in the
mobile device, they cannot be located separately in different parts of the body; rather, the mobile device
needs to be situated in an usual and comfortable position. Another issue related to mobile devices is
the power consumption of the data acquisition tasks. Therefore, the limited battery capacity constraint
needs to be taken into account to ensure a successful continuous operation of mobile applications [31]
by developing and applying lightweight acquisition methods.
The main advantage of using mobile devices for data acquisition is related to the possibility to
acquire data anywhere and at anytime [32]. However, it has some limitations in the performance of
the data acquisition for real-time applications. Multitasking execution patterns differ among mobile
devices, because these depend on their processing ability, memory and power capabilities, which
are very limited, and on the operating system and on the number and type of mobile applications
currently installed and/or running.
The Acquisition Cost-Aware QUery Adaptation (ACQUA) framework, presented in [31], consists of
a query processing engine implemented in an off-the-shelf mobile device that dynamically modifies
both the order and the segments of data streams requested from different data sources (sensors).
The ACQUA framework starts by learning the selectivity properties of several sensor streams and then
utilizes such estimated selectivity values to modify the sequence in which the mobile device acquires
data from the sensors. The authors said that it is supported by some basic automated storage and
retrieval system (ASRS) algorithms for acquisition cost-aware query processing. Query specification
can be made using ASRS query evaluation algorithms, the disjunctive normal form or optimizing for
multiple predicates on the same stream. The main features of ACQUA are the accommodation of the
heterogeneity in sensors’ data rates, the determination of the packet sizes and radio characteristics, the
adaptation of the dynamic changes in query selective properties and the support of multiple queries
and heterogeneous time window semantics. The authors claim that it can result in a 70% reduction in
the energy overhead of continuous query processing, by reducing the volume of sensor data that is
sent over-the-air to a server between a mobile device and its attached sensors, without affecting the
fidelity of the processing logic [33].
Lim et al. [31] reviewed different frameworks to optimize data acquisition. The Orchestrator framework
focuses on resource-sharing between multiple context-aware applications executing queries
independently [34,35]. The ErdOS framework views sensors available in off-the-shelf mobile devices
as a shared resource and seeks to rationally distribute the consumption for all resources [36]. The LittleRock
prototype uses a special low-energy coprocessor to decrease the computational energy spent in embedded
processing of on-board sensor data streams [37]. The Jigsaw continuous sensing engine implements a pipelined
stream processing architecture that adaptively triggers different sensors at different sampling rates to fit
the context accuracy required by different applications [38]. The SociableSense framework combines the
cloud-based computation and adaptive sensor sampling to reduce the computational and sensing overheads
during continuous mobile sensing [39]. The BBQ approach builds a multi-dimensional Gaussian probability
density function of the sensors’ likely data values and then uses conditional probabilities to determine, in an
iterative manner, the next sensor whose value is most likely to resolve a given query [40]. These frameworks
try to improve data acquisition and support the preparation of data for their later processing.
Table 2 presents a summary of the data acquisition frameworks, including their features and
limitations. The data can be acquired by different methods, and the correct definition of the position
of the sensors is the main factor that influences the data acquisition process. The definition of the
best positioning for the sensors improves the reliability of the data acquisition methods implemented
18
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
7 of 27
by the mobile operating systems, improving the data acquisition and data processing techniques.
The selection of the best data acquisition methods depends on the purpose of use, the type of data
acquired and their environment [41–44].
Table 2. Examples of data acquisition methods. ACQUA, Acquisition Cost-Aware QUery Adaptation;
ASRS, automated storage and retrieval system.
Method
Features
Limitations
ACQUA
framework
It is a query processing engine implemented on an off-the-shelf mobile
device that dynamically modifies both the order and the segments of
data streams requested from different data sources; supported by some
basic ASRS algorithms for acquisition-cost-aware query processing; seeks
to additionally reduce the communication energy overheads involved
in acquiring the data wirelessly from additional external sensors; it is
complementary to the Jigsaw and LittleRock frameworks.
It does not exploit correlations, which
means that it lacks the predictive power
of representations based on probabilistic
models.
Orchestrator
framework
Used for resource-sharing between multiple context-aware applications
executing queries independently; enables the platform to host
multiple applications stably, exploiting its full resource capacity in a
holistic manner.
Timing relations are not known.
ErdOS framework
Distributes the consumption for all resources of the off-the-shelf mobile
device; restarts the jobs in case of failure.
Statically configured and non-extensible;
difficult to adapt for each case.
LittleRock prototype
Uses a special low-energy coprocessor to decrease the computational
energy spent in embedded processing of on-board sensor data streams;
loads the interactions with sensors and gives the phone’s main processor
and associated circuitry more time to go into sleep mode; flexible storage.
Limited resources of the mobile devices.
Jigsaw
Balances the performance needs of the application and the resource
demands of continuous sensing on the phone; comprises a set of sensing
pipelines for the accelerometer, microphone and GPS sensors.
Robustness inferences.
SociableSense
framework
Combines the cloud-based computation and adaptive sensor sampling
to reduce the computational and sensing overheads during continuous
mobile sensing.
Limited resources of the mobile devices.
BBQ approach
Builds a multi-dimensional Gaussian probability density function of the
sensors’ likely data values and then uses conditional probabilities to
determine, in an iterative manner, the next sensor whose value is most
likely to resolve a given query; similar to ACQUA.
Similar to ACQUA.
2.3. Data Processing
Data processing is the next step of a multi-sensor mobile system used to process the sensor data.
Data processing is a complex process, which also depends on environmental conditions, the types of
sensor, the capabilities of the mobile device used and the types of data collected. The type of application
also influences the use of data processing methods. Data processing may be executed locally, using the
capabilities of the mobile device, or at the server side, sending the collected data to a remote server,
where the data processing methods are executed [42,45,46]. For server-side processing, the mobile
device is only required to acquire the sensors data and to present the results to the user. As frequent
sensor sampling operations and further data processing can significantly reduce the battery lifetime
and the capacities of the mobile device [47,48], several methods may be used to process different types
of data. For example, while for audio processing, the time-scale modification (TSM) algorithm and, for
medical imaging, remote processing methods are highly recommended, for other sensors, data can
be processed locally in the mobile device without significant consumption of local resources [49–51].
Data processing methods may include a segmentation method, which divides a larger data stream into
smaller chunks appropriate for processing, and a definition of the window size [52].
According to Pejovic and Musolesi [53], the challenges of data processing in mobile devices lie in
the adaptation, context-driven operation, computation, storage and communication. The adaptation
and the context-driven operation have several solutions that include the adaptive sampling, the
hierarchical modality switching, the harnessing domain structure and the use of cloud offloading.
Possible solutions for computation, storage and communication are hierarchical processing, cloud
offloading and hardware co-processing.
Several systems and frameworks to classify and to process sensors’ data have been developed
during the last few years. An option is, after contexts are extracted from the collected data, to discard
19
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
8 of 27
part of the data to reduce memory usage [54]. Methods such as support vector machines (SVM),
artificial neural networks (ANN), Bayes classifiers and k-nearest neighbour (KNN) algorithms, among
others, may be used for data processing. In several research studies, the analysis and processing of
sensor data include complex tasks that need high processing and memory capabilities [55]. In such
cases, these activities need to be performed at the server-side.
In [54], Imai et al. present a rule-based data processing engine for sensor nodes to minimize the
amount of energy needed to process the tasks. The data processing may be performed on a computer
located in the neighbourhood of the sensor nodes, which processes the data and provides the results,
obtaining a good balance between reducing the network load and advanced sensor data analysis
and processing.
Yamada et al. [56] present a location-based information delivery method using StreamSpinner, which
achieves efficient stream data processing based on novel multiple continuous query optimization techniques.
In [57], the Dandelion system is presented, achieving good results with senselet, a smartphone-style,
platform-agnostic programming abstraction for in-sensor data processing. Dandelion provides
a unified programming model for heterogeneous systems that span diverse execution contexts,
supporting data-parallel applications.
Dolui et al. [58] defined two architectures: the Device Data Processing Architecture and the Server
Data Processing Architecture. The Device Data Processing Architecture is designed to acquire the
data from the sensors embedded in an off-the-shelf mobile device and to process the data locally.
This architecture is useful when the processing methods require low resources, such as processing the
accelerometer data, proximity sensor data and others. On the contrary, the Server Data Processing
Architecture consists of the dispatch of the data collected to a remote server allowing the computation
of a large amount of data, as well as computations of a complex nature. This architecture is employed
for instance with data acquired with the GPS receiver and the imaging sensors.
Imai et al. [59] defined a method for data processing using the similarity of motions between
observed persons. The method presented implements sensor data processing using a neighbour host,
executed in two phases: a basic action phase and a changing sensor node settings phase [59]. In the
basic action phase, a neighbouring host receives sensor data from the sensor nodes. Then, this host
analyses and processes the data and sends the results to a different host. In the changing sensor node
settings phase, when analytic results of sensor data fulfill the conditions determined in advance, the
neighbouring host instructs sensor nodes to change the settings. Next, the sensor data processing
method based on similarity is defined, where the system acquires and accumulates a newly-observed
person’s acceleration data, while it estimates the related motions using a similar observed person’s
profile. When the system acquires sufficient acceleration data of a newly-observed person, the system
creates a profile that is added to the knowledge base.
The ACQUA framework [31,33] optimizes the data processing, using ASRS algorithms.
The ACQUA framework can also be implemented in systems with a remote processing architecture or
in systems that process the tasks locally.
Reilent et al. [41] developed an open software architecture for patient monitoring, which supports
semantic data processing and (soft) real-time reasoning. Usually, in medical environments, the context
awareness and decision making are performed in a remote server, which returns the results to the
mobile device [42]. Another option used for processing the collected data is the use of a cloud-based
server that processes the data in real time. However, it makes the mobile device and data processing
dependent on a constant Internet connection [60].
In order to avoid this constraint, some tele-medicine systems have implemented data processing
locally on the mobile device. For instance, Postolache et al. [61] implemented advanced data processing,
data management, human computing interfacing and data communication using a smartphone running
the Android operating system.
Augmented reality is also an application relevant for AAL. Paucher and Turk [32] implemented
a system for image processing that used a remote server. The system implements a nearest
20
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
9 of 27
neighbour search of descriptors using a k-dimensional (KD)-Tree structure, applying it for each image.
Other algorithms for image processing using the camera of a mobile device are presented in [62].
Table 3 summarizes the different architectures for data processing, presenting the most relevant
methods and achievements. It is often described in the literature that the data that can be processed
locally on the mobile device include the motion or positioning sensors. However, the processing of
images or videos is computationally expensive, and due to the low memory and processing capabilities
of mobile devices, it is highly recommended that these tasks use external processing.
Table 3. Data processing: architectures and methods.
Architectures
Methods
Achievements
Device Data
Processing
Architecture
Dandelion
system;
SVM;
ANN;
Bayes classifiers;
KNN algorithms;
location-based information delivery
method using StreamSpinner
Acquisition of the data from the sensors embedded in an off-the-shelf
mobile device; process the data locally; the results are rapidly presented
to the user; processing methods should require low resources; using
segmentation methods, a larger data stream is divided into smaller chunks
improving the methods; the correct definition of the window size is
important for achieve good results.
Server Data
Processing
Architecture
SVM; ANN; Bayes classifiers; KNN
algorithms; nearest neighbour search of
descriptors using a KD-Tree structure.
Dispatching of the data collected to a remote server allowing the
computation of a large amount of data, as well as computations of
complex nature; in some cases, the data processing may be performed on a
computer located in the neighbourhood of the sensor nodes; in server-side
processing, the mobile device and data processing are dependent on a
constant Internet connection.
2.4. Data Imputation
Data acquisition with sensors embedded in off-the-shelf mobile devices for real-time recognition
of ADLs may fail in many circumstances. Firstly, the acquisition process fails when the sensors report
unexpected values. Secondly, due to the multitasking nature of many of the mobile devices, this can
happen anytime and for any given collection scenario.
Depending on the timing of the missing data included in a subset of sensors’ data, the missing
data are classified as missing completely at random (MCAR) when the missing values are randomly
distributed by all time instants [63,64]. The missing data are classified as missing at random (MAR)
when the missing values are randomly distributed by subsets of the collected data [63,64]. Additionally,
the missing data are classified as missing not at random (MNAR) when the missing values are not
randomly distributed [63,64].
Several methods to minimize the effects of missing data have been developed, estimating the
missing values based either on other values correctly obtained or on external factors. In [63], the
Imputation Tree (ITree) method is presented, which is a tree-based algorithm for missing values
imputation. This method constructs a missing pattern tree (MPT), which is a binary classification tree
for identifying the absence of each observation. It uses clustering techniques, e.g., K-means clustering,
to impute missing values and linear regression analysis to improve data imputation.
In [65], a multi-matrices factorization model (MMF) for the missing sensor data estimation
problem is formulated, which uses statistical methods to estimate the missing values. There are other
methods for sensor data imputation that employ KNN-based imputation [66], multiple imputation [67],
hot/cold imputation [68], maximum likelihood and Bayesian estimation [69] and expectation
maximization [70]. In general, these methods are used to verify and increase the consistency of
sensor data.
The KNN method and its variants, such as KNNimpute (K-nearest neighbour imputation),
SKNNimpute (sequential K-nearest neighbour method-based imputation) and MKNNimpute
(K-nearest neighbour imputation method based on Mahalanobis distance), are the most used
methods [71]. Other cluster-based imputation methods exists, such as KMI (K-means-based
imputation) [72] and FCMimpute (fuzzy C-means clustering imputation) [73].
For audio signal imputation purposes, Smaragdis et al. [74] used nearest neighbours and singular
value decomposition (SVD) algorithms. The SVD algorithm is executed after nearest neighbour clustering,
replacing all of the missing values with an initial value, computing the SVD of the resulting matrix and
21
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
10 of 27
replacing the missing values with their prediction according to the SVD decomposition, repeating the
process until the change in the imputed missing data falls below some user-defined threshold.
In [75], a generalized trend diffusion (GTD) method is used to create small datasets. It applies
the so-called shadow data and membership functions for increasing the knowledge when using back
propagate-based (BP) neural networks to predict the missing values.
In [76], random recursive partitioning (RRP) is used to generate a proximity matrix used in
non-parametric matching problems, such as hot-deck missing data imputation and average treatment
effect estimation. In [77], a discrete model is presented to improve data imputation at the time of
collection, reducing the errors and repairing the data if possible.
Other approaches [64] have also been studied to improve data imputation, such as ignoring and
deleting data (e.g., listwise deletion and discarding instances), available-case methods (e.g., pairwise
deletion), non-model based imputation procedures (e.g., unconditional mean imputation, conditional
mean imputation) and model-based imputation procedures. Other implicit and explicit models exist
for data imputation. The implicit models are based on implicit assumptions on the proximity between
individuals belonging to the dataset (e.g., hot deck imputation, cold deck imputation, substitution
method, composite methods), and the explicit models are based on a statistical model to describe the
predictive distribution of missing data (e.g., linear regression, logistic regression, multiple imputation
methods, the expectation-maximization (EM) algorithm, probabilistic neural networks, fuzzy min–max
neural networks, general regression auto associative neural network and distribution free methods,
such as non-parametric regression and tree-based methods).
The use of big data brings other challenges in data processing that consist of the extraction
of relevant information to enable the correct analysis, discovery and interpretation of the data.
Pombo et al. [55] presented a predictive model using the radial basis function neural network (RBFNN)
combined with a filtering technique aiming at the estimation of the electrocardiogram (ECG) waveform,
which supports healthcare professionals on clinical decisions and practices. The study [55] is related to
the PhD thesis, presented in [78], that starts with the analysis of several methods for machine learning
prediction and finalizes with the creation of a clinical decision support system. The methods studied
are the rule-based algorithms (RBA), the ANN, the rough and fuzzy sets (RFS) and the statistical
learning algorithms (SLA).
The data imputation methods can be applied in off-the-shelf mobile devices with some technical
limitations, due to the capacities of these devices. For instance, in [79], neural networks are used to
perform data imputation for a visual system.
Table 4 presents a summary about the data imputation methods analysed on this paper, presenting
the types of data acquired and achievements obtained.
Table 4. Examples of data imputation methods.
22
Types of Data
Models
Achievements
MCAR
listwise deletion; pairwise deletion; ITree method; KNN method
and their variants; KMI; FCMimpute; SVD; GTD method; BP
neural networks; RRP; MMF; MPT; hot/cold imputation; expectation
maximization; Bayesian estimation; unconditional mean imputation;
conditional mean imputation; ANN.
MAR
maximum likelihood; multiple imputation; ITree method; KNN
method and their variants; KMI; FCMimpute; SVD; GTD method; BP
neural networks; RRP; MMF; MPT; hot/cold imputation; expectation
maximization; Bayesian estimation; unconditional mean imputation;
conditional mean imputation; ANN.
These methods improve the identification of
the absence of each observation; the use of
clustering techniques and linear regression
analysis improves the data imputation; the
data imputation also increases the consistency
of the data; other improvements of data
imputation are ignoring and deleting the
unusable data; another measured variable can
be indirectly predicted with the probability of
missingness.
MNAR
selection model; pattern mixture models; maximum likelihood;
multiple imputation; ITree method; KNN method and their
variants; KMI; FCMimpute; SVD; GTD method; BP neural networks;
RRP; MMF; MPT; hot/cold imputation; expectation maximization;
Bayesian estimation; unconditional mean imputation; conditional
mean imputation; ANN.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
11 of 27
However, some devices do not have enough capabilities to perform complex data imputation
methods. In order to solve this, Kim and Prabhakaran [80] presented a new method, named canonical
correlation based k-weighted angular similarity (CkWAS), that maps the missing data with a reference
pattern dataset to assign values to the missing data.
3. Sensor Data Fusion in Mobile Devices
Data fusion is a critical step in the integration of the data collected by multiple sensors. The main
objective of the data fusion process is to increase the reliability of the decision that needs to be made
using the data collected from the sensors, e.g., to increase the reliability of the identification of the
ADL algorithm running in an off-the-shelf mobile device. If a single stream of data cannot eliminate
uncertainty from the output, data fusion will use data from several sources with the goal of decreasing
the uncertainty level of the output. Consequently, the data fusion increases the level of robustness of a
system for the recognition of ADLs, reducing the effects of incorrect data captured by the sensors [81]
or helping to compute solutions when the collected data are not usable for a specific task.
A mobile application implemented by Ma et al. [82] was tested in a Google Nexus 4 and uses the
accelerometer, gyroscope, magnetometer and GPS receiver to evaluate the sensors’ accuracy, precision,
maximum sampling frequency, sampling period, jitter and energy consumption in all of the sensors.
The test results show that the built-in accelerometer and gyroscope sensor data have a standard
deviation of approximately 0.1 to 0.8 units between the measured value and the real value, the compass
sensor data deviate approximately three degrees in the normal sampling rate, and the GPS receiver
data have a deviation lower than 10 meters. Thus, one of the working hypotheses of the research is the
data collected by mobile sensors may be fused to work with more precision towards a common goal.
The data fusion may be performed with mobile applications, accessing the sensors data as a
background process, processing the data and showing the results in a readable format or passing the
results or the data to a central repository or central processing machine for further processing.
Durrant-Whyte et al. [83] described a decentralized data fusion system, which consists of a
network of sensor nodes, each one with its own processing facility. This is a distributed system that
does not require any central fusion or central communication facility, using Kalman filters to perform
the data fusion. Other decentralized systems for data fusion have also been developed, improving
some techniques and the sensors used [84].
The definition of the categories of the data fusion methods has already been been discussed by
several authors [85–87]. According to these authors, the data fusion methods may be categorized as
probabilistic, statistic, knowledge base theory and evidence reasoning methods. Firstly, probabilistic
methods include Bayesian analysis of sensor values with Bayesian networks, state-space models,
maximum likelihood methods, possibility theory, evidential reasoning and, more specifically, evidence
theory, KNN and least square-based estimation methods, e.g., Kalman filtering, optimal theory,
regularization and uncertainty ellipsoids. Secondly, statistic methods include the cross-covariance,
covariance intersection and other robust statistics. Thirdly, knowledge base theory methods include
intelligent aggregation methods, such as ANN, genetic algorithms and fuzzy logic. Finally, the
evidence reasoning methods include Dempster-Shafer, evidence theory and recursive operators.
Depending on the research purpose of the data fusion, these methods have advantages and
disadvantages presented in Table 5. The data fusion methods are influenced with the constraints in
the previous execution of data acquisition, data processing and data imputation. The advantages and
disadvantages also depend on the environmental scenarios and the choice of the correct method to
apply for each research scenario.
23
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
12 of 27
Table 5. Advantages and disadvantages of the sensor data fusion methods.
Methods
Advantages
Disadvantages
Probabilistic
methods
Provide methods for model estimation; allows unsupervised
classification; estimate the state of variables; reduce errors
in the fused location estimate; increase the amount of data
without changing its structure or the algorithm; produce
a fused covariance matrix that better reflects the expected
location error.
Require a priori probabilistic knowledge of information
that is not always available or realistic; classification
depends on the starting point; unsuitable for
large-scale systems; requires a priori knowledge of
the uncertainties’ co-variance matrices related to the
system model and its measurements.
Statistic methods
Accuracy improves from the reduction of the prediction
error; high accuracy compared with other local estimators;
robust with respect to unknown cross-covariance.
Complex and difficult computation is required to
obtain the cross-variance; complexity and larger
computational burden.
Knowledge base
theory methods
Allows the inclusion of uncertainty and imprecision; easy to
implement; learning ability; robust to noisy data and able to
represent complex functions.
The knowledge extraction requires the intervention of
human expertise (e.g., physicians), which takes time
and/or may give rise to interpretation bias; difficulty
in determining the adequate size of the hidden layer;
inability to explain decisions; lack of transparency of
data.
Evidence
reasoning
methods
Assign a degree of uncertainty to each source.
Require assigning a degree of evidence to all concepts.
In [88], data fusion methods are distributed in six categories. The categories of the data fusion
methods include data in–data out, data in–feature out, feature in–feature out, feature in–decision out,
decision in–decision out and data in–decision out.
Performing the data fusion process in real time can be difficult because of the large amount
of data that may need to be fused. Ko et al. [88] proposed a framework, which used dynamic time
warping (DTW), as the core recognizer to perform online temporal fusion on either the raw data or the
features. DTW is a general time alignment and similarity measure for two temporal sequences. When
compared to hidden Markov models (HMMs), the training and recognition procedures in DTW are
potentially much simpler and faster, having a capability to perform online temporal fusion efficiently
and accurately in real time.
The most used method for data fusion is the Kalman filter, developed for linear systems and
then improved to a dynamically-weighted recursive least-squares algorithm [89]. However, as the
sensor data are not linear, the authors in [89] used the extended Kalman filter to linearize the system
dynamics and the measurement function around the expected state and then applied the Kalman filter
as usual. A three-axis magnetometer and a three-axis accelerometer are used for the estimation of
several movements [89].
Other systems employ variants of the Kalman filter to reduce the noise and improve the detection
of movements. Zhao et al. [90] use the Rao-Blackwellization unscented Kalman filter (RBUKF) to fuse
the sensor data of a GPS receiver, one gyroscope and one compass to improve the precision of the
localization. The authors compare the RBUKF algorithm to the extended Kalman filter (EKF) and
unscented Kalman filter (UKF), stating that the RBUKF algorithm improves the tracking accuracy and
reduces computational complexity.
Walter et al. [91] created a system for car navigation by fusing sensor data on an Android
smartphone, using the embedded sensors (i.e., gyroscope) and data from the car (i.e., speed
information) to support navigation via GPS. The developed system employs a controller area network
(CAN)-bus-to-Bluetooth adapter to establish a wireless connection between the smartphone and the
CAN-bus of the car. The mobile application fuses the sensors’ data and implements a strap down
algorithm and an error-state Kalman filter with good accuracy, according to the authors of [91].
Anther application for location inference was built by using the CASanDRA mobile OSGi (Open
Services Gateway Initiative) framework using a LocationFusion enabler that fused the data acquired by
all of the available sensors (i.e., GPS, Bluetooth and WiFi) [92].
Mobile devices allow the development of context-aware applications, and these applications after
use a framework for context information management. In [93], a mobile device-oriented framework
for context information management to solve the problems on context shortage and communication
24
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
13 of 27
inefficiency is proposed. The main functional components of this framework are the data collector, the
context processor, the context manager and the local context consumer. The data collector acquires the
data from internal or external sensors. Then, the context processor and the context manager extract
and manage context. Finally, the local context consumer develops context-aware applications and
provides appropriate services to the mobile users. The authors claim that this framework is able to
run real-time applications with quick data access, power efficiency, personal privacy protection, data
fusion of internal and external sensors and simplicity in usage.
Blum et al. [94] use the GPS receiver, compass and gyroscope embedded in Apple iPhone 4 (iOS)
(Apple, Cupertino, CA, USA), iPhone 4s (iOS) (Apple, Cupertino, CA, USA) and Samsung Galaxy
Nexus (Android) (Samsung, Seul, Korea), for the measurement of the location and augmented reality
situations. Blum et al. analysed the position of the smartphone during the data collection, testing
in three different orientation/body position combinations and in varying environmental conditions,
obtaining results with location errors of 10 to 30 m (with a GPS receiver) and compass errors around
10 to 30°, with high standard deviations for both.
In [95], the problems of data fusion, placement and positioning of fixed and mobile sensors are
focused on, presenting a two-tiered model. The algorithm combines the data collected by fixed sensors
and mobile sensors, obtaining good results. The positioning of the sensors is very important, and
the study implements various optimization models, ending with the creation of a new model for the
positioning of sensors depending on their types. Other models are also studied by the authors, which
consider the simultaneous deployment of different types of sensors, making use of more detailed
sensor readings and allowing for dependencies among sensor readings.
Haala and Böhm [96] created a low-cost system using an off-the-shelf mobile device with several
sensors embedded. The data collected by a GPS receiver, a digital compass and a 3D CAD model of a
region are used for provisioning data related to urban environments, detecting the exact location of a
building in a captured image and the orientation of the image.
A different application where data fusion is required is the recognition of physical activity.
The most commonly-used sensors for the recognition of physical activity include the accelerometer,
the gyroscope and the magnetic sensor. In [26], data acquired with these sensors are fused with a
combined algorithm, composed of Fisher’s discriminant ratio criterion and the J3 criterion for feature
selection [97]. The collection of data related to the physical activity performed is very relevant to
analyse the lifestyle and physiological characteristics of people [98].
Some other applications analyse the user’s lifestyle by fusing the data collected by the sensors
embedded in mobile devices. Yi et al. [99] presented a system architecture and a design flow for remote
user physiological data and movement detection using wearable sensor data fusion.
In the medical scope, Stopczynski et al. [27] combined low-cost wireless EEG sensors with
smartphone sensors, creating 3D EEG imaging, describing the activity of the brain. Glenn and
Monteith [100] implemented several algorithms for the analysis of mental status, by using smartphone
sensors. These algorithms fuse the data acquired from all of the sensors with other data obtained over
Internet, to minimize the constraints of mobile sensors data.
The system proposed in [101] makes use of biosensors for different measurements, such
as a surface plasmon resonance (SPR) biosensor, a smart implantable biosensor, a carbon
nanotube (CNT)-based biosensor, textile sensors and enzyme-based biosensors, combined with
smartphone-embedded sensors and applying various techniques for the fusion of the data and to filter
the inputs to reduce the effects of the position of the device.
Social networks are widely used and promote context-aware applications, as data can be collected
with the user’s mobile device sensors to promote the adaptation of the mobile applications to the
user-specific lifestyle [102]. Context-aware applications are an important research topic; Andò [103]
proposed an architecture to adapt context-aware applications to a real environment, using position,
inertial and environmental sensors (e.g., temperature, humidity, gases leakage or smoke). The authors
25
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
14 of 27
claim this system is very useful for the management of hazardous situations and also to supervise
various physical activities, fusing the data in a specific architecture.
Other authors [104] have been studying this topic with mobile web browsers, using the
accelerometer data and the positional data, implementing techniques to identify different ADLs.
These systems capture the data while the user is accessing the Internet from a mobile device, analysing
the movements and the distance travelled during that time and classifying the ADLs performed.
Steed and Julier [105] created the concept of behaviour–aware sensor fusion that uses the redirected
pointing technique and the yaw fix technique to increase the usability and speed of interaction in
exemplar mixed-reality interaction tasks.
In [106], vision sensors (e.g., a camera) are used and combined with accelerometer data to apply
the depth from focus (DFF) method, which was limited to high precision camera systems for the
detection of movements in augmented reality systems. The vision systems are relevant, as the obtained
information can identify more accurately a moving object. Some mobile devices integrate RFID readers
or cameras to fetch related information about objects and initiate further actions.
Rahman et al. [107] use a spatial-geometric approach for interacting with indoor physical objects
and artefacts instead of RFID-based solutions. It uses a fusion between the data captured by an infrared
(IR) camera and accelerometer data, where the IR cameras are used to calculate the 3D position of the
mobile phone users, and the accelerometer in the phone provides its tilting and orientation information.
For the detection of movement, they used geometrical methods, improving the detection of objects in a
defined space.
Grunerbl et al. [108] report fusing data from acoustic, accelerometer and GPS sensors.
The extraction of features from acoustic sensors uses low-level descriptors, such as root-mean-square
(RMS) frame energy, mel-frequency cepstral coefficients (MFCC), pitch frequency, harmonic-to-noise
ratio (HNR) and zero-crossing-rate (ZCR). They apply the naive Bayes classifier and other pattern
recognition methods and report good results in several situations of daily life.
Gil et al. [109] present other systems to perform sensor fusion, such as LifeMap, which is a
smartphone-based context provider for location-based services, as well as the Joint Directors of
Laboratories (JDL) model and waterfall IF model, which define various levels of abstraction for the
sensor fusion techniques. Gil et al. also present the inContexto system, which makes use of embedded
sensors, such as the accelerometer, digital compass, gyroscope, GPS, microphone and camera, applying
several filters and techniques to recognize physical actions performed by users, such as walking,
running, standing and sitting, and also to retrieve context information from the user.
In [110], a sensor fusion-based wireless walking-in-place (WIP) interaction technique is presented,
creating a human walking detection algorithm based on fusing data from both the acceleration and
magnetic sensors integrated in a smartphone. The proposed algorithm handles a possible data
loss and random delay in the wireless communication environment, resulting in reduced wireless
communication load and computation overhead. The algorithm was implemented for mobile devices
equipped with magnetic, accelerometer and rotation (gyroscope) sensors. During the tests, the
smartphone was adapted to the user’s leg. After some tests, the authors decided to implement the
algorithm with two smartphones and a magnetometer positioned on the user’s body, combining
the magnetic sensor-based walking-in-place and acceleration-based walking in place, in order to
discard the use of a specific model (e.g., gait model) and a historical data accumulated. However, the
acceleration-based technique does not support the correct use in slow-speed walking, and the magnetic
sensor-based technique does not support both normal and fast walking speeds.
Related to the analysis of walking, activities and movements, several studies have employed
the GPS receiver, combined with other sensors, such as the accelerometry, magnetometer, rotation
sensors and others, with good accuracy using some fusion algorithms, such as the naive and oracle
methods [111], machine learning methods and kinematic models [112] and other models adapted to
mobile devices [91]. The framework implemented by Tsai et al. [113] detects the physical activity fusing
data from several sensors embedded in the mobile device and using crowdsourcing, based on the
26
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
15 of 27
methods’ result of the merging of classical statistical detection and estimation theory and uses value
fusion and decision fusion of human sensor data and physical sensor data.
In [114], Lee and Chung created a system with a data fusion approach based on several
discrete data types (e.g., eye features, bio-signal variation, in-vehicle temperature or vehicle speed)
implemented in an Android smartphone, allowing high resolution and flexibility. The study
involves different sensors, including video, ECG, photoplethysmography, temperature and a
three-axis accelerometer, which are assigned as input sources to an inference analysis framework.
A fuzzy Bayesian network includes the eyes feature, bio-signal extraction and the feature
measurement method.
In [81], a sensor weighted network classifier (SWNC) model is proposed, which is composed of
three classification levels. A set of binary activity classifiers consists of the first classification level of
the model proposed. The second classification level is defined by node classifiers, which are decision
making models. The decisions of the model are combined through a class-dependent weighting fusion
scheme structure with a structure defined through several base classifiers. The weighted decisions
obtained by each node classifier are fused on the model proposed. Finally, the last classification level
has a similar constitution of the second classification level. Independently of the level of noise imposed,
when the mobile device stays in a static position during the data acquisition process, a performance up
to 60% can be achieved with the proposed method.
Chen et al. [115] and Sashima et al. [116] created client-server platforms that monitor the activities
in a space with a constant connection with a server. This causes major energy consumption and
decreases the capabilities of the mobile devices in the detection of the activities performed.
Another important measurement with smart phone sensors is related to the orientation of the
mobile device. Ayub et al. [117] have already implemented the DNRF (drift and noise removal filter)
with a sensor fusion of gyroscope, magnetometer and accelerometer data to minimize the drift and
noise in the output orientation.
Other systems have been implemented for gesture recognition. Zhu et al. [118] proposed a
high-accuracy human gesture recognition system based on multiple motion sensor fusion. The method
reduces the energy overhead resulting from frequent sensor sampling and data processing with
a high energy-efficient very large-scale integration (VLSI) architecture. The results obtained have
an average accuracy for 10 gestures of 93.98% for the user-independent case and 96.14% for the
user-dependent case.
As presented in Section 2.4, decentralized systems, cloud-based systems or server-side systems
are used for data processing. As the data fusion is the next stage, van de Ven et al. [119] presented the
Complete ambient assisting living experiment (CAALYX) system that provides continuous monitoring
of people’s health. It has software installed on the mobile phone that uses data fusion for decision
support to trigger additional measurements, classify health conditions or schedule future observations.
In [120], Chen et al. presented a decentralized data fusion and active sensing (D2 FAS) algorithm
for mobile sensors to actively explore the road network to gather and assimilate the most informative
data for predicting the traffic phenomenon.
Zhao et al. [121,122] developed the COUPON (Cooperative Framework for Building Sensing Maps
in Mobile Opportunistic Networks) framework, a novel cooperative sensing and data forwarding
framework to build sensing maps satisfying specific sensing quality with low delay and energy
consumption. This framework implements two cooperative forwarding schemes by leveraging data
fusion; these are epidemic routing with fusion (ERF) and binary spray-and-wait with fusion (BSWF).
It considers that packets are spatial-temporally correlated in the forwarding process and derives the
dissemination law of correlated packets. The work demonstrates that the cooperative sensing scheme
can reduce the number of samplings by 93% compared to the non-cooperative scheme; ERF can reduce
the transmission overhead by 78% compared to epidemic routing (ER); BSWF can increase the delivery
ratio by 16% and reduce the delivery delay and transmission overhead by 5% and 32%, respectively,
compared to binary spray-and-wait (BSW).
27
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
16 of 27
In [123], a new type of sensor node is described with modular and reconfigurable characteristics,
composed of a main board, with a processor, FR (Frequency Response) circuits and a power supply, as
well as an expansion board. A software component was created to join all sensor data, proposing an
adaptive processing mechanism.
In [124], the C-SPINE (Collaborative-Signal Processing in Node Environment) framework
is proposed, which uses multi-sensor data fusion among CBSNs (Collaborative Body Sensor
Networks) to enable joint data analysis, such as filtering, time-dependent data integration and
classification, based on a multi-sensor data fusion schema to perform automatic detection of
handshakes between two individuals and to capture possible heart rate-based emotional reactions due
to individuals meeting.
A wireless, wearable, multi-sensor system for locomotion mode recognition is described in [125],
with three inertial measurement units and eight force sensors, measuring both kinematic and dynamic
signals of the human gait. The system uses a linear discriminant analysis classifier, obtaining
good results for motion mode recognition during the stance phase, during the swing phase and
for sit-to-stand transition recognition.
Chen [126] developed an algorithm for data fusion to track both non-manoeuvring and
manoeuvring targets with mobile sensors deployed in a wireless sensor network (WSN). It applies
the GATING technique to solve the problem of mobile-sensor data fusion tracking (MSDF) for targets.
In WSNs, an adaptive filter (Kalman filter) is also used, consisting of a data association technique
denoted as one-step conditional maximum likelihood.
Other studies have focused on sensor fusion for indoor navigation [127,128]. Saeedi et al. [127]
proposes a context-aware personal navigation system (PNS) for outdoor personal navigation using
a smartphone. It uses low-cost sensors in a multi-level fusion scheme to improve the accuracy and
robustness of the context-aware navigation system [127]. The system developed has several challenges,
such as context acquisition, context understanding and context-aware application adaptation, and it is
mainly used for the recognition of the people’s activities [127]. It uses the accelerometer, gyroscope
and magnetometer sensors and the GPS receiver available on the off-the-shelf mobile devices to
detect and recognize the motion of the mobile device, the orientation of the mobile device and the
location and context of the data acquisition [127]. The system includes a feature-level fusion scheme to
recognize context information, which is applied after the data are processed and the signal’s features
are extracted [127]. Bhuiyan et al. [128] evaluate the performance of several methods for multi-sensor
data fusion focused on a Bayesian framework. A Bayesian framework consists of two steps [128],
such as prediction and correction. In the prediction stage, the current state is updated based on the
previous state and the system dynamics [128]. In the correction stage, the prediction is updated with
the new measurements [128]. In [128], the authors studied different combinations of methods for data
fusion, such as a linear system and Kalman filter and a non-linear system and extended Kalman filter,
implementing some sensor data fusion systems with good results.
In [129], a light, high-level fusion algorithm to detect the daily activities that an individual
performs is presented. The proposed algorithm is designed to allow the implementation of a
context-aware application installed on a mobile device, working with minimum computational cost.
The quality of the estimation of the ADLs depends on the presence of biometric information and the
position and number of available inertial sensors. The best estimation for continuous physical activities
obtained, with the proposed algorithm, is approximately 90%.
The CHRONIOUS system is developed for smart devices as a decision support system, integrating
a classification system with two parallel classifiers. It combines an expert system (rule-based system)
and a supervised classifier, such as SVM, random forests, ANN (e.g., the multi-layer perceptron),
decision trees and naive Bayes [130]. Other systems for the recognition of ADLs have also been
implemented using other classifiers during the data fusion process, e.g., decision tables.
Martín et al. [131] evaluated the accuracy, computational costs and memory fingerprints in the
classifiers mentioned above working with different sensor data and different optimization. The system
28
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
17 of 27
that implements these classifiers encompasses different sensors, such as acceleration, gravity, linear
acceleration, magnetometer and gyroscope, with good results, as presented in [131]. Other systems
were implemented using lightweight methods for sensor data fusion, e.g., the KNOWME system [132],
which implements the autoregressive-correlated Gaussian model for data classification and data fusion.
Other research studies focused on the compensation of the ego-motion of the camera, carried out
with the data fusion of Viola and Jones face detector and inertial sensors, reporting good results [133].
In [134], sensor fusion is used for error detection, implementing the fusion of ECG with the blood
pressure signal, blood pressure with body temperature and acceleration data with ECG.
In [135], Jin et al. proposed a robust dead-reckoning (DR) pedestrian tracking system to be used
with off-the-shelf sensors. It implements a robust tracking task as a generalized maximum a posteriori
sensor fusion problem, and then, they narrow it to a simple computation procedure with certain
assumptions, with a reduction in average tracking error up to 73.7%, compared to traditional DR
tracking methods.
Table 6. Examples of sensor data fusion methods.
Sensors
Methods
Achievements
Accelerometer;
Gyroscope;
Magnetometer;
Compass;
GPS receiver;
Bluetooth;
Wi-Fi;
digital
camera;
microphone; RFID readers;
IR camera.
DR
pedestrian
tracking
system;
Autoregressive-Correlated Gaussian Model;
CASanDRA mobile OSGi framework; Genetic
Algorithms; Fuzzy Logic; Dempster-Shafer;
Evidence Theory; Recursive Operators; DTW
framework; CHRONIOUS system; SVM;
Random Forests; ANN; Decision Trees; Naive
Bayes; Decision Tables; Bayesian analysis.
The use of several sensors reduces the noise effects;
these methods also evaluated the accuracy of sensor data
fusion; the data fusion may be performed with mobile
applications, accessing the sensors data as a background
process, processing the data and showing the results in a
readable format.
Accelerometer;
Gyroscope;
Magnetometer;
Compass;
GPS receiver;
Bluetooth;
WiFi;
digital
camera;
microphone;
low-cost
wireless EEG sensors; RFID
readers; IR camera
Kalman Filtering; C-SPINE framework; DNRF
method; SWNC model; GATING technique;
COUPON framework; CAALYX system; high
energy-efficient very large-scale integration
(VLSI) architecture;
sensor-fusion-based
wireless walking-in-place (WIP) interaction
technique; J3 criterion; DFF method; ERF
method; BSWF method; inContexto system;
RMS frame energy; MFCC method; pitch
frequency; HNR method; ZCR method;
KNN; Least squares-based estimation
methods; Optimal Theory, Regularization;
Uncertainty Ellipsoids.
These methods allow a complex processing of amount
of data acquired, because it is central processed in
a server-side system; using data from several sources
decreases the uncertainty level of the output; performing
the data fusion process in real time can be difficult
because of the large amount of data that may need to
be fused; the data fusion may be performed with mobile
applications, accessing the sensors data as a background
process, processing the data and showing the results in
a readable format or passing the results or the data to
a central repository or central processing machine for
further processing.
Gyroscope;
Compass;
Magnetometer; GPS receiver.
Kalman Filtering; Bayesian analysis.
It is mainly useful for the context-aware localization
systems; defined several recognizer algorithms to
perform online temporal fusion on either the raw data or
the features.
ECG and others
Kalman Filtering.
Using data from several sources decreases the uncertainty
level of the output; defined several recognizer algorithms
to perform online temporal fusion on either the raw data
or the features.
In [136], Grunerbl et al. developed a smartphone-based recognition of states and state changes
in bipolar disorder patients, implemented an optimized state change detection, developing various
fusion methods with different strategies, such as logical AND, OR, and their own weighted fusion,
obtaining results with good accuracy.
In Table 6, a summary of sensor data fusion techniques, their achievements and sensors used is
presented. This table also helps make clear that different methods can be applied to different types of
data collected by different sensors.
In conclusion, sensor fusion techniques for mobile devices are similar to those employed with
other external sensors, because the applications usually involve embedded and external sensors at the
same time. The Kalman filter is the most commonly used, but on these devices, there is a need to use
low processing techniques. Some research applications need a high processing capacity, and in this
case, the mobile device is only used to capture the data. After capturing the data, the captured data
will be sent to a server for later processing. The most important criterion for choosing a data fusion
29
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
18 of 27
method should be based on the limited capacities of the mobile devices, as these devices increase the
possibility to collect physiological data anywhere and at anytime with low cost.
4. Application in Ambient Assisted Living
As described in this paper, the use of data fusion techniques with sensors available in off-the-shelf
mobile devices may be shaped to form different types of systems and architectures. These techniques
are implementable in systems based on mobile technologies, with local processing, server-side
processing or a mix of these, or as distributed systems that work as a cooperative system collecting the
data with different mobile devices.
Mobile systems for the recognition of ADLs have two possible architectures, are where all
processing activities are performed locally on the mobile device or another where the data processing
will be executed totally, or in part, on a remote machine. Several authors have developed algorithms
related to context-aware sensing (e.g., environment recognition, gestural recognition and browsing
history) [88,93,104,109,115], positioning of objects [107] and localization sensing [82,90,91,94,137].
Several works have focused on the development of different methods to recognize some
ADLs [92,111,112,129,132,138,139], such as sitting, walking, running, cycling, standing still, climbing
stairs, using the lift, getting up, falling and having a meeting. In addition, off-the-shelf mobile
devices have several sensors embedded that are capable of recognizing psychological/emotional states,
including the recognition of mental stress [140].
The recognition of the ADLs [81,110,131,141,142] is very important in systems developed
to support and monitor elderly people or people with some disease or impairment [143,144].
When combined with the user’s agenda, they can be employed for monitoring and learning lifestyles
and physical exercise, helping people in emergency situations [116,136] and other situations where the
sensors can improve life quality, e.g., driving monitoring [114].
Healthcare is one of the most important purposes of the recognition of the ADLs, and commonly,
it is performed in smart health environments [4,145], with several technologies that include the relation
between several mobile devices capturing the sensors’ data connected by WSNs and using sensor webs
to aggregate the data collected by several sensors and identify the ADLs with more accuracy [146].
Distributed systems for sensor data fusion with off-the-shelf mobile devices have been used for
tracking [120–122,126,147], location detection [83,84,123,124,148,149], health monitoring [99,103,119],
user monitoring [113,150], automotive systems [151] and other purposes.
In [152], several types of sensors (e.g., complementing the detection of objects using a camera and
inertial sensors) are used to classify objects and detect changes. Using mobile devices to capture sensors’
signals, a problem with the involuntary movements of the mobile device during the performance of
some ADLs is very high. However, the use of data fusion technologies allows one to minimize the
effects of the possible involuntary movements of the mobile device, commonly considered noise [117].
In conclusion, the sensor fusion techniques are very useful to improve the recognition of the
ADLs, which is the main goal of this review. Moreover, the identification of a wide range of ADLs by a
mobile device is a major milestone on the development of a personal digital life coach [153]. The sensor
fusion techniques increase the reliability of these systems, allowing, e.g., to learn the different peoples’
lifestyles. Many mobile applications send data to a remote computer that carries out the processing
and sends back the results to the mobile device. However, the local processing is faster than server-side
processing, reducing the bandwidth usage during all processes.
5. Conclusions
Data acquisition, data processing, data imputation on one sensor data stream and, finally, multiple
sensor data fusion together are the proposed roadmap to achieve a concrete task, and as discussed
in this paper, the task at hand is the identification of activities of daily living. While these steps pose
little challenge when performed in fixed computational environments, where resources are virtually
30
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
19 of 27
illimited, when their execution is intended in mobile off-the-shelf devices, a new type of challenges
arises with the restrictions of the computational environment.
The use of mobile sensors requires a set of techniques to classify and to process the data acquired
to make them usable by software components and to automate the execution of specific tasks. The data
pre-processing and data cleaning tasks are performed at the start of the sensors’ data characterization.
The collected data may have inconsistent and/or unusable data, normally called environmental noise.
However, the application of filters helps with removing the unusable data. The most commonly-used
filters are the Kalman filter and its variants, as these methods are often reported to have good accuracy.
Nevertheless, it must be noted that the correct definition of the important features of the collected data
influences the correct measurement of the accuracy of the Kalman filter. Therefore, and because it is
difficult to assess the accuracy of one particular type of Kalman filter, especially when used regarding
the identification of activities of daily living, it is advisable to explore different sensor data fusion
technologies. The main methods for the identification of the different features of the sensor signal are
machine learning and pattern recognition techniques.
Sensor fusion methods are normally classified into four large groups: probabilistic
methods, statistical methods, knowledge-based theory methods and evidence reasoning methods.
Although several methods have been studied and discussed in this paper, the choice of the best
method for each purpose depends on the quantity and types of sensors used, on the diversity
in the representation of the data, on the calibration of the sensors, on the limited interoperability
of the sensors, on the constraints in the statistical models and, finally, on the limitations of the
implemented algorithms.
Sensors available in off-the-shelf mobile devices can support the implementation of sensor fusion
techniques and improve the reliability of the algorithms created for these devices. However, mobile
devices have limited processing capacity, memory and autonomy. Nevertheless, the fusion of the
sensors’ data may be performed with mobile applications, which access the sensors’ data as a
background task, processing the data collected and showing the results in a readable format to
the user or sending the results or the data to a central repository or central processing machine.
The techniques related to the concepts of sensor and multi-sensor data fusion presented in this
paper have different purposes, including the detection/identification of activities of daily living and
other medical applications. For off-the-shelf mobile devices, the positioning of the device during the
data acquisition process is in itself an additional challenge, as the accuracy, precision and usability of
obtained data are also a function of the sensors’ location. Therefore, it has to be expected that these
systems can fail only because of poor mobile device positioning.
Several research studies have been carried out regarding sensor fusion techniques applied to the
sensors available in off-the-shelf mobile devices, but most of them only detect basic activities, such
as walking. Due to the large market of mobile devices, e.g., smartphones, tablets or smartwatches,
ambient assisted living applications on these platforms become relevant for a variety of purposes,
including tele-medicine, monitoring of elderly people, monitoring of sport performance and other
medical, recreational, fitness or leisure activities. Moreover, the identification of a wide range of
activities of daily living is a milestone in the process of building a personal digital life coach.
The applicability of sensor data fusion techniques for mobile platforms is therefore dependent
on the variable characteristics of the mobile platform itself, as these are very diverse in nature and
features, from local storage capability to local processing power, battery life or types of communication
protocols. Nevertheless, experimenting with the algorithms and techniques described previously, so as
to adapt them to a set of usage scenarios and a class of mobile devices, will render these techniques
usable in most mobile platforms without major impairments.
Acknowledgments: This work was supported by the FCT project UID/EEA/50008/2013 (Este trabalho foi suportado
pelo projecto FCT UID/EEA/50008/2013). The authors would also like to acknowledge the contribution of the COST
Action IC1303–AAPELE - Architectures, Algorithms and Protocols for Enhanced Living Environments.
Author Contributions: All of the authors have contributed to the structure, content and writing of the paper.
31
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
20 of 27
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
32
Yin, G.; Bruckner, D. Daily activity learning from motion detector data for Ambient Assisted Living.
In Technological Innovation for Sustainability; Springer Berlin Heidelberg: Berlin, Germany; Heidelberg,
Germany, 2010. pp. 89–94.
Memon, M.; Wagner, S.R.; Pedersen, C.F.; Beevi, F.H.; Hansen, F.O. Ambient assisted living healthcare
frameworks, platforms, standards, and quality attributes. Sensors 2014, 14, 4312–4341.
Siegel, C.; Hochgatterer, A.; Dorner, T.E. Contributions of ambient assisted living for health and quality
of life in the elderly and care services–A qualitative analysis from the experts’ perspective of care service
professionals. BMC Geriatr. 2014, 14, 112.
Holzinger, A.; Röcker, C.; Ziefle, M., From Smart Health to Smart Hospitals. In Smart Health: State-of-the-Art
and Beyond, Springer Lecture Notes in Computer Science, LNCS 8700; Springer: Berlin, Germany, 2015.
Developers, A. Sensors Overview | Android Developers 2015. Available online: http://developer.android.
com/guide/topics/sensors/sensors_overview.html (accessed on 1 February 2016).
Apple. Apple - iPhone 6 - Technical Specifications 2015. Available online: http://www.apple.com/iphone-6/
specs/ (accessed on 1 February 2016).
Santochi, M.; Dini, G. Sensor Technology in Assembly Systems. CIRP Ann. Manuf. Technol. 1998, 47, 503–524.
White, R. A Sensor Classification Scheme. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 1987, 34, 124–126.
Roebuck, A.; Monasterio, V.; Gederi, E.; Osipov, M.; Behar, J.; Malhotra, A.; Penzel, T.; Clifford, G.D.
A review of signals used in sleep analysis. Physiol. Meas. 2014, 35, R1–R57.
Preece, S.J.; Goulermas, J.Y.; Kenney, L.P.; Howard, D.; Meijer, K.; Crompton, R. Activity identification using
body mounted sensors: a review of classification techniques. Physiol. Meas. 2009, 30, R1–R33.
Mumolo, E.; Nolich, M.; Vercelli, G. Algorithms for acoustic localization based on microphone array in
service robotics. Robot. Auton. Syst. 2003, 42, 69–88.
Guan, D.; Yuan, W.; Jehad Sarkar, A.M.; Ma, T.; Lee, Y.K. Review of Sensor-based Activity Recognition
Systems. IETE Techn. Rev. 2011, 28, 418.
Gentili, M.; Mirchandani, P.B. Locating sensors on traffic networks: Models, challenges and research
opportunities. Transp. Res. Part C: Emerg. Technol. 2012, 24, 227–255.
Lee, S.; Ozsecen, M.; Della Toffola, L.; Daneault, J.F.; Puiatti, A.; Patel, S.; Bonato, P. Activity detection
in uncontrolled free-living conditions using a single accelerometer. In Proceedings of the 2015 IEEE 12th
International Conference on Wearable and Implantable Body Sensor Networks (BSN), Cambridge, MA, USA,
9–12 June 2015; pp. 1-6.
Scheeper, P.R.; van der Donk, A.G.H.; Olthuis, W.; Bergveld, P. A review of silicon microphones.
Sens. Actuators A Phys. 1994, 44, 1–11.
Hoummady, M.; Campitelli, A.; Wlodarski, W. Acoustic wave sensors: Design, sensing mechanisms and
applications. Smart Mater. Struct. 1997, 6, 647.
Wilson, A.D.; Baietto, M. Applications and advances in electronic nose technologies. Sensors 2009,
9, 5099–5148.
Suzuki, T.; Nakauchi, Y. Intelligent medicine case that monitors correct dosing. In Proceedings of the 2010
7th International Symposium on Mechatronics and Its Applications (ISMA), Sharjah, UAE, 20–22 April 2010.
Hill, E.W.; Vijayaragahvan, A.; Novoselov, K. Graphene Sensors. IEEE Sens. J. 2011, 11, 3161–3170.
Lenz, J.; Edelstein, S. Magnetic sensors and their applications. IEEE Sens. J. 2006, 6, 631–649.
Allen, J. Photoplethysmography and its application in clinical physiological measurement. Physiol. Meas.
2007, 28, R1–R39.
Deng, K.-L. Next generation fiber optic sensing and its applications. In Proceedings of the International
Conference on Wireless and Optical Communications, Newark, NJ, USA, 22–23 April 2005.
Khusainov, R.; Azzi, D.; Achumba, I.E.; Bersch, S.D. Real-time human ambulation, activity, and physiological
monitoring: taxonomy of issues, techniques, applications, challenges and limitations. Sensors 2013,
13, 12852–12902.
Zhang, L.; Liu, J.; Jiang, H.; Guan, Y. SensTrack: Energy-Efficient Location Tracking With Smartphone
Sensors. IEEE Sens. J. 2013, 13, 3775–3784.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
21 of 27
Novak, D.; Rebersek, P.; De Rossi, S.M.; Donati, M.; Podobnik, J.; Beravs, T.; Lenzi, T.; Vitiello, N.;
Carrozza, M.C.; Munih, M. Automated detection of gait initiation and termination using wearable sensors.
Med. Eng. Phys. 2013, 35, 1713–1720.
He, Y.; Li, Y. Physical Activity Recognition Utilizing the Built-In Kinematic Sensors of a Smartphone. Int. J.
Distrib. Sens. Netw. 2013, 2013, 1–10.
Stopczynski, A.; Stahlhut, C.; Larsen, J.E.; Petersen, M.K.; Hansen, L.K. The smartphone brain scanner: A
portable real time neuroimaging system. PLoS ONE 2014, 9, e86733.
Daponte, P.; de Vito, L.; Picariello, F.; Riccio, M. State of the art and future developments of measurement
applications on smartphones. Measurement 2013, 46, 3291–3307.
Scalvini, S.; Baratti, D.; Assoni, G.; Zanardini, M.; Comini, L.; Bernocchi, P. Information and communication
technology in chronic diseases: A patient’s opportunity. J. Med. Person 2013, 12, 91–95.
Bersch, S.D.; Azzi, D.; Khusainov, R.; Achumba, I.E.; Ries, J. Sensor data acquisition and processing
parameters for human activity classification. Sensors 2014, 14, 4239–4270.
Lim, L.; Misra, A.; Mo, T. Adaptive data acquisition strategies for energy-efficient, smartphone-based,
continuous processing of sensor streams. Distrib. Parallel Databases 2012, 31, 321–351.
Paucher, R.; Turk, M. Location-based augmented reality on mobile phones. In Proceedings of the 2010
IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPRW),
San Francisco, CA, USA, 13–18 June 2010.
Misra, A.; Lim, L. Optimizing Sensor Data Acquisition for Energy-Efficient Smartphone-Based Continuous
Event Processing. In Proceedings of the 2011 12th IEEE International Conference on Mobile Data
Management (MDM), Lulea, Sweden, 6–9 June 2011.
Kang, S.; Lee, Y.; Min, C.; Ju, Y.; Park, T.; Lee, J.; Rhee, Y.; Song, J. Orchestrator: An active resource
orchestration framework for mobile context monitoring in sensor-rich mobile environments. In Proceedings
of the 2010 IEEE International Conference on Pervasive Computing and Communications (PerCom),
Mannheim, Germany, 29 March–2 April 2010; pp. 135–144.
Agrawal, S.; Kamal, R. Computational Orchestrator: A Super Class for Matrix, Robotics and Control System
Orchestration. Int. J. Comput. Appl. 2015, 117, 12–19.
Vallina-Rodriguez, N.; Crowcroft, J. ErdOS: Achieving Energy Savings in Mobile OS. In Proceedings of the
Sixth International Workshop on MobiArch (MobiArch’11), Bethesda, MD, USA, 28 June 2011; pp. 37–42.
Priyantha, B.; Lymberopoulos, D.; Liu, J. LittleRock: Enabling Energy-Efficient Continuous Sensing on
Mobile Phones. IEEE Pervasive Comput. 2011, 10, 12–15.
Lu, H.; Yang, J.; Liu, Z.; Lane, N.D.; Choudhury, T.; Campbell, A.T. The Jigsaw Continuous Sensing Engine
for Mobile Phone Applications. In Proceedings of the 8th ACM Conference on Embedded Networked Sensor
Systems (SenSys’10), Zurich, Switzerland, 3–5 November 2010; pp. 71–84.
Rachuri, K.K.; Mascolo, C.; Musolesi, M.; Rentfrow, P.J. SociableSense: Exploring the Trade-offs of
Adaptive Sampling and Computation Offloading for Social Sensing. In Proceedings of the 17th Annual
International Conference on Mobile Computing and Networking (MobiCom’11), Las Vegas, NV, USA,
19–23 September 2011; pp. 73–84.
Deshpande, A.; Guestrin, C.; Madden, S.R.; Hellerstein, J.M.; Hong, W. Model-driven Data Acquisition in
Sensor Networks. In Proceedings of the Thirtieth International Conference on Very Large Data Bases-Volume
30. VLDB Endowment (VLDB’04), Toronto, ON, Canada, 31 August–3 September 2004; pp. 588–599.
Reilent, E.; Loobas, I.; Pahtma, R.; Kuusik, A. Medical and context data acquisition system for patient
home monitoring. In Proceedings of the 2010 12th Biennial Baltic Electronics Conference (BEC), Tallinn,
Republic of Estonia, 4–6 October 2010.
Marzencki, M.; Hung, B.; Lin, P.; Huang, Y.; Cho, T.; Chuo, Y.; Kaminska, B. Context-aware physiological
data acquisition and processing with wireless sensor networks. In Proceedings of the 2010 IEEE International
Workshop on Medical Measurements and Applications Proceedings (MeMeA), Ottawa, ON, Canada,
30 April–1 May 2010.
Agoston, K.; Nagy, C. Data acquisition for angular measuring and positioning system. In Proceedings of
the 2012 IEEE International Conference on Automation Quality and Testing Robotics (AQTR), Cluj-Napoca,
Romania, 24–27 May 2012 .
Harvey, P.; Woodward, B.; Datta, S.; Mulvaney, D. Data acquisition in a wireless diabetic and cardiac
monitoring system. IEEE Eng. Med. Biol. Soc. Conf. Proc. 2011, 2011, 3154–3157.
33
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
34
22 of 27
Bieber, G.; Haescher, M.; Vahl, M. Sensor requirements for activity recognition on smart watches.
In Proceedings of the 6th International Conference on PErvasive Technologies Related to Assistive
Environments, Island of Rhodes, Greece, 29–31 May 2013; pp. 1–6.
Sorber, J.M.; Shin, M.; Peterson, R.; Kotz, D. Plug-n-trust: Practical trusted sensing for mhealth.
In Proceedings of the 10th International Conference on Mobile Systems, Applications, and Services,
Lake District, UK, 25–29 June 2012; pp. 309–322.
Zhu, W.; Liu, L.; Yin, S.; Hu, S.; Tang, E.Y.; Wei, S. Motion-sensor fusion-based gesture recognition and its
VLSI architecture design for mobile devices. Int. J. Electron. 2013, 101, 621–635.
Lane, N.; Miluzzo, E.; Lu, H.; Peebles, D.; Choudhury, T.; Campbell, A. A survey of mobile phone sensing.
IEEE Commun. Mag. 2010, 48, 140–150.
Kim, T.h.; Adeli, H.; Robles, R.J.; Balitanas, M. Ubiquitous Computing and Multimedia Applications;
Springer: New York, NY, USA, 2011.
Hwang, Y.C.; Oh, R.D.; Ji, G.H. A Sensor Data Processing System for Mobile Application Based Wetland
Environment Context-aware. In Ubiquitous Computing and Multimedia Applications; Springer: Berlin, Germany;
Heidelberg, Germany, 2011.
Choi, M. A Platform-Independent Smartphone Application Development Framework. In Computer Science
and Convergence; Springer Netherlands: Springer: Dordrecht, The Netherlands, 2012.
Bersch, S.D.; Azzi, D.; Khusainov, R.; Achumba, I.E.; Ries, J. Sensor data acquisition and processing
parameters for human activity classification. Sensors 2014, 14, 4239–4270.
Pejovic, V.; Musolesi, M. Anticipatory Mobile Computing. ACM Comput. Surv. 2015, 47, 1–29.
Imai, S.; Miyamoto, M.; Arai, Y.; Inomata, T. A data processing method for motion estimation considering
network and sensor node loads. In Proceedings of the 2012 IEEE 11th International Conference on Cognitive
Informatics & Cognitive Computing (ICCI*CC), Kyoto, Japan, 22–24 August 2012; pp. 356–362.
Pombo, N.; Garcia, N.; Felizardo, V.; Bousson, K. Big data reduction using RBFNN: A predictive
model for ECG waveform for eHealth platform integration. In Proceedings of the 2014 IEEE 16th
International Conference on e-Health Networking, Applications and Services (Healthcom), Natal, Brazil,
15–18 October 2014.
Yamada, S.; Watanabe, Y.; Kitagawa, H.; Amagasa, T. Location-Based Information Delivery Using Stream
Processing Engine. In Proceedings of the 7th International Conference on Mobile Data Management
(2006 MDM), Nara, Japan, 10–12 May 2006.
Lin, F.X.; Rahmati, A.; Zhong, L. Dandelion: A framework for transparently programming phone-centered
wireless body sensor applications for health. In Proceedings of the WH’10 Wireless Health 2010, San Diego,
CA, USA, 5–7 October 2010.
Dolui, K.; Mukherjee, S.; Datta, S.K. Smart Device Sensing Architectures and Applications. In Proceedings
of the 2013 International Computer Science and Engineering Conference (Icsec), Bangkok, Thailand,
4–7 September 2013; pp. 91–96.
Imai, S.; Miyamoto, M.; Arai, Y.; Inomata, T. Sensor Data Processing Method Based on Observed Person’s
Similarity for Motion Estimation. In Proceedings of the 2013 IEEE 27th International Conference on Advanced
Information Networking and Applications Workshops (Waina), Barcelona, Spain, 25–28 March 2013;
pp. 601–606.
Gurrin, C.; Qiu, Z.; Hughes, M.; Caprani, N.; Doherty, A.R.; Hodges, S.E.; Smeaton, A.F. The smartphone as
a platform for wearable cameras in health research. Am. J. Prev. Med. 2013, 44, 308–313.
Postolache, O.; Girao, P.S.; Ribeiro, M.; Guerra, M.; Pincho, J.; Santiago, F.; Pena, A. Enabling telecare
assessment with pervasive sensing and Android OS smartphone. In Proceedings of the 2011 IEEE
International Workshop on Medical Measurements and Applications Proceedings (MeMeA), Bari, Italy,
30–31 May 2011 .
Wang, G.; Zimmermann, R. Spatial sensor data processing and analysis for mobile media applications.
In Proceedings of the 1st ACM SIGSPATIAL PhD Workshop, Dallas, TX, USA, 4–7 November 2014; pp. 1–5.
Vateekul, P.; Sarinnapakorn, K. Tree-Based Approach to Missing Data Imputation. In Proceedings of the
IEEE International Conference on Data Mining Workshops (ICDMW’09), Miami, FL, USA, 6 December 2009.
D’Ambrosio, A.; Aria, M.; Siciliano, R. Accurate Tree-based Missing Data Imputation and Data Fusion
within the Statistical Learning Paradigm. J. Classif. 2012, 29, 227–258.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
84.
85.
86.
87.
23 of 27
Huang, X.Y.; Li, W.; Chen, K.; Xiang, X.H.; Pan, R.; Li, L.; Cai, W.X. Multi-matrices factorization with
application to missing sensor data imputation. Sensors 2013, 13, 15172–15186.
García-Laencina, P.J.; Sancho-Gómez, J.L.; Figueiras-Vidal, A.R.; Verleysen, M. K nearest neighbours with
mutual information for simultaneous classification and missing data imputation. Neurocomputing 2009,
72, 1483–1493.
Ni, D.; Leonard, J.D.; Guin, A.; Feng, C. Multiple Imputation Scheme for Overcoming the Missing Values
and Variability Issues in ITS Data. J. Transp. Eng. 2005, 131, 931–938.
Smith, B.; Scherer, W.; Conklin, J. Exploring Imputation Techniques for Missing Data in Transportation
Management Systems. Transp. Res. Rec. 2003, 1836, 132–142.
Qu, L.; Zhang, Y.; Hu, J.; Jia, L.; Li, L. A BPCA based missing value imputing method for traffic flow
volume data. In Proceedings of the 2008 IEEE Intelligent Vehicles Symposium, Eindhoven, The Netherlands,
4–6 June 2008 .
Jiang, N.; Gruenwald, L. Estimating Missing Data in Data Streams. In Proceedings of the 12th
International Conference on Database Systems for Advanced Applications (DASFAA 2007), Bangkok,
Thailand, 9–12 April 2007.
Ling, W.; Dong-Mei, F. Estimation of Missing Values Using a Weighted K-Nearest Neighbors Algorithm.
In Proceedings of the International Conference on Environmental Science and Information Application
Technology (ESIAT 2009), Wuhan, China, 4–5 July 2009.
Hruschka, E.R.; Hruschka, E.R.; Ebecken, N.F.F., Towards Efficient Imputation by Nearest-Neighbors: A
Clustering-Based Approach. In AI 2004: Advances in Artificial Intelligence; Springer: Berlin, Germany;
Heidelberg, Germany, 2004; pp. 513–525.
Luo, J.; Yang, T.; Wang. Y. Missing value estimation for microarray data based on fuzzy C-means clustering.
In Proceedings of the Eighth International Conference on High-Performance Computing in Asia-Pacific
Region, Beijing, China, 1 July 2005.
Smaragdis, P.; Raj, B.; Shashanka, M. Missing Data Imputation for Time-Frequency Representations of Audio
Signals. J. Signal Process. Syst. 2010, 65, 361–370.
Liu, Y.; Lv, Z.; Wang, W. An Improved Generalized-Trend-Diffusion-Based Data Imputation for Steel Industry.
Math. Probl. Eng. 2013, 2013, 1–10.
Iacus, S.M.; Porro, G. Missing data imputation, matching and other applications of random recursive
partitioning. Comput. Stat. Data Anal. 2007, 52, 773–789.
Bruni, R. Discrete models for data imputation. Discret. Appl. Math. 2004, 144, 59–69.
Pombo, N.G.C.C. Information Technologies for Pain Management. Ph.D. Thesis, University of Beira Interior,
Covilhã, Portugal, 2014.
Krichmar, J.L.; Snook, J.A. A neural approach to adaptive behavior and multi-sensor action selection in a
mobile device. Rob. Autom. 2002, 4, 3864–3869.
Kim, D.J.; Prabhakaran, B. Faulty and Missing Body Sensor Data Analysis. In Proceedings of the 2013 IEEE
International Conference on Healthcare Informatics (ICHI), Philadelphia, PA, USA, 9–11 September 2013.
Banos, O.; Damas, M.; Pomares, H.; Rojas, I. On the use of sensor fusion to reduce the impact of rotational
and additive noise in human activity recognition. Sensors 2012, 12, 8039–8054.
Ma, Z.; Qiao, Y.; Lee, B.; Fallon, E. Experimental evaluation of mobile phone sensors. In Proceedings of the
Signals and Systems Conference (ISSC 2013), 24th IET Irish, Letterkenny, Ireland, 20–21 June 2013.
Durrant-Whyte, H.; Stevens, M.; Nettleton, E. Data fusion in decentralised sensing networks. In Proceedings
of the 4th International Conference on Information Fusion, Montreal, PQ, Canada, 7–10 August 2001.
Aziz, A.M. A new adaptive decentralized soft decision combining rule for distributed sensor systems with
data fusion. Inf. Sci. 2014, 256, 197–210.
Akhoundi, M.A.A.; Valavi, E. Multi-Sensor Fuzzy Data Fusion Using Sensors with Different Characteristics.
ArXiv E-Prints 2010, arXiv:1010.6096.
Khaleghi, B.; Khamis, A.; Karray, F.O.; Razavi, S.N. Multisensor data fusion: A review of the state-of-the-art.
Inf. Fusion 2013, 14, 28-44.
Pombo, N.; Bousson, K.; Araújo, P.; Viana, J. Medical decision-making inspired from aerospace multisensor
data fusion concepts. Inf. Health Soc. Care 2015, 40, 185–197.
35
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
88.
89.
90.
91.
92.
93.
94.
95.
96.
97.
98.
99.
100.
101.
102.
103.
104.
105.
106.
107.
108.
36
24 of 27
Ko, M.H.; West, G.; Venkatesh, S.; Kumar, M. Using dynamic time warping for online temporal fusion in
multisensor systems. Inf. Fusion 2008, 9, 370–388.
Tanveer, F.; Waheed, O.T.; ur Rehman, A. Design and Development of a Sensor Fusion based Low Cost
Attitude Estimator. J. Space Technol. 2011, 1, 45–50.
Zhao, L.; Wu, P.; Cao, H. RBUKF Sensor Data Fusion for Localization of Unmanned Mobile Platform. Res. J.
Appl. Sci. Eng. Technol. 2013, 6, 3462–3468.
Walter, O.; Schmalenstroeer, J.; Engler, A.; Haeb-Umbach, R. Smartphone-based sensor fusion for improved
vehicular navigation. In Proceedings of the 10th Workshop on Positioning Navigation and Communication
(WPNC), Dresden, Germay, 20–21 March 2013.
Iglesias, J.; Cano, J.; Bernardos, A.M.; Casar, J.R. A ubiquitous activity-monitor to prevent sedentariness.
In Proceedings of the IEEE International Conference on Pervasive Computing and Communications
Workshops (PERCOM Workshops), Seattle, WA, USA, 21–25 March 2011.
Yongkai, Z.; Shuangquan, W.; Zhuang, Z.; Canfeng, C.; Jian, M. A mobile device oriented framework for
context information management. In Proceedings of the IEEE Youth Conference on Information, Computing
and Telecommunication, Beijing, China, 20–21 September 2009.
Blum, J.R.; Greencorn, D.G.; Cooperstock, J.R. Smartphone Sensor Reliability for Augmented Reality
Applications. In Mobile and Ubiquitous Systems: Computing, Networking, and Services; Springer: Berlin,
Germany, 2013.
Neidhardt, A.; Luss, H.; Krishnan, K.R. Data fusion and optimal placement of fixed and mobile sensors.
In Proceedings of the IEEE Sensors Applications Symposium, Atlanta, GA, USA, 12–14 February 2008.
Haala, N.; Böhm, J. A multi-sensor system for positioning in urban environments. ISPRS J. Photogramm.
Remote Sens. 2003, 58, 31–42.
Theodoridis, S. Machine Learning: A Bayesian and Optimization Perspective; Academic Press: Cambridge, MA,
USA, 2015.
Miller, G. The Smartphone Psychology Manifesto. Perspect. Psychol. Sci. 2012, 7, 221–237.
Yi, W.J.; Sarkar, O.; Mathavan, S.; Saniie, J. Wearable sensor data fusion for remote health assessment and
fall detection. In Proceedings of the IEEE International Conference on Electro/Information Technology (EIT),
Milwaukee, WI, USA, 5–7 June 2014.
Glenn, T.; Monteith, S. New measures of mental state and behavior based on data collected from sensors,
smartphones, and the Internet. Curr. Psychiatry Rep. 2014, 16, 1–10.
Ponmozhi, J.; Frias, C.; Marques, T.; Frazão, O. Smart sensors/actuators for biomedical applications: Review.
Measurement 2012, 45, 1675–1688.
Beach, A.; Gartrell, M.; Xing, X.; Han, R.; Lv, Q.; Mishra, S.; Seada, K. Fusing mobile, sensor, and social data
to fully enable context-aware computing. In Proceedings of the Eleventh Workshop on Mobile Computing
Systems Applications, Annapolis, MD, USA, 22–23 February 2010.
Andò, B.; Baglio, S.; Pistorio, A. A Smart Multi-Sensor Approach to Monitoring Weak People in Indoor
Environments. J. Sens. Technol. 2014, 4, 24–35.
Phan, T.; Kalasapur, S.; Kunjithapatham, A. Sensor fusion of physical and social data using Web SocialSense
on smartphone mobile browsers. In Proceedings of the IEEE 11th Consumer Communications and
Networking Conference (CCNC), Las Vegas, NV, USA, 10–13 January 2014.
Steed, A.; Julier, S. Behaviour-aware sensor fusion: Continuously inferring the alignment of coordinate
systems from user behaviour. In Proceedings of the IEEE International Symposium on Mixed and Augmented
Reality (ISMAR), Adelaide, Australia, 1–4 October 2013.
Pucihar, K.C.; Coulton, P.; Hutchinson, D. Utilizing sensor fusion in markerless mobile augmented reality.
In Proceedings of the 13th International Conference on Human Computer Interaction with Mobile Devices
and Services, Stockholm, Sweden, 30 August–2 September 2011.
Rahman, A.S.M.M.; Hossain, M.A.; Saddik, A.E. Spatial-geometric approach to physical mobile interaction
based on accelerometer and IR sensory data fusion. ACM Trans. Multimed. Comput. Commun. Appl. 2010,
6, 1–23.
Grunerbl, A.; Muaremi, A.; Osmani, V.; Bahle, G.; Ohler, S.; Troester, G.; Mayora, O.; Haring, C.; Lukowicz, P.
Smart-Phone Based Recognition of States and State Changes in Bipolar Disorder Patients. IEEE J Biomed.
Health Inf. 2014, doi:10.1109/JBHI.2014.2343154.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
25 of 27
109. Gil, G.B.; Berlanga de Jesus, A.; Molina Lopez, J.M. inContexto: A fusion architecture to obtain mobile
context. In Proceedings of the Proceedings of the 14th International Conference on Information Fusion
(FUSION), Chicago, IL, USA, 5–8 July 2011.
110. Kim, J.; Gracanin, D.; Quek, F. Sensor-fusion walking-in-place interaction technique using mobile devices.
In Proceedings of the 2012 IEEE Virtual Reality Short Papers and Posters (VRW), Costa Mesa, CA, USA,
4–8 March 2012.
111. Abadi, M.J.; Gu, Y.; Guan, X.; Wang, Y.; Hassan, M.; Chou, C.T. Improving Heading Accuracy in
Smartphone-based PDR Systems using Multi-Pedestrian Sensor Fusion. Electr. Eng. 2013 188, 9–35.
112. Altini, M.; Vullers, R.; Van Hoof, C.; van Dort, M.; Amft, O. Self-calibration of walking speed
estimations using smartphone sensors. In Proceedings of the 2014 IEEE International Conference on
Pervasive Computing and Communications Workshops (PERCOM Workshops), Budapest, Hungary,
24–28 March 2014.
113. Tsai, P.H.; Lin, Y.J.; Ou, Y.Z.; Chu, E.T.H.; Liu, J.W.S. A Framework for Fusion of Human Sensor and Physical
Sensor Data. IEEE Trans. Syst. Man Cybern. Syst. 2014, 44, 1248–1261.
114. Lee, B.G.; Chung, W.Y. A smartphone-based driver safety monitoring system using data fusion. Sensors
2012, 12, 17536–17552.
115. Chen, D.; Schmidt, A.; Gellersen, H.W. An Architecture for Multi-Sensor Fusion in Mobile Environments.
Available online: http://www.cs.cmu.edu/ datong/Fusion99.pdf (accessed on 1 February 2016).
116. Sashima, A.; Ikeda, T.; Kurumatani, K. Toward Mobile Sensor Fusion Platform for Context Aware Services.
Available online: http://cdn.intechopen.com/pdfs-wm/6799.pdf (accessed on 1 February 2016).
117. Ayub, S.; Bahraminisaab, A.; Honary, B. A sensor fusion method for smart phone orientation estimation.
In Proceedings of the 13th Annual Post Graduate Symposium on the Convergence of Telecommunications,
Networking and Broadcasting, Liverpool, UK, 25–26 June 2012
118. Zhu, W.; Liu, L.; Yin, S.; Hu, S.; Tang, E.Y.; Wei, S. Motion-sensor fusion-based gesture recognition and its
VLSI architecture design for mobile devices. Int. J. Electron. 2013, 101, 621–635.
119. van de Ven, P.; Bourke, A.; Tavares, C.; Feld, R.; Nelson, J.; Rocha, A.; O Laighin, G. Integration of a suite
of sensors in a wireless health sensor platform. In Proceedings of the 2009 IEEE Sensors, Christchurch,
New Zealand, 25–28 October 2009
120. Chen, J.; Low, K.H.; Tan, C.K.Y.; Oran, A.; Jaillet, P.; Dolan, J.M.; Sukhatme, G.S. Decentralized Data Fusion
and Active Sensing with Mobile Sensors for Modeling and Predicting Spatiotemporal Traffic Phenomena
arXiv preprint arXiv:1206.6230 2012, arXiv:1206.6230.
121. Zhao, D.; Ma, H.; Tang, S. COUPON: Cooperatively Building Sensing Maps in Mobile Opportunistic
Networks. In Proceedings of the 2013 IEEE 10th International Conference on Mobile Ad-Hoc and Sensor
Systems (MASS), Hangzhou, China, 14–16 October 2013.
122. Zhao, D.; Ma, H.; Tang, S.; Li, X.Y. COUPON: A Cooperative Framework for Building Sensing Maps in
Mobile Opportunistic Networks. IEEE Trans. Parallel Distrib. Syst. 2015, 26, 392–402.
123. Deyun, G.; Tao, Z.; Dong, P.; Sidong, Z. A general multi-sensor node in wireless sensor networks.
In Proceedings of the 2009. ICCTA ’09. IEEE International Conference on Communications Technology and
Applications, Beijing, China, 16–18 October 2009.
124. Fortino, G.; Galzarano, S.; Gravina, R.; Li, W. A framework for collaborative computing and multi-sensor
data fusion in body sensor networks. Inf. Fusion 2015, 22, 50–70.
125. Zheng, E.; Chen, B.; Wang, X.; Huang, Y.; Wang, Q. On the Design of a Wearable Multi sensor System for
Recognizing Motion Modes and Sit to stand Transition. Int. J. Adv. Robot. Syst. 2014, doi:10.5772/57788.
126. Chen, J.I.Z. An Algorithm of Mobile Sensors Data Fusion Orientation tracking for Wireless Sensor Networks.
Wirel. Pers. Commun. 2009, 58, 197–214.
127. Saeedi, S.; Moussa, A.; El-Sheimy, N. Context-aware personal navigation using embedded sensor fusion in
smartphones. Sensors 2014, 14, 5742–5767.
128. Bhuiyan, M.Z.H.; Kuusniemi, H.; Chen, L.; Pei, L.; Ruotsalainen, L.; Guinness, R.; Chen, R.
Performance Evaluation of Multi-Sensor Fusion Models in Indoor Navigation. Eur. J. Navig. 2013, 11, 21–28.
129. Martin, H.; Bernardos, A.M.; Tarrio, P.; Casar, J.R. Enhancing activity recognition by fusing inertial and
biometric information. In Proceedings of the 14th International Conference on Information Fusion, Chicago,
IL, USA, 5–8 July 2011.
37
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
26 of 27
130. Bellos, C.; Papadopoulos, A.; Rosso, R.; Fotiadis, D.I. Heterogeneous data fusion and intelligent techniques
embedded in a mobile application for real-time chronic disease management. Conf. Proc. IEEE Eng. Med.
Biol. Soc. 2011, 2011, 8303–8306.
131. Martín, H.; Bernardos, A.M.; Iglesias, J.; Casar, J.R. Activity logging using lightweight classification
techniques in mobile devices. Pers. Ubiquitous Comput. 2012, 17, 675–695.
132. Thatte, G.; Li, M.; Lee, S.; Emken, B.A.; Annavaram, M.; Narayanan, S.; Spruijt-Metz, D.; Mitra, U.
Optimal Time-Resource Allocation for Energy-Efficient Physical Activity Detection. IEEE Trans. Signal
Process. 2011, 59, 1843–1857.
133. Scheuermann, B.; Ehlers, A.; Riazy, H.; Baumann, F.; Rosenhahn, B. Ego-motion compensated face detection
on a mobile device. In Proceedings of the 2011 IEEE Computer Society Conference on Computer Vision and
Pattern Recognition Workshops (CVPRW), Colorado Springs, CO, USA, 20–25 June 2011.
134. Klingeberg, T.; Schilling, M. Mobile wearable device for long term monitoring of vital signs. Comput. Methods
Progr. Biomed. 2012, 106, 89–96.
135. Jin, Y.Y.; Toh, H.S.; Soh, W.S.; Wong, W.C. A Robust Dead-Reckoning Pedestrian Tracking System with
Low Cost Sensors. In Proceedings of the 2011 IEEE International Conference on Pervasive Computing and
Communications (PerCom), Seattle, WA, USA, 21–25 March 2011.
136. Grunerbl, A.; Muaremi, A.; Osmani, V.; Bahle, G.; Ohler, S.; Troester, G.; Mayora, O.; Haring, C.; Lukowicz, P.
Smart-Phone Based Recognition of States and State Changes in Bipolar Disorder Patients. IEEE J. Biomed.
Health Inf. 2014, doi:10.1109/JBHI.2014.2343154.
137. García, F.; Jiménez, F.; Anaya, J.J.; Armingol, J.M.; Naranjo, J.E.; de la Escalera, A. Distributed Pedestrian
Detection Alerts Based on Data Fusion with Accurate Localization. Sensors 2013, 13, 11687–11708.
138. Ou, S.; Fagg, A.H.; Shenoy, P.; Chen, L. Application Of Reinforcement Learning In Multisensor Fusion
Problems With Conflicting Control Objectives. Intell. Autom. Soft Comput. 2009, 15, 223–235.
139. Wang, J.; Chen, G.; Kotz, D. A sensor-fusion approach for meeting detection. In Proceedings of the Workshop
on Context Awareness at the Second International Conference on Mobile Systems, Applications, and Services,
Boston, MA, USA, 6–9 June 2004.
140. Seoane, F.; Mohino-Herranz, I.; Ferreira, J.; Alvarez, L.; Buendia, R.; Ayllón, D.; Llerena, C.; Gil-Pita, R.
Wearable Biomedical Measurement Systems for Assessment of Mental Stress of Combatants in Real Time.
Sensors 2014, 14, 7120–7141.
141. Luque, R.; Casilari, E.; Moron, M.J.; Redondo, G. Comparison and characterization of Android-based fall
detection systems. Sensors 2014, 14, 18543–18574.
142. Salah, O.; Ramadan, A.A.; Sessa, S.; Ismail, A.A.; Fujie, M.; Takanishi, A. ANFIS-based Sensor Fusion System
of Sit- to- stand for Elderly People Assistive Device Protocols. Int. J. Autom. Comput. 2014, 10, 405–413.
143. Kleinberger, T.; Becker, M.; Ras, E.; Holzinger, A.; Müller, P. Ambient Intelligence in Assisted Living: Enable
Elderly People to Handle Future Interfaces. In Proceedings of the 4th International Conference on Universal
Access in Human-computer Interaction: Ambient Interaction, Beijing, China, 22–27 July 2007.
144. Holzinger, A.; Searle, G.; Pruckner, S.; Steinbach-Nordmann, S.; Kleinberger, T.; Hirt, E.; Temnitzer, J.
Perceived usefulness among elderly people: Experiences and lessons learned during the evaluation of a
wrist device. In Proceedings of the 2010 4th International Conference on Pervasive Computing Technologies
for Healthcare (PervasiveHealth), Munich, Germany, 22–25 March 2010.
145. Ziefle, M.; Rocker, C.; Holzinger, A. Medical Technology in Smart Homes: Exploring the User’s Perspective
on Privacy, Intimacy and Trust. In Proceedings of the 2011 IEEE 35th Annual Computer Software and
Applications Conference Workshops (COMPSACW), Munich, Germany, 18–22 July 2011.
146. Macias, E.; Suarez, A.; Lloret, J. Mobile sensing systems. Sensors 2013, 13, 17292–17321.
147. Volkov, A.S. Accuracy bounds of non-Gaussian Bayesian tracking in a NLOS environment. Signal Process.
2015, 108, 498–508.
148. Gulrez, T.; Kavakli, M. Precision Position Tracking in Virtual Reality Environments using Sensor Networks.
In Proceedings of the IEEE International Symposium on Industrial Electronics, 2007. ISIE 2007, Vigo, Spain,
4–7 June, 2007.
149. Chen, C.A.; Chen, S.L.; Huang, H.Y.; Luo, C.H. An asynchronous multi-sensor micro control unit for wireless
body sensor networks (WBSNs). Sensors 2011, 11, 7022–7036.
38
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2016, 16, 184
27 of 27
150. Castro Garrido, P.; Luque Ruiz, I.; Gomez-Nieto, M.A. AGATHA: Multiagent system for user monitoring.
In Proceedings of the 2012 IEEE International Conference on Consumer Electronics - Berlin (ICCE-Berlin),
Berlin, Germany, 3–5 September 2012.
151. Broggi, A.; Debattisti, S.; Panciroli, M.; Grisleri, P.; Cardarelli, E.; Buzzoni, M.; Versari, P. High performance
multi-track recording system for automotive applications. Int. J. Autom. Technol. 2011, 13, 123–132.
152. Dong, J.; Zhuang, D.; Huang, Y.; Fu, J. Advances in multi-sensor data fusion: algorithms and applications.
Sensors 2009, 9, 7771–7784.
153. Garcia, N.M. A Roadmap to the Design of a Personal Digital Life Coach. ICT Innovations 2016, 399, 21–27.
© 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons by Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).
39
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
40
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
2. Validation Techniques for Sensor Data in Mobile
Health Applications
The following article is the second part of the chapter 2.
Validation Techniques for Sensor Data in Mobile Health Applications
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta and Natalia Díaz
Rodríguez
Journal of Sensors (Hindawi Publishing Corporation), published, 2016.
According to 2017 Journal Citation Reports published by Thomson Reuters in 2018, this journal
has the following performance metrics:
ISI Impact Factor (2017): 2.057
ISI Article Influence Score (2017): 0.3
Journal Ranking (2017): 259/644 (Electrical and Electronic Engineering)
41
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
42
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Hindawi Publishing Corporation
Journal of Sensors
Volume 2016, Article ID 2839372, 9 pages
http://dx.doi.org/10.1155/2016/2839372
Review Article
Validation Techniques for Sensor Data in
Mobile Health Applications
Ivan Miguel Pires,1,2,3 Nuno M. Garcia,1,3,4 Nuno Pombo,1,3
Francisco Flórez-Revuelta,5 and Natalia Díaz Rodríguez6
1
Instituto de Telecomunicações, Universidade of Beira Interior, Covilhã, Portugal
Altranportugal, Lisbon, Portugal
3
Assisted Living Computing and Telecommunications Laboratory (ALLab), Computer Science Department,
Universidade of Beira Interior, Covilhã, Portugal
4
ECATI, Universidade Lusófona de Humanidades e Tecnologias, Lisbon, Portugal
5
Department of Computer Technology, Universidad de Alicante, Alicante, Spain
6
Department of Computer Science and Artificial Intelligence, CITIC-UGR, University of Granada, Granada, Spain
2
Correspondence should be addressed to Ivan Miguel Pires; impires@it.ubi.pt
Received 27 February 2016; Revised 25 August 2016; Accepted 4 September 2016
Academic Editor: Francesco Dell’Olio
Copyright © 2016 Ivan Miguel Pires et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
Mobile applications have become a must in every user’s smart device, and many of these applications make use of the device sensors’
to achieve its goal. Nevertheless, it remains fairly unknown to the user to which extent the data the applications use can be relied
upon and, therefore, to which extent the output of a given application is trustworthy or not. To help developers and researchers and
to provide a common ground of data validation algorithms and techniques, this paper presents a review of the most commonly
used data validation algorithms, along with its usage scenarios, and proposes a classification for these algorithms. This paper also
discusses the process of achieving statistical significance and trust for the desired output.
1. Introduction
There has been an increase of the number of mobile applications that make use of sensors to achieve a plethora of goals.
Many of these applications are designed and developed by
amateur programmers, and that in itself is good as it confirms
an increase in the overall set of skills of the developer
community. Nevertheless, and even when the applications are
developed by professionals or by companies, there are not
many applications that publicize or disclose how the sensors’
data is processed. This is a problem, in particular when these
applications are meant to be used in a scenario where they
can influence their users’ lives, as for example, when the data
is expected to be used to identify Activities of Daily Living
(ADLs) or, to an extreme, when the applications are used in
medical scenarios.
Due to the nature of the mobile device itself, multiprocessing, with limited computational power and limited
battery life, the data that is collected from the sensors is often
unusable in its primary form, requiring further processing
to allow it to be representative of the event or object that it
is supposed to measure. The recording of sensor data and
the sequent processing of this data need to include validation
subtasks that guarantee that the data are suitable to be fed into
the higher-level algorithms.
Moreover, the use of the sensors’ data to feed higherlevel algorithms needs to guarantee a minimum degree of
error, with this error being the difference between the output
of these applications, built on limited computational mobile
platforms, and the output of a golden standard. To achieve
a minimum degree of error, statistical methods need to be
applied to ensure that the output of the mobile application is
to maximum extent similar to the output given by the relevant
golden standard, if and when this is possible.
To mitigate this problem, this paper presents and
discusses the most used data validation algorithms and
43
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
2
techniques and their usage in a mobile application that relies
on the sensors’ data to give meaningful output to its user. The
algorithms are listed and their use is discussed. The discussion
of the statistical process to ensure maximum reliability of the
results is also presented.
The remainder of this paper is organized as follows: this
paragraph concludes Section 1, where a short introduction
to the problem and a proposal to achieve its mitigation are
disclosed; Section 2 presents the most commonly found data
validation methods, along with a critical comparison of its
usage scenarios; Section 3 deepens the analysis presenting
a classification of the data validation methods; Section 4
discusses the applicability of these methods, including the
discussion of the degree of trust the data can be expected to
provide; finally, Section 5 presents relevant conclusions.
2. Data Validation Methods
Sensor data validation is an important process executed
during the data acquisition and data processing modules of
the multisensor mobile system. This process consists of the
validation of the external conditions of the data and the
validity of the data for specific purpose, in order to obtain
accurate and reliable results. The sequence of this validation
may be applied not only in data acquisition but also in data
processing since increase, as these increase the degree of
confidence of the systems, with the confidence in the output
being of great importance, especially for systems involved in
medical diagnosis, but also for the identification of ADLs or
sports monitoring.
In addition, data validation methods must be used during
the different phases of the conception of a new system,
such as design, development, tests, and validation. Therefore,
the data validation methods with verified reliability during
the conception should be also used to validate the data
automatically during the execution time.
One of the causes for the presence of incorrect values during the data acquisition process may be existence
of environmental noise. Even when the data is correctly
collected, the data may still be incorrect because of noise.
Therefore, very often the data captured or processed has to
be cleaned, treated, or imputed to obtain better and reliable
results. Following the existence of missing values at random
instants of time, the causes may be the mechanical problems
or power failures of sensors. At this case, data correction
methods should be applied, including data imputation and
data cleaning. The data validation process may be simplified
as presented in Figure 1.
The selection of the best technique for sensor data
validation also depends on the type of data collected, the
purpose of its application, and the computational platform
where the algorithm will be run. Data validation techniques
are commonly composed by statistical methods. Due to the
characteristics of mobile devices, data validation techniques
can be executed locally in the mobile device or at the
server-side, depending on the amount of data to validate
simultaneously, the frequency of the validation tasks, and
the computational, communication, and storage resources
needed for the validation. The characteristics of the sensors
44
Journal of Sensors
are also important for the selection of the best techniques,
which may be separated in three large groups, which are
sensor performance characteristics, pervasive metrics, and
environmental characteristics [1].
While data validation is important for improving the
reliability of a system, it also depends on other factors, such
as power instability, temperature changes, out-of-range data,
internal and external noises, and synchronization problems
that occur when multiple sensors are integrated into a
system [2]. However, the reconstruction of the data and
correction for the correct measurement is also important,
and several research studies have proposed systems, methods,
models, and frameworks to improve the data validation and
reconstruction [3, 4].
Sensor data validation methods can be separated in three
large groups, such as faulty data detection methods, data
correction methods, and other assisting techniques or tools
[5].
Firstly, faulty data detection methods may be either
simple test based methods or physical or mathematical
model based methods, and they are classified in valid data
and invalid or missingness data [6, 7]. For the detection
of faulty data, the authors in [7] presented an order of
methods that should be applied to obtain better results,
which are as follows: zero value detection, flat line detection,
minimum and maximum values detection, minimum and
maximum thresholds based on last values, statistical tests
that follow certain distributions, multivariate statistical tests,
artificial neural networks (ANNs) [8], one-class support
vector machine (SVM) [9], and classification and physical
models. On the one hand, simple test based methods include
different techniques, such as physical range check, local
realistic range detection, detection of gaps in the data, constant value detection, the signals’ gradient test, the tolerance
band method, and the material redundancy detection [7,
10, 11]. On the other hand, physical or mathematical model
based methods include extreme value check using statistics,
drift detection by exponentially weighted moving averages,
the spatial consistency method, the analytical redundancy
method, gross error detection, the multivariate statistical test
using Principal Component Analysis (PCA), and data mining
methods [7, 12, 13].
Secondly, data correction methods can be carried out by
interpolation, smoothing, data mining, and data reconciliation [10, 12, 14]. For the application of the interpolation, the
authors of [11] proposed the use of the value measured from
the last measurement or the use of the trend from previous
sets of measurements. The smoothing methods, for example,
moving average and median, may be used to filter out the
random noise and convert the data into a smooth curve that
is relatively unbiased by outliers [10]. The application of data
mining techniques allows the replacement of the faulty values
by the measurements performed with several methods, for
example, ANNs [14]. The data reconciliation methods, for
example, PCA, are used for the calculation of a minimal
correction to the measured variables, according to the several
constraints of the model [13].
Thirdly, the other assisting techniques or tools are,
namely, the checking of the status of the sensors, the checking
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
3
Journal of Sensors
Sensors
Soft sensors
Origin: data acquisition/data processing
Data cleaning/noise removal
Data is
valid?
Yes
No
Data is
complete?
Yes
No
Data imputation
Discard the data
No
Data is
valid?
Yes
Next stages: data processing/data fusion
Data discarded
Figure 1: Sequence of activities performed during the data validation process.
of the duration after sensor maintenance, data context classification, the calibration of measuring systems, and the
uncertainty consideration [6, 7, 10].
Several research studies have been performed, using data
validation techniques. In [15], PCA is used for the compression of linearly correlated data. The authors compared the
Auto-Associative Neural Network (AANN) and the Kernel
PCA (KPCA) methods for data validation, creating a new
approach named as Hybrid AANN-KPCA that uses these two
methods. When compared with AANN and KPCA methods, the Hybrid AANN-KPCA achieves better performance
results for the prediction or correction of inconsistent data.
In [16], the authors proposed that the data validation
may be performed with Kalman filtering and linear predictive
coding (LPC), showing that the results using Kalman filtering
are better than LPC using several types of data, but the LPC
reported a smaller energy consumption.
Several studies proposed the use of ANNs, for example,
the Multilayer Perceptron (MLP), that can be trained to
perform the identification of faulty sensors using prototype
data and used to determine the near optimal subset of sensor
data to produce the best results [2, 17–19]. Besides, the sensor
data validation may be performed with other probabilistic
methods, such as Bayesian Networks, Propagation in Trees,
Probabilistic Causal Methods, and Learning Algorithms [20].
The authors of [20] proposed the anytime sensor validation
algorithms that combine several probabilistic methods. On
the contrary, [21] proposed the validation of data using the
Sparse Bayesian Learning and the Relevance Vector Machine
(RVM), which are an specialization of SVM.
For the estimation of the values during data validation,
the authors of [22] analysed the use of the Kalman filter,
which was implemented in two methods: Algorithmic Sensor
Validation (ASV), and Heuristic Sensor Validation (HSV).
The ASV method implements different statistical methods,
for example, mean, standard deviation, and sensor confidence that represent the uncertain nature of sensors. HSV
identifies faulty sensor readings as attributable to a sensor or
system failure. As an example, the authors of [23] proposed
the use of the Kalman filter for the validation of the GPS
data.
Other used methods are the grey models, which consists
of differential equations describing the behaviour of an
accumulated generating operation (AGO) data sequence. As
an example, [4] presented a novel self-validating strategy
using grey bootstrap method (GBM) for data validation and
dynamic uncertainty estimation of self-validating sensor. The
GBM can evaluate the measurement uncertainty due to poor
information and small sample.
In [2], the Autoregressive Moving Averages (ARMA)
transform the process for determining the validity of the
acquired data, evaluating the levels of noise and providing
a timely warning from the expected signals. The model
created for ARMA includes linear regression techniques to
predict the invalid values with Autoregressive (AR) and
Moving Average (MA) models. Sensor Data Validation in
Aeroengine Vibration Tests also implements the Autoregressive (AR) Model, complemented with the Empirical Mode
Decomposition (EMD) [24]. Another method presented is
the sensor validation and fusion of the Nadaraya-Watson
statistical estimator [25], using a Fuzzy Logic model [26].
These methods and others, including the use of Gaussian
distributions and error detection methods, may be also used
to improve the quality of the measurements [27, 28].
45
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
4
Intelligent sensor systems are able to perform the capture
and validation of the sensors’ data. Staroswiecki [29] argues
that the data validation is important to increase the confidence level of these systems, proposing two types of validation, such as technological and functional. Technological
validation consists on the analysis of the conditions of the
hardware resources of the sensors, but it does not guarantee
that the estimation produced by the sensor is correct, but
only that the operating conditions were not against possible
correctness. On the contrary, functional validation consists
of Fault Detection and Isolation (FDI) procedures, which
consists of the use of algorithms to complement the Technological Validation. The authors of [30] also agreed with
Staroswiecki in the separation of the data validation in two
types, presenting a real time algorithm based on probabilistic
methods. Other studies have been researched and developed,
including the data validation techniques using intelligent
sensor systems [31].
Another powerful technique for data validation consists
of the use of self-validating (SEVA) sensors, which provide an
estimation of the error bounds during the measurements [32].
SEVA are widely researched in literature. An example, using
a Back-Propagation (BP) model, is applied into a system to
obtain an estimated value and then a fault detection method
called SPRT (sequential probability ratio test), identifying the
validity of the system [33]. For the use of SEVA technologies,
the authors of [34] also proposed the validated random fuzzy
variable (VRFV) based uncertainty evaluation strategy for
the online validated uncertainty (VU) estimation. In [35],
the authors presented a novel strategy of using polynomial
predictive filters coupled with VRFV which is proposed for
the online measurements validation and validated uncertainty estimation of multifunctional self-validating sensors.
These authors also performed a research about the use of
some fuzzy logic rules, comparing the predicted values with
the actual measurements to obtain the confidence evaluation
[36]. In [37], the authors proposed an approach of sensor data
validation using self-reporting, including the measurement
based on the data quality, that is, validating the data loss
measured by periodic sensors, the timing of data collection,
and the accuracy of the detection of changes. ANNs may
be used for SEVA with self-organizing maps (SOM) [38],
which are trained using unsupervised learning techniques to
produce a low-dimensional, discretized representation of the
input space of the training samples [39].
The use of valid data is important for the developments
of intelligent sensor systems, which may be used for health
purposes and, consequently, for the detection of the ADLs
[40–45]. The use of mobile devices allows the data acquisition
anywhere and at anytime, but these devices have several constraints, such as low memory, processing power, and battery
resources, but data validation may help for increasing of the
performance of the measurements, reducing the resources
needed [46–48]. In general, these systems use probabilistic
methods to detect the failures at real-time to obtain better
results.
Table 1 presents a summary of the data validation
methods included on each category. The methods that are
mainly implemented use statistical and artificial intelligence
46
Journal of Sensors
techniques, such as PCA, RVM, ANNs, and others, increasing
the reliability of the data acquisition and data processing
algorithms. In spite of the SVM and the ANN working in a
slightly different manner, their foundations are quite similar.
In fact the SVM without kernel is a single neural network
neuron with a different cost function. Congruently, when the
SVM has a kernel it is comparable with a 2-layer ANN.
Following the methods presented at Table 1, the most
studied scenarios for data validation are mainly related to
health sciences, laboratory experiments, and other undifferentiated tasks. However, only a minor part of studies is
related to the use of mobile devices, smart sensors, and other
devices used daily. Besides, the development of healthcare
solutions based on the sensors available on the mobile
devices increases the requirement of the validation of the
data collected by the sensors available on the mobile devices.
Depending on the types of the data, for some complex data
acquired, such as images, videos, GPS signal, and other
complex types of data, the validation of the data should be
accomplished by other auxiliary systems working at the same
time, validating the data at the server-side, but a constant
network connection must be available. Other topologies of
systems may be susceptible for the implementation of data
validation techniques. The Wireless Sensor Networks (WSN)
are an example of systems where the different nodes of the
network may perform the validation of the data collected for
the neighbourhood nodes, and these nodes may be composed
of different types of sensors. However, the main topology for
the implementation with mobile devices is the self-validation
using only the sensors and the data available on the mobile
device.
The data validation may be executed automatically and
transparently for the mobile devices’ user and, commonly,
at least one of the methods for each stage should be implemented in a system to perform the validation of the sensors’
data. Firstly, for faulty data detection methods, the ANNs
are the most used methods for the training of the data
and for the detection of the inconsistent values. Secondly,
for data correction methods, the most used method is the
Kalman filter. Thirdly, the other assisting techniques that
are commonly applied are the data context classification,
the checking of the status of sensors, and the uncertainty
considerations. Applying the data validation techniques correctly, the reliability and acceptability of the systems may be
increased.
3. Classification of Data Validation Methods
Data validation methods may be classified in three large
groups [5] as follows: faulty data detection methods, data
correction methods, and other assisting techniques or tools.
The faulty data detection methods and the data correction
methods may be executed sequentially in a multisensor
system in order to obtain the results based on valid data. The
other assisting techniques or tools mainly consist of the validation of the working state of the sensors, and this validation
may be executed at the same time of the execution of faulty
data detection and data correction methods, because these
types of failures invalidated the results of the algorithms.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
5
Journal of Sensors
Table 1: Classification of the data validation methods by functionality.
Groups of data validation methods Methods included
ANNs
(i) MLP; AANN; BP algorithm; SVM;
Instance based
(i) SOM
Gaussian distributions
Statistical methods
(i) ASV; HSV
Faulty data detection methods
Probabilistic methods
(i) Bayesian Networks; Propagation in Trees;
Probabilistic Causal Methods; Learning Algorithms;
Sparse Bayesian Learning; RVM; SPRT
Dimensionality Reduction
(i) Fuzzy logic; PCA; KPCA;
others
(i) Hybrid AANN-KPCA
Kalman filter
LPC
ARMA
(i) AR; MA; EMD
Data correction methods
Nadaraya-Watson statistical estimator
Interpolation
Smoothing
Data mining techniques
Data reconciliation techniques
Checking of the status of the sensors
Checking of the duration after sensor maintenance
Data context classification
Calibration of measuring systems
Other assisting techniques or tools Uncertainty consideration
Grey models
(i) GBM; dynamic uncertainty estimation of
self-validating sensor
VRFV method
These different approaches are based on either mathematical
methods, for example, statistical or probabilistic methods, or
complex analysis, for example, artificial intelligence methods.
According to [49], the data validation methods may be
classified in several types of methods, which are presented in
Figure 2.
As depicted in Figure 2, the faulty data detection methods, used to detect failures on the sensors’ signal, may
include ANNs, dimensional reduction methods, instance
based methods, probabilistic and statistical methods, and
Bayesian methods. On the contrary, the data correction
methods include the following methods: filtering, regression,
estimation, interpolation, smoothing, data mining, and data
reconciliation. These methods work specifically with the
sensors’ data and the selection of the methods that can
be applied by a system should consider the system’s usage
scenarios.
Finally, the other assisting techniques or tools are mainly
related to detection of problems originated by either hardware components or its working environment. In addition,
on real-time systems, these problems should be verified
constantly to prevent the existence of failures in the data
captured.
Description
Consisting of the detection of faulty or
incorrect values discovered during the
data acquisition and processing stages
Consisting of the estimation of faulty or
incorrect values obtained during the data
acquisition and processing stages
These are different approaches created for
the correct validation of the data
4. Applicability of the Sensor Data
Validation Methods
Mobile devices have a plethora of sensors available for the
measurement of several parameters, including the identification of the ADLs. Examples of these sensors include the
accelerometer, the gyroscope, the magnetometer, the GPS,
and the microphone.
The data acquisition using accelerometers may fail
because of several problems, including problems related with
the internal electronic amplifier of the Integrated Electronic
Piezoelectric (IEPE) device, the exposure to temperatures
beyond the accelerometer working range, failure related with
electrical components, capture of environmental noise, the
multitasking and multithreading capabilities of the mobile
devices that may cause irregular sampling rates, the positioning of the accelerometer, the low processing and memory
power, and the battery consumption [50]. The causes of
failure of an accelerometer are similar to the causes of the
failure of a gyroscope, a magnetometer, or a microphone [51].
In addition, the GPS has another failure cause, which consists
of the low connectivity of satellites in indoor environments
[52].
47
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
6
Journal of Sensors
MLP
AANN
Artificial neural networks
BP
Hybrid AANN-KPCA
PCA
KPCA
Hybrid AANN-KPCA
Dimensionality reduction
Instance based
SOM
SPRT
Faulty data
detection
methods
Probabilistic Causal Methods
Learning Algorithms
Probabilistic
Propagation in Trees
ASV
Statistic
HSV
Artificial intelligence
Fuzzy logic
SVM
RVM
Bayesian Networks
Sparse Bayesian Learning
Bayesian
Gaussian distributions
Data
validation techniques
Filtering
Kalman filter
LPC
ARMA
Regression
Data
correction
methods
AR
MA
EMD
Nadaraya-Watson statistical estimator
Interpolation
Smoothing
Data mining techniques
Data reconciliation techniques
Checking the status of the sensors
Checking the duration after sensors maintenance
Data context classification
Other
assisting
techniques
or tools
Calibration of measuring systems
Grey models
Uncertainty consideration
Dynamic uncertainty estimation of selfvalidating sensors
VRFV method
Grey models
Ensemble
GBM
Figure 2: Different categories of the data validation methods.
The validation of the data is important, but, for critical
systems, for example, clinical systems, not only should the
input data be validated, but also the results should be validated to guarantee the reliability, accuracy and, consequently,
acceptance of the system. The validation of the system may
consist of the detection of failures and the methods that may
48
be applied are the faulty detection methods. As presented
in Section 3, the methods that may be included in this
category are probabilistic and statistical methods, among
others, which may be used to validate the results of the
system. This validation can be performed by comparing the
results obtained by an equivalent system which is considered
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7
Journal of Sensors
to be a gold standard [53] with the results obtained by
the developed methods implemented by different sensors or
devices, for example, a mobile device.
Once estimated the initial error of the system, that is, how
different the obtained results are from the results obtained
by the gold standard system, the validation of the results of
the system consist of three steps, such as the definition of
the confidence level needed for the acceptance of the system,
the determination of the minimum number of experiments
needed to validate the application with confidence level
defined, and the validation of the results when compared
to a golden-standard [54]. The definition of the degree of
confidence of the system is a choice of the development
team. The system design leader may define what system
needs to have a maximum 5% error 95% of the times.
Using these parameters, a minimum number of calibration
experiments need to be performed to allow the fine tuning
of the algorithm. The minimum number of experiments may
be measured by several statistical tests, for example, Student’s
𝑡-test [55].
After the calibration of the algorithms in the system,
further tests and comparison with golden-standard systems
can be done to insure that the results reported by the
system have a 5% maximum error when compared to the
golden standard results, for 95% of the time. Note that
the 5% and 95% values are merely indicative. Moreover,
the data collection stage must hold into consideration the
limits for the optimal functioning of the sensors. As these
limits are extremely dependent on the task the sensors must
perform, we do not discuss them in this paper, for example,
if the application is supposed to track the movements of a
sportsperson in an open environment, it is possible that a
thermal sensor reports an environment temperature of −5∘ C,
yet, for an application that tracks the indoor activity of an
elder, such value should raise an alarm. In this extreme case,
it is even possible that more robust systems need to contain
different types of sensors.
5. Conclusion
The validation of the data collected by sensors in a mobile
device is an important issue for two main reasons: the first
one is the increasing number of devices and the applications
that make use of the devices’ sensors; the other is that also
increasingly users rely on these devices and applications to
collect information and make decisions that may be critical
for the user’s life and well-being.
Despite the fact that there is a wide array and types of
data validation algorithms, there is also a lack of published
information on the validity of many mobile applications.
Also, it is impossible to present a critical comparison of the
discussed methods, even within their respective categories,
as their efficiency is extremely dependent on their particular
usage; for example, the efficiency of a specific method may
be very dependent on the number and type of features
the algorithm selects on the signal to be processed, and of
course these features are chosen in view of the intended
purpose of the application. Additionally, it is possible that
even with the same chosen method and the same chosen set of
features, different authors report different efficiency ratios; for
example, their base population sample varies in size and/or
type using different population sizes or using populations that
are homogenous in age (elders or youngsters).
This paper has presented a discussion on the different
types of data validation methods such as faulty data detection,
data correction, and assisting techniques or tools. Furthermore, a classification of these methods in accordance with
its functionalities was provided. Finally, the relevance of the
data validation methods for critical systems in terms of its
reliability, accuracy and acceptance was highlighted. Complementary studies should be addressed aiming at providing an
overview on the use of valid data for the identification of the
ADLs.
Competing Interests
The authors declare that there is no conflict of interests
regarding the publication of this paper.
Acknowledgments
This work was supported by FCT project UID/EEA/50008/
2013 (Este trabalho foi suportado pelo projecto FCT UID/
EEA/50008/2013). The authors would also like to acknowledge the contribution of the COST Action IC1303 Architectures, Algorithms and Protocols for Enhanced Living
Environments (AAPELE).
References
[1] A. Braun, R. Wichert, A. Kuijper, and D. W. Fellner, “A benchmarking model for sensors in smart environments,” in Ambient
Intelligence: European Conference, (AmI ’14), Eindhoven, The
Netherlands, November 2014. Revised Selected Papers, E. Aarts,
B. de Ruyter, P. Markopoulos et al., Eds., pp. 242–257, Springer,
Cham, Switzerland, 2014.
[2] J. Garza-Ulloa, H. Yu, and T. Sarkodie-Gyan, “A mathematical
model for the validation of the ground reaction force sensor in
human gait analysis,” Measurement, vol. 45, no. 4, pp. 755–762,
2012.
[3] M. À. Cugueró, M. Christodoulou, J. Quevedo, V. Puig, D.
Garcı́a, and M. P. Michaelides, “Combining contaminant event
diagnosis with data validation/reconstruction: application to
smart buildings,” in Proceedings of the 22nd Mediterranean
Conference on Control and Automation (MED ’14), pp. 293–298,
June 2014.
[4] Y. Chen, J. Yang, and S. Jiang, “Data validation and dynamic
uncertainty estimation of self-validating sensor,” in Proceedings
of the IEEE International Instrumentation and Measurement
Technology Conference (I2MTC ’15), pp. 405–410, Pisa, Italy,
May 2015.
[5] S. Sun, J. Bertrand-krajewski, A. Lynggaard-Jensen et al., “Literature review for data validation methods,” SciTechnol, vol. 47,
no. 2, pp. 95–102, 2011.
[6] J.-L. Bertrand-Krajewski, S. Winkler, E. Saracevic, A. Torres,
and H. Schaar, “Comparison of and uncertainties in raw sewage
COD measurements by laboratory techniques and field UVvisible spectrometry,” Water Science and Technology, vol. 56, no.
11, pp. 17–25, 2007.
49
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
8
[7] N. Branisavljević, Z. Kapelan, and D. Prodanović, “Improved
real-time data anomaly detection using context classification,”
Journal of Hydroinformatics, vol. 13, no. 3, pp. 307–323, 2011.
[8] M. H. Hassoun, Fundamentals of Artificial Neural Networks,
MIT Press, 1995.
[9] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training algorithm
for optimal margin classifiers,” in Proceedings of the 5th Annual
Workshop on Computational Learning Theory, pp. 144–152,
ACM Press, Pittsburgh, Pa, USA, 1992.
[10] M. Mourad and J.-L. Bertrand-Krajewski, “A method for automatic validation of long time series of data in urban hydrology,”
Water Science and Technology, vol. 45, no. 4-5, pp. 263–270,
2002.
[11] G. Olsson, M. K. Nielsen, Z. Yuan, and A. Lynggaard-Jensen,
Instrumentation, Control and Automation in Wastewater Systems, IWA, London, UK, 2005.
[12] F. Edthofer, J. Van den Broeke, J. Ettl, W. Lettl, and A. Weingartner, “Reliable online water quality monitoring as basis for fault
tolerant control,” in Proceedings of the 1st Conference on Control
and Fault-Tolerant Systems (SysTol ’10), pp. 57–62, October 2010.
[13] S. J. Qin and W. Li, “Detection, identification, and reconstruction of faulty sensors with maximized sensitivity,” AIChE
Journal, vol. 45, no. 9, pp. 1963–1976, 1999.
[14] M. A. Kramer, “Nonlinear principal component analysis using
autoassociative neural networks,” AIChE Journal, vol. 37, no. 2,
pp. 233–243, 1991.
[15] R. Sharifi and R. Langari, “A hybrid AANN-KPCA approach to
sensor data validation,” in Proceedings of the 7th Conference on
7th WSEAS International Conference on Applied Informatics and
Communications, vol. 7, pp. 85–91, 2007.
[16] C. C. Castello, J. R. New, and M. K. Smith, “Autonomous
correction of sensor data applied to building technologies
using filtering methods,” in Proceedings of the 1st IEEE Global
Conference on Signal and Information Processing (GlobalSIP ’13),
pp. 121–124, December 2013.
[17] M. Kasinathan, B. S. Rao, N. Murali, and P. Swaminathan,
“An artificial neural network approach for the discordance
sensor data validation for SCRAM parameters,” in Proceedings
of the 1st International Conference on Advancements in Nuclear
Instrumentation, Measurement Methods and their Applications
(ANIMMA ’09), pp. 1–5, Marseille, France, June 2009.
[18] E. V. Zabolotskikh, L. M. Mitnik, L. P. Bobylev, and O. M.
Johannessen, “Neural networks based algorithms for sea surface
wind speed retrieval using SSM/I data and their validation,” in
Proceedings of the IEEE International Geoscience and Remote
Sensing Symposium (IGARSS ’99), vol. 2, pp. 1010–1012, Hamburg, Germany, June-July 1999.
[19] E. Gaura, R. Newman, M. Kraft, A. Flewitt, and W. de Lima
Monteiro, Smart MEMS and Sensor Systems, World Scientific,
Singapore, 2006.
[20] P. H. Ibarguengoytia, Any Time Probabilistic Sensor Validation,
University of Salford, Salford, UK, 1997.
[21] M. E. Tipping, “Sparse Bayesian learning and the relevance
vector machine,” Journal of Machine Learning Research, vol. 1,
no. 3, pp. 211–244, 2001.
[22] S. Alag, A. M. Agogino, and M. Morjaria, “A methodology for
intelligent sensor measurement, validation, fusion, and fault
detection for equipment monitoring and diagnostics,” Artificial
Intelligence for Engineering Design, Analysis and Manufacturing,
vol. 15, no. 4, pp. 307–320, 2001.
50
Journal of Sensors
[23] E. Shi, “An improved real-time adaptive Kalman filter for
low-cost integrated GPS/INS navigation,” in Proceedings of
the International Conference on Measurement, Information and
Control (MIC ’12), vol. 2, pp. 1093–1098, May 2012.
[24] Z. Wu and N. E. Huang, “Ensemble empirical mode decomposition: a noise-assisted data analysis method,” Advances in
Adaptive Data Analysis, vol. 1, no. 1, pp. 1–41, 2009.
[25] S. J. Wellington, J. K. Atkinson, and R. P. Sion, “Sensor validation
and fusion using the Nadaraya-Watson statistical estimator,” in
Proceedings of the 5th International Conference on Information
Fusion, vol. 1, pp. 321–326, Annapolis, Md, USA, 2002.
[26] K. E. Holbert, A. S. Heger, and N. K. Alang-Rashid, “Redundant
sensor validation by using fuzzy logic,” Nuclear Science and
Engineering, vol. 118, no. 1, pp. 54–64, 1994.
[27] Y. Ai, X. Sun, C. Zhang, and B. Wang, “Research on sensor
data validation in aeroengine vibration tests,” in Proceedings
of the International Conference on Measuring Technology and
Mechatronics Automation (ICMTMA ’10), vol. 3, pp. 162–166,
March 2010.
[28] F. Pfaff, B. Noack, and U. D. Hanebeck, “Data validation in the
presence of stochastic and set-membership uncertainties,” in
Proceedings of the 16th International Conference on Information
Fusion (FUSION ’13), pp. 2125–2132, Istanbul, Turkey, July 2013.
[29] M. Staroswiecki, “Intelligent sensors: a functional view,” IEEE
Transactions on Industrial Informatics, vol. 1, no. 4, pp. 238–249,
2005.
[30] P. H. Ibargiengoytia, L. E. Sucar, and S. Vadera, “Real time intelligent sensor validation,” IEEE Transactions on Power Systems,
vol. 16, no. 4, pp. 770–775, 2001.
[31] J. Rivera-Mejýa, E. Arzabala-Contreras, and Á. G. Leýn-Rubio,
“Approach to the validation function of intelligent sensors based
on error’s predictors,” in Proceedings of the IEEE Instrumentation and Measurement Technology Conference (I2MTC ’10), pp.
1121–1125, IEEE, May 2010.
[32] M. P. Henry and D. W. Clarke, “The self-validating sensor: rationale, definitions and examples,” Control Engineering Practice,
vol. 1, no. 4, pp. 585–610, 1993.
[33] B. Mounika, G. Raghu, S. Sreelekha, and R. Jeyanthi, “Neural network based data validation algorithm for pressure
processes,” in Proceedings of the International Conference on
Control, Instrumentation, Communication and Computational
Technologies (ICCICCT ’14), pp. 1223–1227, July 2014.
[34] Q. Wang, Z. Shen, and F. Zhu, “A multifunctional self-validating
sensor,” in Proceedings of the IEEE International Instrumentation
and Measurement Technology Conference (I2MTC ’13), pp. 1283–
1288, Minneapolis, Minn, USA, May 2013.
[35] Z. Shen and Q. Wang, “Data validation and validated uncertainty estimation of multifunctional self-validating sensors,”
IEEE Transactions on Instrumentation and Measurement, vol.
62, no. 7, pp. 2082–2092, 2013.
[36] Z. Shen and Q. Wang, “Data validation and confidence of selfvalidating multifunctional sensor,” in Proceedings of the Sensors,
pp. 1–4, Taipei, Taiwan, October 2012.
[37] J. Doyle, A. Kealy, J. Loane et al., “An integrated home-based
self-management system to support the wellbeing of older
adults,” Journal of Ambient Intelligence and Smart Environments,
vol. 6, no. 4, pp. 359–383, 2014.
[38] T. Kohonen, “The self-organizing map,” Proceedings of the IEEE,
vol. 78, no. 9, pp. 1464–1480, 1990.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Journal of Sensors
[39] B. Lamrini, E.-K. Lakhal, M.-V. Le Lann, and L. Wehenkel,
“Data validation and missing data reconstruction using selforganizing map for water treatment,” Neural Computing and
Applications, vol. 20, no. 4, pp. 575–588, 2011.
[40] A. Pantelopoulos and N. G. Bourbakis, “A survey on wearable
sensor-based systems for health monitoring and prognosis,”
IEEE Transactions on Systems, Man and Cybernetics Part C:
Applications and Reviews, vol. 40, no. 1, pp. 1–12, 2010.
[41] S. Helal, J. W. Lee, S. Hossain, E. Kim, H. Hagras, and D. Cook,
“Persim—simulator for human activities in pervasive spaces,”
in Proceedings of the 7th International Conference on Intelligent
Environments (IE ’11), pp. 192–199, Nottingham, UK, July 2011.
[42] H. Eren, “Assessing the health of sensors using data historians,”
in Proceedings of the IEEE Sensors Applications Symposium (SAS
’12), pp. 208–211, February 2012.
[43] S. Oonk, F. J. Maldonado, and T. Politopoulos, “Distributed
intelligent health monitoring with the coremicro Reconfigurable Embedded Smart Sensor Node,” in Proceedings of the
IEEE AUTOTESTCON, pp. 233–238, Anaheim, Calif, USA,
September 2012.
[44] C.-M. Chen, R. Kwasnicki, B. Lo, and G. Z. Yang, “Wearable
tissue oxygenation monitoring sensor and a forearm vascular
phantom design for data validation,” in Proceedings of the 11th
International Conference on Wearable and Implantable Body
Sensor Networks (BSN ’14), pp. 64–68, June 2014.
[45] F. J. Maldonado, S. Oonk, and T. Politopoulos, “Enhancing
vibration analysis by embedded sensor data validation technologies,” IEEE Instrumentation and Measurement Magazine,
vol. 16, no. 4, pp. 50–60, 2013.
[46] E. Miluzzo, Smartphone Sensing, Dartmouth College, Hanover,
New Hampshire, 2011.
[47] F. J. Maldonado, S. Oonk, and T. Politopoulos, “Optimized
neuro genetic fast estimator (ONGFE) for efficient distributed
intelligence instantiation within embedded systems,” in Proceedings of the International Joint Conference on Neural Networks
(IJCNN ’13), pp. 1–8, August 2013.
[48] B. Wallace, R. Goubran, F. Knoefel et al., “Automation of
the validation, anonymization, and augmentation of big data
from a multi-year driving study,” in Proceedings of the IEEE
International Congress on Big Data, pp. 608–614, New York, NY,
USA, June 2015.
[49] J. Brownlee, A Tour of Machine Learning Algorithms, Machine
Learning Mastery, 2013.
[50] R. Denton, “Sensor Reliability Impact on Predictive Maintenance Program Costs,” 2010, Wilcoxon Research.
[51] O. Brand, G. K. Fedder, C. Hierold, J. G. Korvink, O. Tabata,
and T. Tsuchiya, Reliability of MEMS: Testing of Materials and
Devices, John Wiley & Sons, New York, NY, USA, 2013.
[52] G. Heredia, A. Ollero, M. Bejar, and R. Mahtani, “Sensor
and actuator fault detection in small autonomous helicopters,”
Mechatronics, vol. 18, no. 2, pp. 90–99, 2008.
[53] G. J. Prescott and P. H. Garthwaite, “A simple Bayesian analysis
of misclassified binary data with a validation substudy,” Biometrics, vol. 58, no. 2, pp. 454–458, 2002.
[54] V. R. Basili, R. W. Selby Jr., and T.-Y. Phillips, “Metric analysis
and data validation across fortran projects,” IEEE Transactions
on Software Engineering, vol. 9, no. 6, pp. 652–663, 1983.
[55] K. R. Sundaram and A. Jose, “Teaching: estimation of minimum
sample size and the impact of effect size and altering the
type-I & II errors on IT, in clinical research,” in Data and
Context in Statistics Education: Towards an Evidence-Based
9
Society. Proceedings of the Eighth International Conference on
Teaching Statistics (ICOTS8, July, 2010), Ljubljana, Slovenia, C.
Reading, Ed., International Statistical Institute, Voorburg, The
Netherlands, 2010.
51
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
52
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
3. Audio Fingerprinting Techniques for detection of
the environment of Activities of Daily Living (ADL):
A Systematic Review
The following article is the third part of the chapter 2.
Audio Fingerprinting Techniques for detection of the environment of Activities of Daily Living
(ADL): A Systematic Review
Ivan Miguel Pires, Rui Santos, Nuno Pombo, Nuno M. Garcia, Francisco Flórez-Revuelta,
Susanna Spinsante, Rossitza Goleva and Eftim Zdravevski
Sensors (MDPI), published, 2018.
According to 2016 Journal Citation Reports published by Thomson Reuters in 2017, this journal
has the following performance metrics:
ISI Impact Factor (2016): 2.677
ISI Article Influence Score (2016): 0.6
Journal Ranking (2016): 104/642 (Electrical and Electronic Engineering)
53
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
54
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
sensors
Review
Recognition of Activities of Daily Living Based on
Environmental Analyses Using Audio Fingerprinting
Techniques: A Systematic Review
Ivan Miguel Pires 1,2,3 ID , Rui Santos 1,3 , Nuno Pombo 1,3,4 , Nuno M. Garcia 1,3,4, *
Francisco Flórez-Revuelta 5 , Susanna Spinsante 6 ID , Rossitza Goleva 7 ID
and Eftim Zdravevski 8 ID
1
2
3
4
5
6
7
8
*
ID
,
Instituto de Telecomunicações, Universidade da Beira Interior, 6201-001 Covilhã, Portugal;
impires@it.ubi.pt (I.M.P.); rui_17_santos@hotmail.com (R.S.); ngpombo@ubi.pt (N.P.)
Altranportugal, 1990-096 Lisbon, Portugal
ALLab—Assisted Living Computing and Telecommunications Laboratory, Computing Science Department,
Universidade da Beira Interior, 6201-001 Covilhã, Portugal
ECATI, Universidade Lusófona de Humanidades e Tecnologias, 1749-024 Lisbon, Portugal
Department of Computer Technology, Universidad de Alicante,
03690 Sant Vicent del Raspeig, Alicante, Spain; francisco.florez@ua.es
Department of Information Engineering, Marche Polytechnic University, 60121 Ancona, Italy;
s.spinsante@univpm.it
Department of Informatics, New Bulgarian University, 1618 g.k. Ovcha kupel 2 Sofia, Bulgaria;
rgoleva@gmail.com
Faculty of Computer Science and Engineering, University Ss Cyril and Methodius, 1000 Skopje, Macedonia;
eftim.zdravevski@finki.ukim.mk
Correspondence: ngarcia@di.ubi.pt; Tel.: +351-966-637-9785
Received: 28 November 2017; Accepted: 5 January 2018; Published: 9 January 2018
Abstract: An increase in the accuracy of identification of Activities of Daily Living (ADL) is very
important for different goals of Enhanced Living Environments and for Ambient Assisted Living (AAL)
tasks. This increase may be achieved through identification of the surrounding environment. Although
this is usually used to identify the location, ADL recognition can be improved with the identification
of the sound in that particular environment. This paper reviews audio fingerprinting techniques that
can be used with the acoustic data acquired from mobile devices. A comprehensive literature search
was conducted in order to identify relevant English language works aimed at the identification of the
environment of ADLs using data acquired with mobile devices, published between 2002 and 2017.
In total, 40 studies were analyzed and selected from 115 citations. The results highlight several audio
fingerprinting techniques, including Modified discrete cosine transform (MDCT), Mel-frequency
cepstrum coefficients (MFCC), Principal Component Analysis (PCA), Fast Fourier Transform (FFT),
Gaussian mixture models (GMM), likelihood estimation, logarithmic moduled complex lapped
transform (LMCLT), support vector machine (SVM), constant Q transform (CQT), symmetric pairwise
boosting (SPB), Philips robust hash (PRH), linear discriminant analysis (LDA) and discrete cosine
transform (DCT).
Keywords: acoustic sensors; fingerprint recognition; data processing; artificial intelligence; mobile
computing; signal processing algorithms; systematic review; Activities of Daily Living (ADL)
Sensors 2018, 18, 160; doi:10.3390/s18010160
www.mdpi.com/journal/sensors
55
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
2 of 23
1. Introduction
The identification of Activities of Daily Living (ADL) [1] is of utmost importance to build Enhanced
Living Environment and Ambient Assisted Living solutions [2,3], or to allow the development of
Personal Digital Life Coaching systems [4]. To achieve this, several authors have proposed the
development of solutions based on mobile devices (e.g., smartphones) [5–8] for several reasons,
the most prominent being the adoption ratios of these devices, its increasing computing power and
memory, and the fact that these devices already come equipped with a plethora of sensors that can be
used to sense and feed data to ADL identification systems.
Despite the increasing complexity of ADL identification systems, the recognition of the
surrounding environment is limited because of the restrictions of some location sensors. For instance,
Global Positioning System (GPS) sensors, can only be reliably and accurately used in outdoor scenarios.
Likewise, proximity sensors, radar sensors, Passive Infra-Red (PIR) sensors and alike require significant
installation effort, thus are not widely used in real scenarios which require ADL identification.
As proposed in previous works [9–11], an ADL identification framework should also be able to
integrate data from the sound of the environment into the ADL identification module in order to allow
the system to sense the environmental sounds, to determine the type of environment, and to increase
the accuracy of the overall ADL identification solution.
Most mobile devices are equipped with a microphone that can capture an acoustic signal.
This signal can be processed using audio fingerprinting techniques, allowing the system to find
a match between the collected signal and a database of well-known audio fingerprints. This might
facilitate an increase in the accuracy of recognition of the environment where ADLs are performed.
Several methods may be used to carry out audio fingerprinting, performing the pre-processing
of the acoustic data (e.g., Fast Fourier Transform (FFT)), extracting relevant features, and after that,
obtaining a classification or recognition (e.g., Support Vector Machine (SVM)).
This review summarizes the existing methods in the literature related to audio fingerprinting
techniques for the application in a system that uses mobile technology for the recognition of
the environment. While acknowledging that the methods here presented are very diverse and have
been tested with different data sets and different feature extraction techniques, in order to estimate
which method may provide better results in a mobile computational device, this paper also presents a
comparison between the different methods and features.
The remainder of this paper is organized as follows: Section 2 presents the methodology for
this review; the methods discovered in the literature are presented in Section 3; Section 4 discusses
different methods, and finally, Section 5 present conclusions of this review.
2. Methodology
2.1. Research Questions
The primary questions of this review were as follows: (RQ1) What is audio fingerprinting? (RQ2)
Which audio fingerprinting techniques are useful to identify the environment of daily activities? (RQ3)
Which are the audio fingerprinting techniques feasible for their use in mobile devices?
2.2. Inclusion Criteria
Studies assessing ADLs using audio fingerprinting techniques were included in this review if they
met the following criteria: (1) audio fingerprinting techniques adapted to mobile devices; (2) audio
fingerprinting techniques used for the detection of the environment of ADL; (3) using mobile devices;
(4) the accuracies of the audio fingerprinting techniques presented are reported; (5) were published
between 2002 and 2017; and (6) were written in English.
56
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
3 of 23
2.3. Search Strategy
The team searched for studies meeting the inclusion criteria in the following electronic databases:
IEEE Xplore, and ACM Digital Library. Every study was independently evaluated by eight reviewers
(IP, RS, NP, NG FR, SP, RG and EZ), and its suitability was determined with the agreement of all parties.
The studies were examined to identify the characteristics of audio fingerprint and its suitability for
application with mobile devices for the identification of ADL.
2.4. Extraction of Study Characteristics
The following data was extracted from the studies and tabulated (see Tables 1 and 2): year of
publication, population for the application of the algorithm, purpose of the study, devices used, and
study outcomes of the algorithm for audio fingerprinting. For all cited studies in Tables 1 and 2,
the experiments were conducted in laboratory settings. We additionally verified whether the raw data
and source code are available, either publically or per request, by emailing the corresponding author
of each study.
Table 1. Study Analysis.
Devices
Raw Data
Available
Source
Code
Available
To search for audio in
the database by the
content rather than by
name
Mobile
Phone
(Android)
No
No
100,000 MP3
fragments
To create an MP3 sniffer
system that includes
audio fingerprinting
Not
mentioned
Yes
Only for
feature
extraction
2011
10,000 MP3
fragments
Proposes an MP3
fingerprint system for
the recognition of
several clips
Not
mentioned
The same
data as [13]
The same
source code
as [13]
Tsai et al. [15]
2016
Multi-channel audio
recordings of 75 real
research group
meetings,
approximately 72 h
of meetings in total
Proposes an adaptive
audio fingerprint based
on spectrotemporal
eigenfilters
Mobile
phones,
tablets or
laptop
computers
Yes
No
Casagranda et al.
[16]
2015
1024 samples
Proposes an audio
fingerprinting method
that uses GPS and
acoustic fingerprints
Smartphone
No
No
Approximately
1,518,177 min
(25,303 h) of songs
Proposes a method to
accelerate audio
fingerprinting
techniques by skipping
the search for irrelevant
signal sections
Not
mentioned
Yes
No
1062 10 s clips
Proposes a method to
analyze and classify
daily activities in
personal audio
recordings
Not
mentioned
Yes
No
Year of
Publication
Population
Sui et al. [12]
2014
2500 pieces of 8 s
advertisement
audios, and
randomly select
200 pieces of audio
in the existing
database and
50 pieces of other
irrelevant audio as
test audio
Liu [13]
2012
Liu et al. [14]
Paper
Purpose of the Study
ACM
IEEE
Nagano et al.
[17]
Ziaei et al. [18]
2015
2015
57
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
4 of 23
Table 1. Cont.
Paper
Year of
Publication
Population
Purpose of the Study
Devices
Raw Data
Available
Source
Code
Available
George et al.
[19]
2015
1500 audio files
Proposes an audio
fingerprinting method
based on landmarks in
the audio spectrogram
Computer
No
No
Kim et al. [20]
2015
6000 television
advertisements with
a total time of 1110 h
Proposes a television
advertisement search
based on audio
fingerprinting in real
environments
Television
No
No
Seo [21]
2014
1000 songs with
classic, jazz, pop,
rock, and hip-hop
Proposes a binary audio
fingerprint matching,
using auxiliary
information
Not
mentioned
No
No
Rafii et al. [22]
2014
Several songs with a
duration between
6 and 9 s
Proposes an audio
fingerprinting method
for recognition of some
clips
Computer
and
Smartphone
No
No
Not
mentioned
No
No
Naini et al. [23]
2014
1000 songs
Proposes an audio
fingerprinting method
based on maximization
of the mutual
information across the
distortion channel
Yang et al. [24]
2014
200,000 songs
Proposes a music
identification system
based on space-saving
audio fingerprints
Not
mentioned
No
No
Yin et al. [25]
2014
958 randomly
chosen query
excerpts
Proposes an audio
fingerprinting
algorithm that uses
compressed-domain
spectral entropy
Not
mentioned
No
No
Wang et al. [26]
2014
100,000 songs
Proposes an audio
fingerprinting method
that uses GPUs
Not
mentioned
No
No
3000 TV
advertisements
Proposes a
high-performance
audio fingerprint
extraction method for
identifying Television
commercial
advertisement
Television
No
No
Smartphone,
tablet,
notebook,
desktop, or
another
mobile
device
No
No
Lee et al. [27]
Shibuya et al.
[28]
2013
1374 television
programs
(792 h in total)
Proposes a method of
identifying media
content from an audio
signal recorded in
reverberant and noisy
environments using a
mobile device
Bisio et al. [29]
2013
20 sounds
Proposes the Improved
Real-Time TV-channel
Recognition (IRTR)
method
Smartphone
No
No
2013
1000 songs as
positive samples
and 999 songs as
negatives
Proposes a method that
speeds up the search
process, reducing the
number of database
accesses
Not
mentioned
No
No
Lee et al. [30]
58
2014
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
5 of 23
Table 1. Cont.
Paper
Year of
Publication
Population
Purpose of the Study
Devices
Raw Data
Available
Source
Code
Available
Bisio et al. [31]
2012
100,000 songs
Proposes an audio
fingerprint algorithm
adapted to mobile
devices
Smartphone
No
No
Several datasets
Proposes an audio
fingerprinting
algorithm that encodes
the local spectral
energies around salient
points selected among
the main spectral peaks
in a given signal
Not
mentioned
No
No
300 real-world
recordings in a
living room
Proposes an audio
fingerprinting method
that combines the
Fingerprinting
technique with
Generalized cross
correlation
iPad
No
No
20 music clips
with 5 s
Proposes an audio
fingerprinting
algorithm for
recognition of some
clips
Not
mentioned
No
No
835 popular songs
Proposes an audio
fingerprinting
algorithm based on
dynamic subband
locating and
normalized spectral
subband centroid (SSC)
Not
mentioned
No
No
Not
mentioned
No
No
Anguera et al.
[32]
Duong et al. [33]
Wang et al. [34]
Xiong et al. [35]
2012
2012
2012
2012
Deng et al. [36]
2011
100 audio files
Proposes an audio
fingerprinting
algorithm based on
harmonic enhancement
and SSC of audio signal
Pan et al. [37]
2011
62-h audio database
of 1000 tracks
Proposes an audio
feature in spectrum,
local energy centroid,
for audio fingerprinting
Not
mentioned
No
No
2011
3600 s of several
real-time tests
Presents an audio
fingerprinting method
with a low-cost
embedded
reconfigurable platform
Computer
No
No
1000 songs
Proposes an indexing
scheme and a search
algorithm based on the
index
Computer
No
Only
pseudo-code
for
fingerprint
matching
7500 experiments
Proposes an audio
fingerprinting method
for the recognition of
some clips
Computer
No
No
500 popular songs
Proposes an audio
fingerprinting method
using sub-fingerprint
masking based on the
predominant pitch
extraction
Mobile
devices
Yes
No
Martinez et al.
[38]
Cha [39]
Schurmann et al.
[40]
Son et al. [41]
2011
2011
2010
59
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
6 of 23
Table 1. Cont.
Paper
Chang et al. [42]
Umapathy et al.
[43]
Kim et al. [44]
Sert et al. [45]
Ramalingam et al.
[46]
Ghouti et al. [47]
60
Population
Purpose of the Study
Devices
Raw Data
Available
Source
Code
Available
17,208 audio clips
Presents a sub-Nyquist
audio fingerprinting
system for music
recognition, which
utilizes Compressive
Sampling (CS) theory
Not
mentioned
No
No
2007
213 audio signals
Proposes an audio
feature extraction and a
multi-group
classification using the
local discriminant bases
(LDB) technique
Not
mentioned
No
No
2007
100 Korean
broadcast TV
programs
Proposes an audio
fingerprinting method
for identification of
bookmarked audio
segments
Computer
No
No
2006
approximately
45 min of pop, rock,
and country songs
Proposes an audio
fingerprinting method
from the most
representative section of
an audio clip
Not
mentioned
No
No
2006
250 audio files
Proposes and audio
fingerprinting method
using several features
Not
mentioned
No
No
2006
Two audio contents
perceptually similar
Proposes an audio
fingerprinting
algorithm that uses
balanced multiwavelets
(BMW)
Not
mentioned
No
No
PDA or
computer
Yes
No
Year of
Publication
2010
Cook et al. [48]
2006
7,106,069
fingerprints
Proposes an audio
fingerprinting
algorithm for the fast
indexing and searching
of a metadata database
Seo et al. [49]
2005
8000 classic, jazz,
pop, rock, and
hip-hop songs
Proposes an audio
fingerprinting method
based on normalized
SSC
Not
mentioned
No
No
Not
mentioned
No
No
Not
mentioned
No
No
Haitsma et al.
[50]
2003
256 sub-fingerprints
Proposes to solve larger
speed changes by
storing the fingerprint
at multiple speeds in
the database or
extracting the
fingerprint query at
multiple speeds and
then to perform
multiple queries on the
database
Haitsma et al.
[51]
2002
256 sub-fingerprints
Proposes an audio
fingerprinting system
for recognition of some
clips
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
7 of 23
Table 2. Study summaries.
Paper
Outcomes
ACM
Sui et al. [12]
The authors propose a two-level audio fingerprint retrieval algorithm to satisfy the
demand of accurate and efficient search for advertisement audio. Based on clips with 8 s of
advertisements, the authors build a database with 2500 audio fingerprints. The results
show that the algorithm implemented with parallel processing yields a precision of 100%.
Liu [13]
The authors create an MP3 sniffer system and test it with multi-resolution local
descriptions. The system has a database of 100,000 MP3 tones and authors report that the
system has high performance, because 100 queries for identifying unknown MP3 tones
took less than 2 s to be processed
Liu et al. [14]
The authors describe an MP3 fingerprinting system that compares the normalized distance
between two MP3 fingerprints to detect a false identification. The authors identify the
possible features of the song and build a large database. For the identification, the authors
test the near neighbor searching schemes and compare with the indexing scheme, which
utilizes the PCA technique, the QUery Context (QUC)-tree, and the MP3 signatures.
The conclusions show that the system has a maximum average error equals to 4.26%.
IEEE
Tsai et al. [15]
The authors propose a method for aligning a set of overlapping meeting recordings, which
uses an audio fingerprint representation based on spectrotemporal eigenfilters that are
learned on-the-fly in an unsupervised manner. The proposed method is able to achieve
more than 99% alignment accuracy at a reasonable error tolerance of 0.1 s.
The authors propose an audio fingerprinting algorithm based on the spectral features of
Casagranda et al.
the audio samples. The authors reported that the algorithm is noise tolerant, which is a key
[16]
feature for audio based group detection.
Nagano et al.
[17]
The authors propose an approach to accelerate fingerprinting techniques and apply it to
the divide-and-locate (DAL) method. The reported results show that DAL3 can reduce the
computational cost of DAL to approximately 25%.
Ziaei et al. [18]
The authors propose a method to analyze and classify daily activities in personal audio
recordings (PARs), which uses speech activity detection (SAD), speaker diarization, and a
number of audio, speech and lexical features to characterize events in daily audio streams.
The reported overall accuracy of the method is approximately 82%.
George et al.
[19]
The authors propose an audio fingerprinting method that is tolerant to time-stretching and
is scalable. The proposed method uses three peaks in the time slice, unlike Shazam, which
uses only one. The additive noise deteriorates the lowest frequency bin, decreasing the
performance of the algorithm at higher additive noise, compared to other algorithms.
Kim et al. [20]
The authors propose a Television advertisement search based on audio peak-pair hashing
method. The reported results show that the proposed method has respectable results
compared to other methods.
Seo [21]
The authors propose an asymmetric fingerprint matching method which utilizes an
auxiliary information obtained while extracting fingerprints from the input unknown
audio. The experiments carried out with one thousand songs against various distortions
compare the performance of the asymmetric matching with the conventional Hamming
distance. Reported results suggest that the proposed method has better performance than
the conventional Hamming distance.
Rafii et al. [22]
The authors propose an audio fingerprinting system with two stages: fingerprinting and
matching. The system uses CQT and a threshold method for fingerprinting stage, and the
Hamming similarity and the Hough Transform for the matching stage, reporting an
accuracy between 61% and 81%.
Naini et al. [23]
The authors present a method for designing fingerprints that maximizes a mutual
information metric, using a greedy optimization method that relies on the information
bottleneck (IB) method. The results report a maximum accuracy around 65% in the
recognition.
Yang et al. [24]
The authors propose an efficient music identification system that utilizes a kind of
space-saving audio fingerprints. The experiments were conducted on a database of 200,000
songs and a query set of 20,000 clips compressed in MP3 format with different bit rates.
The author’s report that compared to other methods, this method reduces the memory
consumption and keeps the recall rate at approximately 98%.
61
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
8 of 23
Table 2. Cont.
Paper
Outcomes
IEEE
62
Yin et al. [25]
The authors propose a compressed-domain audio fingerprinting algorithm for MP3 music
identification in the Internet of Things. The algorithm achieves promising results on
robustness and retrieval precision rates under various time-frequency audio signal
distortions including the challenging pitch shifting and time-scale modification.
Wang et al. [26]
The authors propose parallelized schemes for audio fingerprinting over GPU. In the
experiments, the speedup factors of the landmark lookup and landmark analysis are
verified and the reported overall response time has been reduced.
Lee et al. [27]
The authors propose a salient audio peak pair fingerprint extraction based on CQT.
The reported results show that the proposed method has better results compared to other
methods, and is suitable for many practical portable consumer devices.
Shibuya et al.
[28]
The authors develop a method that uses the quadratically interpolated FFT (QIFFT) for the
audio fingerprint generation in order to identify media content from an audio signal
recorded in a reverberant or noisy environment with an accuracy around 96%.
Bisio et al. [29]
The authors present an improvement of the parameter configuration used by the Philips
audio fingerprint computation algorithm in order to reduce the computational load and
consequent energy consumption in the smartphone client. The results show a significant
reduction of computational time and power consumption of more than 90% with a limited
decrease in recognition performance.
Lee et al. [30]
The authors propose an audio fingerprint search algorithm for music retrieval from large
audio databases. The results of the proposed method achieve 80–99% search accuracy for
input audio samples of 2–3 s with signal-to-noise ratio (SNR) of 10 dB or above.
Bisio et al. [31]
The authors present an optimization of the Philips Robust Hash audio fingerprint
computation algorithm, in order to adapt it to run on a smartphone device. In the
experiments, the authors report that the proposed algorithm has an accuracy of 95%.
Anguera et al.
[32]
The authors present a novel local audio fingerprint called Masked Audio Spectral
Keypoints (MASK) that is able to encode, with few bits, the audio information of any kind
in an audio document. MASK fingerprints encode the local energy distribution around
salient spectral points by using a compact binary vector. The authors report an accuracy
around 58%.
Duong et al. [33]
The authors presented a new approach based on audio fingerprinting techniques.
The results of this study indicate that a high level of synchronization accuracy can be
achieved for a recording period as short as one second.
Wang et al. [34]
The authors present an audio fingerprinting algorithm, where the audio fingerprints are
produced based on 2-Dimagel, reporting an accuracy between 88% and 99%.
Xiong et al. [35]
The authors propose an improved audio fingerprinting algorithm based on dynamic
subband locating and normalized Spectral Subband Centroid (SSC). The authors claim that
the algorithm can recognize unknown audio clips correctly, even in the presence of severe
noise and distortion.
Deng et al. [36]
The authors propose an audio fingerprinting algorithm based on harmonic enhancement
and Spectral Subband Centroid (SSC). The authors build a database with 100 audio files,
and also implement several techniques to reduce the noise and other degradations,
proving the reliability of the method when severe channel distortion is present. The results
report an accuracy between 86% and 93%.
Pan et al. [37]
The authors propose a method for fingerprinting generation using the local energy
centroid (LEC) as a feature. They report that the method is robust to different noise
conditions and, when the linear speed is not changed, the audio fingerprint method based
on LEC obtains an accuracy of 100%, reporting better results than Shazam’s fingerprinting.
Martinez et al.
[38]
The authors present a music information retrieval algorithm based on audio fingerprinting
techniques. The size of frame windows influences the performance of the algorithm, e.g.,
the best size of the frame window for shorts audio tracks is between 32 ms to 64 ms, and
the best size of the frame window for audio tracks is 128 ms.
Cha [39]
The author proposes an indexing scheme for large audio fingerprint databases.
The method shows a higher performance than the Haitsma-Kalker method with respect to
accuracy and speed.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
9 of 23
Table 2. Cont.
Paper
Outcomes
IEEE
Schurmann et al.
[40]
The authors propose a fuzzy-cryptography scheme that is adaptable in its noise tolerance
through the parameters of the error correcting code used and the audio sample length. In a
laboratory environment, the authors utilized sets of recordings for five situations at three
loudness levels and four relative positions of microphones and audio source. The authors
derive the expected Hamming distance among audio fingerprints through 7500
experiments. The fraction of identical bits is above 0.75 for fingerprints from the same
audio context, and below 0.55 otherwise.
Son et al. [41]
The authors present an audio fingerprinting algorithm to recognize songs in real noisy
environments, which outperforms the original Philips algorithm in recognizing
polyphonic music in real similar environments.
Chang et al. [42]
The authors introduce the Compressive Sampling (CS) theory to the audio fingerprinting
system for music recognition, by proposing a CS-based sub-Nyquist audio fingerprinting
system. Authors claim that this system achieves an accuracy of 93.43% in reducing the
sampling rate and in the extraction of musical features.
Umapathy et al.
[43]
The authors present a novel local discriminant bases (LDB)-based audio classification
scheme covering a wide range of audio signals. After the experiments, the obtained results
suggest significant potential for LDB-based audio classification in auditory scene analysis
or environment detection.
Kim et al. [44]
The authors develop a system that retrieves desired bookmarked video segments using
audio fingerprint techniques based on the logarithmic modified Discrete Cosine Transform
(DCT) modulation coefficients (LMDCT-MC) feature and two-stage bit vector searching
method. The author’s state that the search accuracy obtained is 99.67%.
Sert et al. [45]
The authors propose an audio fingerprinting model based on the Audio Spectrum Flatness
(ASF) and Mel Frequency Cepstral Coefficients (MFCC) features, reporting and accuracy of
93% and 91%, respectively.
The authors propose a method to create audio fingerprints by Gaussian Mixtures using
Ramalingam et al. features extracted from the short-time Fourier transform (STFT) of the signal. The
[46]
experiments were performed on a database of 250 audio files, obtaining the highest
identification rate of 99.2% with spectral centroid.
Ghouti et al.
[47]
The authors propose a framework for robust identification of audio content by using short
robust hashing codes, which applies the forward balanced multiwavelet (BMW) to
transform each audio frame using 5 decomposition levels, and after the distribution of the
subbands’ coefficients into 32 different blocks, the estimation quantization (EQ) scheme
and the hashes are computed.
Cook et al. [48]
The authors propose a system that allows audio content identification and association of
metadata in very restricted embedded environments. The authors report that the system
has better performance than the method based on a more traditional n-dimensional
hashing scheme, but it achieves results with 2% less accuracy.
Seo et al. [49]
The authors propose an audio fingerprinting method based on the normalized Spectral
Subband Centroid (SSC), where the match is performed using the square of the Euclidean
distance. The normalized SSC obtains better results than the widely-used features, such as
tonality and Mel Frequency Cepstral Coefficients (MFCC).
Haitsma et al.
[50]
The authors present an approach to audio fingerprinting, but it has negligible effects on
other aspects, such as robustness and reliability. They proved that the developed method is
robust in case of linear speed changes.
Haitsma et al.
[51]
The authors present an approach to audio fingerprinting, in which the fingerprint
extraction is based on the extraction of a 32-bit sub-fingerprint every 11.8 millis. They also
develop a fingerprint database and implement a two-phase search algorithm, achieving an
excellent performance, and allowing the analytical modeling of false acceptance rates.
3. Results
As illustrated in Figure 1, our review identified 115 papers that included three duplicates, which
were removed. The remaining 112 works were evaluated in terms of title, abstract, and keywords,
resulting in the exclusion of 50 citations. Full text evaluation of the remaining 62 papers resulted in
the exclusion of 22 papers that did not match the defined criteria. The remaining 40 papers were
63
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
10 of 23
included in the qualitative synthesis and the quantitative synthesis. In summary, our review examined
40 papers.
Figure 1. Flow diagram of identification and inclusion of papers.
We suggest that interested readers refer to the original cited works to find relevant information
about the details of the methods analyzed in this review. Table 1 shows the year of publication,
population, purpose of the study, devices, and settings of the selected papers. Table 2 shows study
aims and results. As shown in Table 1, all studies have been performed in controlled environments
(laboratory). The major part of the studies was performed between 2011 and 2016 with a total of
29 studies (73%), where five studies were in 2011 (13%), five studies in 2012 (13%), four studies
in 2013 (10%), eight studies in 2014 (20%), six studies in 2015 (15%), and one study in 2016 (3%).
Some studies indicate the devices used: eight studies used computer microphones (23%), 10 studies
used mobile devices (25%), and two studies used a television (5%).
Methods for Audio Fingerprinting
In [12], the authors created a system that implements the framing, Fast Fourier Transform (FFT),
calculation of the spectrum modules, extraction of two kinds of audio fingerprinting, and two level
search of two kinds of audio fingerprinting. The two kinds were extracted calculating the sum of
the spectrum modulus of every frame, getting the sum of global spectrum modulus in two stages.
The authors reported that, when the signal noise rate (SNR) is 10 dB, the two level algorithm, with
parallel processing, reports a precision of 100% [12].
In [14], several MP3 features were extracted, such as the Single local description, the Multiple local
description, the Modified discrete cosine transform (MDCT), the Mel-frequency cepstrum coefficients
(MFCC), the MPEG-7 descriptors, and the chroma vectors, using the Principal Component Analysis
(PCA) technique to reduce the dimensionality and QUery Context (QUC)-tree to search for songs.
64
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
11 of 23
The tests of the methods described in [14] were performed with 10,000 MP3 fragments, reporting
a maximum average error equals to 4.26%, which represents an accuracy around 96%. In [13],
the same authors extracted the same features and used the same techniques, but they also used
the MP3 signatures for the implementation of the audio fingerprinting method, performing tests
with 100,000 MP3 fragments, which reported the errors and accuracies obtained are equals to the
reported in [14].
Tsai et al. [15] presented a method to calculate audio fingerprints with 6 steps, namely compute
spectrogram, collect context frames, apply eigenfilters, compute deltas, apply threshold, and bit
packing. The authors reported that the developed method is more robust than the three other
fingerprints (e.g., Shazam, Masked Audio Spectral Keypoints (MASK), and Panako), achieving an
accuracy of 99.2% [15].
Another audio feature named local energy centroid (LEC) is used in [37] to obtain a representation
of audio signals in noisy condition. The method for audio fingerprinting has several steps. First,
the audio is downsampled to 8 kHz and segmented into frames, and then FFT is employed to obtain the
spectrum. Later, the spectrum is optimized by applying weighted window functions with different size.
Then, the LEC is saved and the amplitude components are removed, obtaining an audio spectrum that
can be represented by sparse LEC set of coordinates [37]. The authors reported that the method is
robust to different noise conditions, and when the linear speed is not changed, the audio fingerprint
method based on LEC reports an accuracy of 100% [37].
In [36], the authors proposed an audio fingerprinting algorithm that starts with the application of
low-pass filter to the audio signal and resampling to eliminate the high-frequency noise and other audio
components that are perceptually insignificant for human auditory system. Afterwards, the audio is
framed and weighted by Window function, and the FFT is applied [36]. Next, the Spectral Subband
Centroid (SSC) is calculated and the approach of harmonic enhancement is adopted to estimate the
predominant pitch of audio signal [36]. Finally, the normalized SSC is masked by the predominant
pitch, and the proposed algorithm is resistant to some kinds of signal degradations in varying degrees,
reporting an accuracy between 86% and 93% [36]. The authors of [35] also used the normalized SSC
for the creation of an audio fingerprinting algorithm. The algorithm is structured in several phases,
such as: pre-processing, framing, implementing the FFT to transform audio signals from time to
frequency domain, implementing the dynamic subband locating, and applying the normalized SSC,
obtaining, at the end, the audio fingerprint [35]. With the fingerprints created, the authors reported an
accuracy up to 80% in normal conditions [35]. The authors of [49] also proposed an audio fingerprinting
algorithm using SSC, starting with the conversion to mono and downsampling the audio to 11,025 Hz.
After the downsampling, the audio signal is windowed by Hamming window (typically 371.5 ms)
with 50% overlap and transformed into the frequency domain using FFT [49]. Afterwards, the audio
spectrum is divided into 16 critical bands, and the frequency centroids of the 16 critical bands are used
as a fingerprint of the audio frame [49], reporting an accuracy around 60% with MP3 and Random
start, and an accuracy around 100% with Equalization.
Another algorithm is presented in [50] that consists of the modification of an existing algorithm
named Streaming Audio Fingerprinting (SAF), which apply the framing and the FFT, create energy
33 bands, and then, apply a filter and a threshold. The modification of the algorithm consists of
increasing the number of the energy bands, and three new steps between the creation of energy bands
and the application of a filter and threshold: auto-correction, filter and the creation of a subsample [50].
The authors reported that the algorithm is robust in case of linear speed changes [50].
In [28], the audio fingerprinting methods proposed has several steps, these are framing,
application of FFT or quadratically interpolated FFT (QIFFT), time averaging, peak detection, quadratic
interpolation, sinusoidal quantification, frequency-axial discretization, and time-axial warping.
A fingerprint that represents the distribution of pseudosinusoidal components in the time-frequency
domain is generated, showing results with an accuracy around 96% and precision of 100% [28].
65
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
12 of 23
In [51], the authors proposed a new fingerprint algorithm based on the streaming approach,
where the audio signal is segmented into overlapping frames, the FFT should be applied, and after
that, the Human Auditory System (HAS) is used, reporting an accuracy of 100% for the recognition
of pop-music.
In [45] is proposed a system for audio fingerprinting that starts with preprocessing and framing
of the audio signal. Afterwards, a general feature extraction paradigm, extended with a descriptor
based on structural similarity analysis with MPEG-7 Audio Spectrum Flatness (ASF), is applied to the
signal [45]. The last step, before the fingerprint construction, consists of the structural analysis that
results only the feature vector of the expressive audio piece [45]. At the end, the authors reduce the
dimension of the ASF feature vector in the fingerprint construction stage based on the MPEG-7 Audio
Signature (AS) description scheme [45], reporting an accuracy around 93%.
The authors of [46] proposed an audio fingerprinting scheme with several stages, such as
preprocessing, framing, feature extraction, Gaussian mixture models (GMM) modelling, likelihood
estimation, and comparison with a fingerprinting database. In the preprocessing stage, the audio signal
is converted to a standard format (16-bit, pulse code modulation (PCM)) [46]. In the framing stage,
the audio signals are divided into frames of length equals to 23 ms [46]. During feature extraction,
the authors used the STFT, extracting several features, such as Shannon entropy, Rényi entropy, Spectral
centroid, Spectral bandwidth, Spectral band energy, Spectral flatness measure, Spectral crest factor,
and Mel-frequency cepstral coefficients (MFCC) [46]. Afterwards, the GMM models are applied, using
the probability density function (PDF), and the Expectation-Maximization (EM) [46]. Among the
features used, spectral centroid gives the highest identification rate of 99.2% [46].
The authors of [47] presented an audio fingerprint extraction algorithm, consisting of:
downsampling of the input audio content of 3 s to obtain a sampling rate of 5512 Hz; applying
the framing division on the downsampled content using Hamming window with an overlap factor of
31/32; computing the forward balanced multiwavelet (BMW) to transform for each audio frame using
five decomposition levels; dividing the subbands’ coefficients into 32 different blocks; applying the
estimation quantization (EQ) scheme using a neighbouring window of five audio samples; computing
the log variances of the magnitudes of the subbands’ coefficients; computing the mean value of all the
log variances for each audio frame; and at the end, extracting the sub-hash bit. Authors report that the
performance of the algorithm degrades as the compression rate increases.
In [48], the authors proposed an algorithm with two stages named indexing and search.
The indexing is based in the construction of zone tables using the Search by Range Reduction (SRR)
threshold values [48]. The search is based on the SRR test, calculating the Itakura distance between
two fingerprints, and comparing it with values in the zone tables [48]. An accuracy of around 98%
is reported.
The authors of [43] proposed an algorithm with training and testing phases. For the training
phase, the authors started with the wavelet packet decomposition, and developed a local discriminant
bases (LDBs)-based automated multigroup audio classification system, which focuses on identifying
discriminatory time-frequency subspaces [43]. The testing phase consists of the construction of a
new wavelet tree, feature extraction, and implementation of a linear discriminant analysis (LDA) [43].
The extracted features include MFCC, spectral similarity, timbral texture, band periodicity, linear
prediction coefficient derived cepstral coefficients (LPCCs), zero crossing rate, MPEG-7 descriptors,
entropy, and octaves [43]. The authors of [43] reported that the average classification accuracy was
between 91% and 99% [43].
The authors of [44] presents a video retrieval system (VRS) for Interactive-Television as like
internet protocol television (IPTV), which implements an audio fingerprint feature of long-term
logarithmic modified discrete cosine transform (DCT) modulation coefficients (LMDCT-MC) for audio
indexing and retrieval, and implements two-stage search (TSS) algorithm for fast searching. In the
first stage of TSS, candidate video segments are roughly determined with audio index bit vectors (IBV)
66
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
13 of 23
and then the optimal video clip is obtained by fingerprint bit vectors (FBV). An accuracy of 99.67% is
reported in [44].
In [41] an audio fingerprint method using sub-fingerprint masking based on the predominant
pitch extraction is proposed. It increases the accuracy of the audio fingerprinting system in a noisy
environment dramatically, while requiring much less computing power compared to the expanded
hash table lookup method. When applied to an audio signal without noise, the reported accuracy
is 97.4%.
The authors of [42] presented a sub-Nyquist audio fingerprinting system for music recognition,
which utilizes Compressive Sampling (CS) theory to generate a compact audio fingerprint, and to
achieve significant reduction of the dimensionality of the input signal, compared to Nyquist sampling
methods [42]. The average accuracy of this method is 93.43% under various distorted environments.
In [38], the authors presented an algorithm based on fingerprinting techniques implemented in a
low-cost embedded reconfigurable platform. It utilizes the FFT implementation from the CUFFT library,
based on the Fastest Fourier Transform in the West (FFTW) algorithm. This approach normalizes and
frames the audio signal, computes the correlation and cross correlation, and applies a derivative of the
audio signal. An accuracy of 94% is reported.
The authors of [39] created a fingerprint database of songs and focused on the problem of effective
and efficient database search. The authors proposed a new indexing scheme that overcomes the
limitations of Haitsma-Kalker’s method and Miller’s k-ary tree method, adopting the inverted file
as the underlying index structure and developing the techniques to apply it to the effective and
efficient audio fingerprinting problem. An accuracy higher than 97% is reported in [39], which is the
performance of the Haitsma-Kalker’s method.
The authors of [40] explored a common audio-fingerprinting approach with the implementation
of FFT, and taken into account the noise in the derived fingerprints by employing error correcting
codes and applying statistical tests. Testing with several sample windows of Network Time Protocol
(NTP)-based synchronization recordings, authors of [40] reported an accuracy between 60% and 70%.
The authors of [31] created a system based on a client-server architecture able to recognize a live
television show using audio fingerprinting. To create audio fingerprints, FFT is computed to obtain the
power spectrum, which is integrated over a pre-defined set of non-overlapping, logarithmically spaced
frequency bins and eventually squared to obtain an energy measure [31]. The likelihood estimation
based on the cross-correlation function was used for comparison of the audio fingerprints. An accuracy
of around 95% is reported in [31].
The authors of [32] presented an audio fingerprint method named Masked Audio Spectral
Keypoints (MASK), which encodes the acoustic information existent in audio documents and
discriminates between transformed versions of the same acoustic documents and other unrelated
documents. The MASK fingerprint extraction method is composed of several tasks: time-to-frequency
transformation, where the input signal is transformed from the time domain to the spectral domain,
and transformed into Mel-scale; salient spectral points search; local mask application around each of
the salient points; grouping of the different spectrogram values into regions; and the MASK fingerprint
encoding and storage. The averaged energy values of each one of these spectrogram regions are
compared to construct a fixed length binary descriptor. Authors of [32] report an accuracy around 58%.
In [33], the authors implemented an audio fingerprinting algorithm based on fingerprint extraction
and matching search, adapting the well-known Philips’ algorithm. The fingerprint extraction derives
and encodes a set of relevant audio features, which need to be invariant to various kinds of signal
distortion, including background noise, audio compression, and A/D conversion [33]. The matching
search finds the best match between these fingerprints and those stored in the database [33].
The implemented audio fingerprint extraction method uses FFT, and extracts several features, such as:
mel-frequency cepstral coefficients (MFCC), spectral centroid or spectral flatness [33]. The audio
fingerprinting method reports an accuracy of 95% and a precision of 100% [33].
67
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
14 of 23
The authors of [34] proposed an audio fingerprinting system with several characteristics, including
robustness, granularity, and retrieval speed, reporting an accuracy between 88% and 99%. The structure
of the audio fingerprinting implemented is the same as all other algorithms presented in [34], applying
the FFT and an High-pass filter. The authors used the local maximum chroma energy (LMCE) to
extract the perception features of Tempo-Frequency domain [34].
The work presented in [29] reviews the state-of-the-art methods for improving the power
consumption and computation speed to make the smartphone implementation. It also proposed
the Improved Real-Time TV-channel Recognition (IRTR), which is a fingerprint extraction method
aimed at recognizing in real time what people are watching on TV without any active user interaction.
The identification using the audio fingerprint is performed using a likelihood estimation [29].
The audio fingerprinting method implements linear transform and feature extraction, with several
steps: the audio is recorded and divided into frames with overlap factor; each frame is filtered by means
of a Hamming window function; the application of the FFT and the squared modulus; the spectrum is
divided into logarithmically spaced frequency bins and the energy is computed for each bin; and the
nervy of band of each frame is denoted. An accuracy about 95% is reported in [29].
In [30], an audio fingerprinting algorithm is proposed for efficient retrieval of corresponding
or similar items from large audio databases, which improves the of the database search compared
to the algorithm used in Haitsma’s method, without impairing the accuracy of the search results.
The approach implements the FFT, the extraction of candidate songs via lookup table, the assignment
of weights to candidate songs, and the database search [30], while reporting an average accuracy
around 81%.
The authors of [21] proposed an audio fingerprinting algorithm, which improves binary audio
fingerprint matching performance by utilizing auxiliary information. The proposed matching method
is based on Philips robust hash (PRH) for audio signal; Asymmetric Fingerprint Matching for PRH
using the Magnitude Information, which consists of Normalization of the Subband-Energy Difference;
and Fingerprint Matching Based on the Likelihood Ratio Test [21]. The proposed method yields better
performance than the conventional Hamming distance [21].
The authors of [22] proposed an audio fingerprinting constituted by two stages: fingerprinting
and matching. The fingerprinting module uses a log-frequency spectrogram based on the Constant
Q Transform (CQT), and an adaptive thresholding method based on two-dimensional median
filtering [22]. The matching uses the Hamming similarity and the Hough Transform [22]. The reported
accuracy is between 61% and 81%.
The authors of [23] presented a method for the construction of audio fingerprints based on:
maximization of the mutual information across the distortion channel; using the information bottleneck
method to optimize the filters; and quantizers that generate these fingerprints. The method starts
with the application of the short time Fourier transform (STFT), and capturing the Spectral Sub-band
Centroids (SSC) using 16 bins on the Bark scale. The generated features with [23] result in a maximum
accuracy of around 65%.
The authors of [24] implemented an audio fingerprinting algorithm composed by several steps:
downsampling to 5 kHz, segmenting frames every 11.6 ms, applying the FFT, calculating the frequency
bands energies, and finally, calculating the fingerprints. A recall around 98% is reported.
In [25], the authors presented an audio fingerprinting algorithm based on the compressed-domain
spectral entropy as audio features, showing strong robustness against various audio signal distortions
such as recompression, noise interference, echo addition, equalization, band-pass filtering, pitch
shifting, moderate time-scale modification, among others. The algorithm includes four steps: granule
grouping, frequency alignment between long and short windows, coefficients selection and subband
division, and MDCT spectral entropy calculation and fingerprint modelling [25]. It reports an accuracy
above 90%.
In [26], the authors presented the implementation of an audio fingerprinting system, using graphic
processing units (GPUs). The system starts with the extraction of landmarks using FFT, and continues
68
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
15 of 23
with the landmark extraction, lookup, and analysis. The authors explored the use of one thread for
one hash key, and one block for one hash key, reporting an accuracy around 80.96%, when there are
100,000 songs in the database [26].
The authors of [27] proposed a high-performance audio fingerprint extraction method for
identifying TV commercial advertisement. The audio fingerprint extraction consists of a salient
audio peak pair fingerprints based on constant Q transform (CQT). The algorithm obtains the audio
fingerprints through five main steps: preprocessing; application of the CQT; application of the Mean
Subtraction of Logarithmic CQT Spectrum; application of the CQT Based Salient Peak Detection using
forward and Backward Filtering; and finally, application of the Fingerprint Generation using CQT Peak
Pair. The reported recognition accuracy of the method based on CQT, presented in [27], is around 89.8%.
The authors of [16] used a smartphone and create an audio fingerprinting algorithm based on the
joint usage of GPS and acoustic fingerprints. The authors created an audio fingerprinting algorithm
with noise tolerance, assessing it under several conditions [16]. The algorithm starts with the calculation
of the audio sample spectrogram using the STFT, and then calculates audio sample spectrogram using
the Hamming window and a high overlap [16]. Next, it takes only the first 40 frequency bins, as most
of the useful audio features are in that bandwidth, averaging the logarithmic amplitude in each bin [16].
Afterwards, for each frequency bin, a 16-bit fingerprint is calculated [16]. The 16-bits fingerprint is then
stored with the associated frequency and time [16]. For the comparison of the audio fingerprints, the
Hamming distances between each fingerprint are calculated, looking for a minimum [16]. An accuracy
of around 86% is reported.
In [17], the authors proposed an approach to accelerate fingerprinting techniques by skipping the
search for irrelevant sections of the signal and demonstrate its application to the divide and locate (DAL)
audio fingerprint method. The method in DAL starts with the extraction of the time-frequency power
spectral applied for the signals, normalizing each logarithmic power [17]. Afterwards, the normalized
data is decomposed into a number of small time-frequency components of uniform size, and thus,
the computational cost and memory usage are reduced in the fingerprint data [17]. The authors verified
that with a reduced search threshold, the accuracy of the recognition is around 100% [17].
The authors of [18] created a method to analyze and classify daily activities in personal audio
recordings (PARs). The method applies: speech activity detection (SAD), speaker diarization systems,
and computing the number of audio speech and lexical features [18]. It uses a TO-Combo-SAD
(Threshold Optimized Combo SAD) algorithm for separating speech from noise [18]. The Principal
Component Analysis (PCA) is first applied for dimensionality reduction, and then, the remaining
features are supplied to a multi-class support vector machine (SVM) with radial basis function (RBF)
kernel for model training and evaluation [18]. The authors performed recognition of faculty meeting,
research meeting, staff meeting, alone time, and conference call, reporting accuracies between 62.78%
and 84.25% [18].
In [19], the authors proposed an audio fingerprinting method, based on landmarks in the audio
spectrogram. The algorithm is based on the audio hashing of frequency peaks in the spectrogram [19].
It starts with the application of the FFT, thresholding the data, applying a high pass filter, identifying
the local maximums and finding the peaks of the spectrogram [19]. The performance of the algorithm
decreases at higher additive noise in comparison with other algorithms [19], reporting an accuracy
around 96.71%.
In [20], the authors proposed a robust TV advertisement search based on audio fingerprinting
in real environments. This algorithm has several steps, such as preprocessing, logarithmic moduled
complex lapped transform (LMCLT), two-are segmentation using adaptive thresholding based on
median filtering, detection of prominent LMCLT spectral peaks, and fingerprinting generation using
LMCLT peak pair [20]. The method applies adaptive peak-picking thresholding method to extract
more salient and distinct peak pairs for comparing the query fingerprint with the original fingerprints,
and the authors reported an accuracy of 86.5% [20].
69
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
16 of 23
4. Discussion
This review confirms the findings of previous studies related to the use of audio fingerprinting
techniques for identification of the environment related to the different ADLs. We consider that
many of the reviewed works raise important issues regarding the concept of Open Science, including,
but not limited to, Reproducibility and Verifiability of the research results, and Comparability of
similar research. Many of them were evaluated on unpublished data and did not publish their source
code, although when commercial solutions are in perspective, a necessary degree of confidentiality is
understandable. Regarding validation and comparability, only six studies used raw data available
online or published its research data online. Likewise, only three studies presented some parts of the
code used of the experiments. In addition, the studies that used data that is now publicly available,
did not publish the research source code, making the validation of the results and further comparative
research an impossible task. Therefore, we suggest to the audio fingerprinting community to become
better at sharing raw data and algorithms, so as to be able to recreate and evaluate the soundness of
previous studies.
Nevertheless, assuming the results of the presented research studies are comparable, Tables 3–5
present a summary of the Features and Methods ordered by the number of identified studies that use
these features and methods.
Tables 3 and 4 present the distribution of the extracted features and methods implemented in the
analyzed studies, verifying that FFT is one of the most widely used feature extraction method, because
it extracts the frequencies from the audio signal, and the other most used features include thresholding,
normalized Spectral Subband Centroid (SSC), Mel-frequency cepstrum coefficients (MFCC), maximum,
local peaks and landmarks, Shannon entropy, Rényi entropy, MPEG-7 descriptors, Spectral bandwidth,
Spectral flatness measure, Modified discrete cosine transform (MDCT), Constant Q Transform (CQT),
Short-time Fourier transform (STFT), average, and the maximum and minimum. These features
were used in a large part of the analyzed studies [12,14,19,24,26,28–31,33–35,38,49–51], and with them,
the reported accuracy is greater than 80%, as presented in Table 3.
For Tables 3 and 4, the accuracies that are equal or higher than 99% are shown in a different
background color (yellow).
Table 3. Distribution of the features extracted in the studies.
70
Features
Average Accuracy
of Features
Number of
Studies
Fast Fourier Transform (FFT)
Thresholding
Normalized spectral subband centroid (SSC)
Mel-frequency cepstrum coefficients (MFCC)
Maximum
Local peaks and landmarks
Shannon entropy
Rényi entropy
MPEG-7 descriptors
Spectral bandwidth
Spectral flatness measure
Modified discrete cosine transform (MDCT)
Constant Q transform (CQT)
Short-time Fourier transform (STFT)
Average
Minimum
93.85%
90.49%
93.44%
97.30%
87.57%
82.32%
99.10%
99.10%
97.50%
97.10%
97.10%
93.00%
85.40%
84.50%
83.00%
83.00%
16
6
5
4
3
3
2
2
2
2
2
2
2
2
2
2
Sum of the spectrum modulus of every frame
Sum of global spectrum modulus in two stages
Local energy centroid (LEC)
Time-frequency power spectral
100.00%
100.00%
100.00%
100.00%
1
1
1
1
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
17 of 23
Table 3. Cont.
Features
Average Accuracy
of Features
Number of
Studies
Long-term logarithmic modified discrete cosine transform (DCT)
modulation coefficients (LMDCT-MC)
99.67%
1
Bit packing
Spectral band energy
Spectral crest factor
Spectral similarity
Timbral texture
Band periodicity
Linear prediction coefficient derived cepstral coefficients (lpccs)
Zero crossing rate
Octaves
Single local description
Multiple local description
Chroma vectors
MP3 signatures
Time averaging
Quadratic interpolation
Sinusoidal quantification
Frequency-axial discretization
Time-axial warping
Logarithmic moduled complex lapped transform spectral peaks
Correlation coefficient
Matching score
99.20%
99.20%
99.20%
99.00%
99.00%
99.00%
99.00%
99.00%
99.00%
96.00%
96.00%
96.00%
96.00%
96.00%
96.00%
96.00%
96.00%
96.00%
86.50%
70.00%
70.00%
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
Table 4. Distribution of the methods implemented in the studies.
Methods
Average Accuracy
of Methods
Number of
Studies
Other methods
Two level search algorithm
Likelihood estimation
Principal Component Analysis (PCA)
Hamming distances between each fingerprint
Streaming audio fingerprinting (SAF)
Human auditory system (HAS)
Divide and locate (DAL)
Gaussian mixture models (GMM) modelling
90.78%
99.84%
97.10%
90.13%
83.50%
100.00%
100.00%
100.00%
99.20%
15
2
3
2
2
1
1
1
1
Local discriminant bases (LDBS)-based automated multigroup audio
classification system
99.00%
1
Linear discriminant analysis (LDA)
Local maximum chroma energy (LMCE)
Expanded hash table lookup method
Query Context (QUC)-tree
Improved Real-Time TV-channel Recognition (IRTR)
Sub-Nyquist fudio fingerprinting system
Logarithmic moduled complex lapped transform (LMCLT) peak pair
TO-Combo-SAD (Threshold Optimized Combo SAD) algorithm
Support vector machine (SVM)
Hough Transform between each fingerprint
Masked audio spectral keypoints (MASK)
99.00%
99.00%
97.40%
96.00%
95.00%
93.43%
86.50%
84.25%
84.25%
81.00%
58.00%
1
1
1
1
1
1
1
1
1
1
1
71
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
18 of 23
On the other hand, as verified in Table 4, a large part of the analyzed
studies [12,14,19,24,26,28–31,33–35,38,49–51] do not mention the name of the method applied,
presenting only the features used. Regarding the undifferentiated methods, the most used methods are
the two level search algorithm, the likelihood estimation, the Principal Component Analysis (PCA),
and the Hamming distances between each fingerprint, reporting accuracies also higher than 80%.
Table 5 presents in a matrix format, the average of the averages of the accuracies in Methods vs.
its Features. This table is a mere comparison exercise, as there are not enough studies to sustain a valid
analysis of the use of different features with different methods. On the other hand, this table assumes
that these results are comparable, and moreover, that any method or algorithm can be used with any
set of features, which of course, is a very wide, and possibly not true assumption. Nevertheless, Table 5
shows, in a colored background the match between features and methods. For example, for method
SAF (Streaming Audio Fingerprinting) the set of used features are Fast Fourier Transform, Thresholds
and Energy bands, whose mean accuracies in the found studies are not higher than 99%. Also,
for example for the method GMM (Gaussian Mixture Models Modelling), besides the 4 highlighted
features that were used, this method uses additionally 5 other sets of features.
Taking Table 5 into consideration, one can identify Shannon’s Entropy as the feature that is most
used in the most accurate number of studies. Arguably, this table may propose new combinations of
features and methods that can be used to devise audio-fingerprinting solutions.
For a particular use, the methods to be implemented must be chosen according to their complexity,
the computational power of the use case scenario, and to the purpose of its intended use. This review is
focused on the use of mobile devices, but only three of the reviewed works argue that they use methods
that need low resources (see Table 1). Only 19 studies compared the implemented methods with other
methods published in the literature and present their accuracy, claiming an increased accuracy in the
recognition of the environment using audio fingerprinting.
According to the results of this review, the use of the mobile devices for the application of audio
fingerprinting techniques is limited, because of the restrictions these devices impose, i.e., low power
processing and battery capacity. Thus, only 10 of the analyzed studies utilize mobile devices with local
processing or server-side processing of the data acquired from the mobile devices. In the case of the
server-side processing, the use of the mobile devices implies a constant and stable network connection,
which is not a trivial requirement both from technical perspective, but also because of battery life
implications. To some extent, the using Fog and Mist Computing paradigms could overcome the
challenges of the client-server architectures. The creation of lightweight techniques should be explored,
as they could be executed on mobile devices (i.e., edge-nodes). The models could be recalibrated
offline on the server occasionally, and then, as pre-trained models to be seamlessly redeployed on
mobile devices.
In conclusion, only one of the reviewed studies [38] can achieve reliable performance with reduced
computational cost and memory usage. It utilizes the FFT implementation in the CUFFT library, divide
and locate (DAL) audio fingerprint method, and sub-fingerprint masking based on the predominant
pitch extraction methods. However, other methods could be implemented on mobile devices with
some restrictions. Nonetheless, they could be amended to utilize more lightweight implementations of
the underlying libraries, or by sacrificing floating point precision, for instance.
72
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
19 of 23
Table 5. Potential accuracies for the top most accurate methods vs. top most mean accurate features (mean accuracies equal or higher than 99%, according to its
authors).
SAF
HAS
DAL
TLS
GMM
LDBS
LDA
LMCE
Local energy centroid (LEC)
Sum of global spectrum modulus in two stages
Sum of the spectrum modulus of every frame
Time-frequency power spectral
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
100.00%
99.84%
99.92%
99.92%
99.92%
99.92%
99.20%
99.60%
99.60%
99.60%
99.60%
99.00%
99.50%
99.50%
99.50%
99.50%
99.00%
99.50%
99.50%
99.50%
99.50%
99.00%
99.50%
99.50%
99.50%
99.50%
Long-term logarithmic modified discrete cosine transform (DCT)
modulation coefficients (LMDCT-MC)
99.67%
99.84%
99.84%
99.84%
99.76%
99.44%
99.34%
99.34%
99.34%
Bit packing
Spectral band energy
Spectral crest factor
Rényi entropy
Shannon entropy
Band periodicity
Linear prediction coefficient derived cepstral coefficients (lpccs)
Octaves
Spectral similarity
Timbral texture
Zero crossing rate
99.20%
99.20%
99.20%
99.10%
99.10%
99.00%
99.00%
99.00%
99.00%
99.00%
99.00%
99.60%
99.60%
99.60%
99.55%
99.55%
99.50%
99.50%
99.50%
99.50%
99.50%
99.50%
99.60%
99.60%
99.60%
99.55%
99.55%
99.50%
99.50%
99.50%
99.50%
99.50%
99.50%
99.60%
99.60%
99.60%
99.55%
99.55%
99.50%
99.50%
99.50%
99.50%
99.50%
99.50%
99.52%
99.52%
99.52%
99.47%
99.47%
99.42%
99.42%
99.42%
99.42%
99.42%
99.42%
99.20%
99.20%
99.20%
99.15%
99.15%
99.10%
99.10%
99.10%
99.10%
99.10%
99.10%
99.10%
99.10%
99.10%
99.05%
99.05%
99.00%
99.00%
99.00%
99.00%
99.00%
99.00%
99.10%
99.10%
99.10%
99.05%
99.05%
99.00%
99.00%
99.00%
99.00%
99.00%
99.00%
99.10%
99.10%
99.10%
99.05%
99.05%
99.00%
99.00%
99.00%
99.00%
99.00%
99.00%
73
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
20 of 23
5. Conclusions
This review identified and described the methodologies used for audio fingerprinting that can
be applied to mobile technologies. Forty-seven studies were examined and the main findings are
summarized as follows:
•
•
•
(RQ1) the audio fingerprinting is defined as the ability to recognize the scenario in which a given
audio was collected and involved in, based on various methods.
(RQ2) Several techniques have been applied to implement audio fingerprinting methods,
including Fast Fourier Transform (FFT). Support Vector Machine (SVM), QUery Context
(QUC)-tree, spectral subband centroid (SSC), Streaming Audio Fingerprinting (SAF), Human
Auditory System (HAS), Gaussian mixture models (GMM) modelling, likelihood estimation,
linear discriminant analysis (LDA), Compressive Sampling (CS) theory, Philips robust hash (PRH),
Asymmetric Fingerprint Matching, and TO-Combo-SAD (Threshold Optimized Combo SAD).
These techniques yield high accuracy, and the use of mobile devices does not influence the
predictive performance, allowing the use of these techniques anywhere, anytime.
(RQ3) All of the methods presented in RQ2 can be implemented on mobile devices, but the
methods that require lower computational resources are FFT with the CUFFT library, divide and
locate (DAL) audio fingerprint method, and sub-fingerprint masking based on the predominant
pitch extraction.
In addition, this review highlights the application of audio fingerprinting techniques on mobile
or other devices with limited computational and battery resources. Some limitations of this review
should be mentioned. First, the authors chose to exclude studies that are not focused on audio
fingerprinting techniques. Second, the studies that do not utilize mobile devices have been excluded.
These exclusions were performed with the analysis of the abstract and then, the full text of the papers.
Finally, only English-language publications were included.
Based on the analysis, we conclude that the most used methods are undifferentiated
methods, two level search algorithms, likelihood estimation, Principal Component Analysis (PCA),
and Hamming distances between each fingerprint. The conclusion is that the use of statistical methods
reports results with an accuracy higher than 80%. Furthermore, the most used features are Fast
Fourier Transform (FFT), Thresholding, normalized spectral subband centroid (SSC), Mel-frequency
cepstrum coefficients (MFCC), maximum, local peaks and landmarks, Shannon entropy, Rényi entropy,
MPEG-7 descriptors, Spectral bandwidth, Spectral flatness measure. Modified discrete cosine transform
(MDCT), Constant Q Transform (CQT), Short-time Fourier transform (STFT), average, and minimum,
which also result in accuracies greater than 80%.
As future work, the extraction of features based on audio fingerprinting will be implemented in
order to develop a system for the recognition of ADLs and their environments, presented in [9–11].
As presented in Table 3, the accuracy is always higher than 80%. Then, we should consider the most
used features, including FFT, MFCC, average, maximum, and minimum, in order to better handle the
recognition of the environment. The implementation of this framework is part of the development of a
personal digital life coach [4].
Acknowledgments: This work was supported by FCT project UID/EEA/50008/2013 (Este trabalho foi suportado pelo
projecto FCT UID/EEA/50008/2013). The authors would also like to acknowledge the contribution of the COST
Action IC1303—AAPELE—Architectures, Algorithms and Protocols for Enhanced Living Environments.
Author Contributions: All the authors have contributed with the structure, content, and writing of the paper.
Conflicts of Interest: The authors declare no conflict of interest.
74
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
21 of 23
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
Foti, D.; Koketsu, J.S. Pedretti’s Occupational Therapy: Practical Skills for Physical Dysfunction, 7th ed.; Activities
of daily living; Mosby: St. Louis, MI, USA, 2013; pp. 157–232.
Garcia, N.M.; Rodrigues, J.J.P. Ambient Assisted Living; CRC Press: Boca Raton, FL, USA, 2015.
Dobre, C.; Mavromoustakis, C.X.; Goleva, R.L. Ambient Assisted Living and Enhanced Living Environments:
Principles, Technologies and Control; Butterworth-Heinemann: Oxford, UK, 2016; p. 552.
Garcia, N.M. A Roadmap to the Design of a Personal Digital Life Coach. In ICT Innovations 2015; Springer:
Cham, Switzerland, 2016.
Da Silva, J.R.C. Smartphone Based Human Activity Prediction; Faculdade de Engenharia: Porto, Portugal, 2013.
Bieber, G.; Luthardt, A.; Peter, C.; Urban, B. The Hearing Trousers Pocket—Activity Recognition by
Alternative Sensors. In Proceedings of the 4th International Conference on PErvasive Technologies Related
to Assistive Environments, Crete, Greece, 25–27 May 2011.
Kazushige, O.; Miwako, D. Indoor-outdoor activity recognition by a smartphone. In Proceedings of the 2012
ACM Conference on Ubiquitous Computing, Pittsburgh, PA, USA, 5–8 September 2012; p. 537.
Ganti, R.K.; Srinivasan, S.; Gacic, A. Multisensor Fusion in Smartphones for Lifestyle Monitoring.
In Proceedings of the 2010 International Conference on Body Sensor Networks, Singapore, 7–9 June 2010.
Pires, I.M.; Garcia, N.M.; Pombo, N.; Flórez-Revuelta, F. From Data Acquisition to Data Fusion:
A Comprehensive Review and a Roadmap for the Identification of Activities of Daily Living Using Mobile
Devices. Sensors 2016, 16, 184. [CrossRef] [PubMed]
Pires, I.M.; Garcia, N.M.; Flórez-Revuelta, F. Multi-sensor data fusion techniques for the identification
of activities of daily living using mobile devices. In Proceedings of the ECMLPKDD 2015 Doctoral
Consortium, European Conference on Machine Learning and Principles and Practice of Knowledge Discovery
in Databases, Porto, Portugal, 7–11 September 2015.
Pires, I.M.; Garcia, N.M.; Pombo, N.; Flórez-Revuelta, F. Identification of Activities of Daily Living Using
Sensors Available in off-the-shelf Mobile Devices: Research and Hypothesis. In Ambient Intelligence-Software
and Applications, Proceedings of the 7th International Symposium on Ambient Intelligence (ISAmI 2016), Seville,
Spain, 1–3 June 2016; Springer International Publishing: Cham, Switzerland, 2016.
Sui, D.; Ruan, L.; Xiao, L. A Two-level Audio Fingerprint Retrieval Algorithm for Advertisement Audio.
In Proceedings of the 12th International Conference on Advances in Mobile Computing and Multimedia,
Kaohsiung, Taiwan, 8–10 December 2014; pp. 235–239.
Liu, C.-C.; Chang, P.-F. An efficient audio fingerprint design for MP3 music. In Proceedings of the 9th
International Conference on Advances in Mobile Computing and Multimedia, Ho Chi Minh City, Vietnam,
5–7 December 2011; pp. 190–193.
Liu, C.-C. MP3 sniffer: A system for online detecting MP3 music transmissions. In Proceedings of the 10th
International Conference on Advances in Mobile Computing, Bali, Indonesia, 3–5 December 2012; pp. 93–96.
Tsai, T.J.; Stolcke, A. Robust and efficient multiple alignment of unsynchronized meeting recordings.
IEEE/ACM Trans. Audio Speech Lang. Proc. 2016, 24, 833–845. [CrossRef]
Casagranda, P.; Sapino, M.L.; Candan, K.S. Audio assisted group detection using smartphones.
In Proceedings of the 2015 IEEE International Conference on Multimedia & Expo Workshops (ICMEW),
Torino, Italy, 29 June–3 July 2015.
Nagano, H.; Mukai, R.; Kurozumi, T.; Kashino, K. A fast audio search method based on skipping irrelevant
signals by similarity upper-bound calculation. In Proceedings of the 2015 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), Brisbane, Australia, 19–24 April 2015.
Ziaei, A.; Sangwan, A.; Kaushik, L.; Hansen, J.H.L. Prof-Life-Log: Analysis and classification of activities in
daily audio streams. In Proceedings of the 2015 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), Brisbane, Australia, 19–24 April 2015.
George, J.; Jhunjhunwala, A. Scalable and robust audio fingerprinting method tolerable to time-stretching.
In Proceedings of the 2015 IEEE International Conference on Digital Signal Processing (DSP), Singapore,
21–24 July 2015.
Kim, H.G.; Cho, H.S.; Kim, J.Y. TV Advertisement Search Based on Audio Peak-Pair Hashing in Real
Environments. In Proceedings of the 2015 5th International Conference on IT Convergence and Security
(ICITCS), Kuala Lumpur, Malaysia, 24–27 August 2015.
75
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
76
22 of 23
Seo, J.S. An Asymmetric Matching Method for a Robust Binary Audio Fingerprinting. IEEE Signal
Process. Lett. 2014, 21, 844–847.
Rafii, Z.; Coover, B.; Han, J. An audio fingerprinting system for live version identification using image
processing techniques. In Proceedings of the 2014 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP), Florence, Italy, 4–9 May 2014.
Naini, R.; Moulin, P. Fingerprint information maximization for content identification. In Proceedings of the
2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Florence, Italy,
4–9 May 2014.
Yang, G.; Chen, X.; Yang, D. Efficient music identification by utilizing space-saving audio fingerprinting
system. In Proceedings of the 2014 IEEE International Conference on Multimedia and Expo (ICME), Chengdu,
China, 14–18 July 2014.
Yin, C.; Li, W.; Luo, Y.; Tseng, L.-C. Robust online music identification using spectral entropy in the
compressed domain. In Proceedings of the Wireless Communications and Networking Conference
Workshops (WCNCW), Istanbul, Turkey, 6–9 April 2014.
Wang, C.C.; Jang, J.S.R.; Liou, W. Speeding up audio fingerprinting over GPUs. In Proceedings of the 2014
International Conference on Audio, Language and Image Processing (ICALIP), Shanghai, China, 7–9 July 2014.
Lee, J.Y.; Kim, H.G. Audio fingerprinting to identify TV commercial advertisement in real-noisy environment.
In Proceedings of the 2014 14th International Symposium on Communications and Information Technologies
(ISCIT), Incheon, South Korea, 24–26 September 2014.
Shibuya, T.; Abe, M.; Nishiguchi, M. Audio fingerprinting robust against reverberation and noise based on
quantification of sinusoidality. In Proceedings of the 2013 IEEE International Conference on Multimedia and
Expo (ICME), San Jose, CA, USA, 15–19 July 2013.
Bisio, I.; Delfino, A.; Lavagetto, F.; Marchese, M. A Television Channel Real-Time Detector using Smartphones.
IEEE Trans. Mob. Comput. 2015, 14, 14–27. [CrossRef]
Lee, S.; Yook, D.; Chang, S. An efficient audio fingerprint search algorithm for music retrieval. IEEE Trans.
Consum. Electron. 2013, 59, 652–656. [CrossRef]
Bisio, I.; Delfino, A.; Luzzati, G.; Lavagetto, F.; Marchese, M.; Fra, C.; Valla, M. Opportunistic estimation of
television audience through smartphones. In Proceedings of the 2012 International Symposium on Performance
Evaluation of Computer and Telecommunication Systems (SPECTS), Genoa, Italy, 8–11 July 2012.
Anguera, X.; Garzon, A.; Adamek, T. MASK: Robust Local Features for Audio Fingerprinting. In Proceedings
of the 2012 IEEE International Conference on Multimedia and Expo, Melbourne, Australia, 9–13 July 2012.
Duong, N.Q.K.; Howson, C.; Legallais, Y. Fast second screen TV synchronization combining audio fingerprint
technique and generalized cross correlation. In Proceedings of the 2012 IEEE International Conference on
Consumer Electronics (ICCE-Berlin), Berlin, Germany, 3–5 September 2012.
Wang, H.; Yu, X.; Wan, W.; Swaminathan, R. Robust audio fingerprint extraction algorithm based on 2-D
chroma. In Proceedings of the 2012 International Conference on Audio, Language and Image Processing
(ICALIP), Shanghai, China, 16–18 July 2012.
Xiong, W.; Yu, X.; Wan, W.; Swaminathan, R. Audio fingerprinting based on dynamic subband locating
and normalized SSC. In Proceedings of the 2012 International Conference on Audio, Language and Image
Processing (ICALIP), Shanghai, China, 16–18 July 2012.
Jijun, D.; Wan, W.; Yu, X.; Pan, X.; Yang, W. Audio fingerprinting based on harmonic enhancement and
spectral subband centroid. In Proceedings of the IET International Communication Conference on Wireless
Mobile and Computing (CCWMC 2011), Shanghai, China, 14–16 November 2011.
Pan, X.; Yu, X.; Deng, J.; Yang, W.; Wang, H. Audio fingerprinting based on local energy centroid.
In Proceedings of the IET International Communication Conference on Wireless Mobile and Computing
(CCWMC 2011), Shanghai, China, 14–16 November 2011.
Martinez, J.I.; Vitola, J.; Sanabria, A.; Pedraza, C. Fast parallel audio fingerprinting implementation in
reconfigurable hardware and GPUs. In Proceedings of the 2011 VII Southern Conference on Programmable
Logic (SPL), Cordoba, Argentina, 13–15 April 2011.
Cha, G.H. An Effective and Efficient Indexing Scheme for Audio Fingerprinting. In Proceedings of the
2011 5th FTRA International Conference on Multimedia and Ubiquitous Engineering (MUE), Crete, Greece,
28–30 June 2011.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 160
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
23 of 23
Schurmann, D.; Sigg, S. Secure Communication Based on Ambient Audio. IEEE Trans. Mob. Comput. 2013,
12, 358–370. [CrossRef]
Son, W.; Cho, H.-T.; Yoon, K.; Lee, S.-P. Sub-fingerprint masking for a robust audio fingerprinting system in
a real-noise environment for portable consumer devices. IEEE Trans. Consum. Electron. 2010, 56, 156–160.
[CrossRef]
Chang, K.K.; Pissis, S.P.; Jang, J.-S.R.; Iliopoulos, C.S. Sub-nyquist audio fingerprinting for music recognition.
In Proceedings of the Computer Science and Electronic Engineering Conference (CEEC), Colchester, UK,
8–9 September 2010.
Umapathy, K.; Krishnan, S.; Rao, R.K. Audio Signal Feature Extraction and Classification Using Local
Discriminant Bases. IEEE Trans. Audio Speech Lang. Proc. 2007, 15, 1236–1246. [CrossRef]
Kim, H.G.; Kim, J.Y.; Park, T. Video bookmark based on soundtrack identification and two-stage search for
interactive-television. IEEE Trans. Consum. Electron. 2007, 53, 1712–1717.
Sert, M.; Baykal, B.; Yazici, A. A Robust and Time-Efficient Fingerprinting Model for Musical Audio.
In Proceedings of the 2006 IEEE International Symposium on Consumer Electronics, St Petersburg, Russia,
28 June–1 July 2006.
Ramalingam, A.; Krishnan, S. Gaussian Mixture Modeling of Short-Time Fourier Transform Features for
Audio Fingerprinting. IEEE Trans. Inf. Forensics Secur. 2006, 1, 457–463. [CrossRef]
Ghouti, L.; Bouridane, A. A fingerprinting system for musical content. In Proceedings of the 2006 14th
European Signal Processing Conference, Florence, Italy, 4–8 September 2006.
Cook, R.; Cremer, M. A Tunable, Efficient, Specialized Multidimensional Range Query Algorithm.
In Proceedings of the 2006 IEEE International Symposium on Signal Processing and Information Technology,
Vancouver, BC, Canada, 28–30 August 2006.
Seo, J.S.; Jin, M.; Lee, S.; Jang, D.; Lee, S.; Yoo, C.D. Audio fingerprinting based on normalized spectral
subband centroids. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal
Processing (ICASSP 05), Philadelphia, PA, USA, 23 March 2005.
Haitsma, J.; Kalker, T. Speed-change resistant audio fingerprinting using auto-correlation. In Proceedings of
the 2003 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 03), Hong Kong,
China, 6–10 April 2003.
Haitsma, J.; Kalker, T.; Oostveen, J. An efficient database search strategy for audio fingerprinting.
In Proceedings of the 2002 IEEE Workshop on Multimedia Signal Processing, St.Thomas, VI, USA,
9–11 December 2002.
© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).
77
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
78
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 3
Framework for the Identification of Activities of
Daily Living
This chapter is related to the definition of the framework for the recognition of ADL and
environments, and it is composed by one article.
1. Approach for the Development of a Framework for
the Identification of Activities of Daily Living Using
Sensors in Mobile Devices
The following article is the second part of the chapter 3.
Approach for the Development of a Framework for the Identification of Activities of Daily Living
Using Sensors in Mobile Devices
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta and Susanna
Spinsante
Sensors (MDPI), published, 2018.
According to 2016 Journal Citation Reports published by Thomson Reuters in 2017, this journal
has the following performance metrics:
ISI Impact Factor (2016): 2.677
ISI Article Influence Score (2016): 0.6
Journal Ranking (2016): 104/642 (Electrical and Electronic Engineering)
79
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
80
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Article
Approach for the Development of a Framework for
the Identification of Activities of Daily Living Using
Sensors in Mobile Devices
Ivan Miguel Pires 1,2,3,*, Nuno M. Garcia 1,3,4, Nuno Pombo 1,3,4, Francisco Flórez-Revuelta 5
and Susanna Spinsante 6
Instituto de Telecomunicações, Universidade da Beira Interior, 6201-001 Covilhã, Portugal;
ngarcia@di.ubi.pt (N.M.G.); ngpombo@ubi.pt (N.P.)
2 Altranportugal, 1990-096 Lisbon, Portugal
3 ALLab - Assisted Living Computing and Telecommunications Laboratory, Computing Science
Department, Universidade da Beira Interior, 6201-001 Covilhã, Portugal
4 ECATI, Universidade Lusófona de Humanidades e Tecnologias, 1749-024 Lisbon, Portugal
5 Department of Computer Technology, Universidad de Alicante, 03690 Sant Vicent del Raspeig, Alicante,
Spain; francisco.florez@ua.es
6 Department of Information Engineering, Marche Polytechnic University, 60121 Ancona, Italy;
s.spinsante@univpm.it
* Correspondence: impires@it.ubi.pt; Tel.: +351-966-379-785
1
Received: 7 January 2018; Accepted: 19 February 2018; Published: 21 February 2018
Abstract: Sensors available on mobile devices allow the automatic identification of Activities of
Daily Living (ADL). This paper describes an approach for the creation of a framework for the
identification of ADL, taking into account several concepts, including data acquisition, data
processing, data fusion, and pattern recognition. These concepts can be mapped onto different
modules of the framework. The proposed framework should perform the identification of ADL
without Internet connection, performing these tasks locally on the mobile device, taking in account
the hardware and software limitations of these devices. The main purpose of this paper is to present
a new approach for the creation of a framework for the recognition of ADL, analyzing the allowed
sensors available in the mobile devices, and the existing methods available in the literature.
Keywords: Activities of Daily Living (ADL); environment; sensors; mobile devices; framework; data
acquisition; data processing; data fusion; pattern recognition; machine learning
1. Introduction
Sensors embedded in off-the-shelf mobile devices, e.g., accelerometers, gyroscopes,
magnetometers, microphones, and Global Positioning System (GPS) receivers [1], may be used in the
development of algorithms for the recognition of Activities of Daily Living (ADL) [2] and the
environments in which they are carried out. These algorithms are part of the development of a
Personal Digital Life Coach (PDLC) [3]. According to [3], a PDLC “(…) will monitor our actions and
activities, be able to recognize its user state of mind, and propose measures that not only will allow
the user to achieve his/her stated goals, but also to act as an intermediate health and well-being agent
between the user and his/her immediate care givers (…)”. This work is related to the development of
ambient assisted living (AAL) systems, and, due to the increasing demands in our society, it is a field
with high importance [4]. Due to recent advances in technology, there is an increasing number of
research studies in this field for the monitoring of people with impairments and older people in a
Sensors 2018, 18, 640; doi:10.3390/s18020640
www.mdpi.com/journal/sensors
81
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
2 of 22
plethora of situations by using AAL technologies, including mobile devices and smart environments
[5].
Multi-sensor data fusion technologies may be implemented with mobile devices, because they
incorporate several sensors, such as motion sensors, magnetic/mechanical sensors, acoustic sensors,
and location sensors [6], improving the accuracy of the recognition of several types of activities, e.g.,
walking, running, going downstairs, going upstairs, watching TV, and standing, and environments,
e.g., bar, classroom, gym, library, kitchen, street, hall, living room, and bedroom. The selection of the
activities and environments that will be included in the framework was based in the activities
previously recognized with best accuracies, and, in the case of the environments, there are a lack of
studies related to the environment recognition, taking into account some of the environments
previously recognized and the most common environments [7]. The recognition of ADL may be
performed with motion, magnetic/mechanical and location sensors, and the environments may be
recognized with acoustic sensors. In order to improve the recognition of the ADL, the environment
recognized may be fused with the other features extracted from the other sensors.
In accordance with previous works [6,8,9], the main motivation of this paper is to present the
architecture of a framework for the recognition of ADL and their environments, which takes
advantage of the use of a wide set of sensors available in a mobile device, also aiming at reducing the
current complexity and constraints in the development of these systems. The test and validation of this
framework is currently the subject of another step of this research plan [9], which includes the acquisition
of a dataset that contains approximately 2.7 hours of data collected from the accelerometer,
gyroscope, magnetometer, microphone and GPS receiver, related to each activity and environment.
During the collection phase, the data were acquired with the mobile device located in the front pocket
of the trousers by 25 subjects aged between 16 and 60 years old and different lifestyles (10 mainly
active and 15 mainly sedentary) and gender (10 female and 15 male). The activities performed and
the environments frequented were labelled by the user. The subjects used their personal mobile
phones with their applications running, where the mainly used device was a BQ Aquarius device
[10].
The identification of ADL and environments using sensors has been studied during the last
years, and several methods and frameworks [11–16] have been implemented using smartphones.
However, this is a complex problem that should be separated into different stages, such as data
acquisition, processing, and fusion; and artificial intelligence systems. The frameworks developed in
previous studies are commonly only focused on some specific parts of the problem. For example, the
Acquisition Cost-Aware QUery Adaptation (ACQUA) framework [17] has been designed for data
acquisition and data processing, but it does not include all the steps needed for data processing.
There are no predefined standards for the creation of a framework for the recognition of the
ADL [18–20], and the most implemented methods for the recognition of ADL are related to the use
of motion sensors. However, there are methods and sensors that can be fused for the creation of a
structured framework as a holistic approach to the identification of the ADL and environments
presented in this paper.
Around the concept of sensors’ data fusion, the selection of the sensors to use is the first step for
the creation of the framework, defining a method for the acquisition of the data, and, consequently,
their processing. The processing of the data includes data cleaning, data imputation, and extraction
of the features. Data segmentation techniques are not considered, as this study was designed for local
execution on mobile devices and, due to the low memory and power processing restrictions of these
devices, only a short sample of the sensors’ data can be used (initial research points to 5 s samples).
This strategy makes it unsuitable to apply data segmentation techniques while still making it possible
to deploy the framework in scarce resource devices. The final step in the proposed framework is the
selection of the best features, and then the application of artificial intelligence techniques, i.e., the
implementation of three types of Artificial Neural Networks (ANN), such as Multilayer Perceptron
(MLP) with Backpropagation, Feedforward Neural Networks (FNN) with Backpropagation and
Deep Neural Networks (DNN), in order to choose the best method for the accurate recognition of the
ADL and the environments.
82
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
3 of 22
The remaining sections of this paper are organized as follows: Section 2 presents the state of the
art in this topic, presenting a set of methods for each module/stage. Section 3 presents the framework
for the identification of ADL using the sensors available in off-the-shelf mobile devices, the sensors
and the methods that may be used. Section 4 presents a discussion and conclusions about the new
approach proposed.
2. Related Work
Following previous research works related to the identification of ADL and the environment in
which they are carried out, this Section reviews the state of the art on the sensors available on mobile
devices (Section 2.1), data acquisition (Section 2.2), processing (Section 2.3), fusion (Section 2.4),
artificial intelligence techniques (Section 2.5), and, finally, in Section 2.6, the methods to merge
sensors’ data with users’ agenda.
2.1. Sensors
Sensors are small components that allow the acquisition of data when they are excited
responding to stimuli, often external to the device. Available in many mobile devices, namely, in
smartphones, sensors can be used to infer an ADL, and the combination of the data from multiple
sensors can increase the efficiency of ADL identification, and environment recognition [9]. The
number and types of sensors available on mobile devices is different for each mobile platform. In
general, the sensors available in mobile devices are magnetic/mechanical sensors, environmental
sensors, location sensors, motion sensors, imaging/video sensors, proximity sensors, acoustic sensors,
optical sensors, and force sensors, being able to capture different types of signals, such as electrical,
mechanical, acoustic and others [1,21].
Based on the classification presented in [6], sensors available on Android devices include
microphones, accelerometers, gyroscopes, magnetometers, altimeters, humidity sensors, ambient
light sensors, temperature sensors, GPS receivers, touch screens, microphones, and cameras [22,23].
In addition to platform-dependent restrictions in the use of sensors, the hardware differences
between devices can influence the availability of specific sensors. Thus, the sensors available in most
of the mobile devices, presented in Table 1, are the accelerometer, the gyroscope, the magnetometer,
the GPS, the microphone, the touch screen, and the camera.
Table 1. List of sensors available in mobile devices.
Categories:
Motion sensors
Magnetic/mechanical sensors
Location sensors
Acoustic sensors
Force sensors
Imaging/video sensors
Sensors:
Accelerometer
Gyroscope
Magnetometer
GPS
Microphone
Touch screen
Camera
Availability
Always present
Present in most models
Present in most models
Always present
Always present
Always present
Always present
2.2. Data Acquisition
Data acquisition consists in the process of receiving the different types of data from the sensors
available in the mobile devices. There are some possible problems that occur during the data
acquisition process, including the influence of the unpredictable and uncontrolled external
environment, the variability of the sampling rate of sensors, the number of tasks performed by the
mobile device during the data acquisition, and the variability of the sensors chosen as input for a
given developed framework [24]. Related to the variability of the position of the smartphone when
carried by a user, to the best of the authors’ knowledge, there are no studies that solve this issue. As
a standard method was not previously defined for the correct data acquisition and processing, and
the sensors and capabilities of the mobile devices are different between manufacturers, the authors
assumed that the results are nonetheless comparable.
83
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
4 of 22
In order to improve the data acquisition process, several frameworks have been developed,
including Acquisition Cost-Aware QUery Adaptation (ACQUA) framework [17], Orchestrator
framework [25], ErdOS framework [26], LittleRock prototype [27], Jigsaw continuous sensing engine
[28], SociableSense framework [29], Continuous Hand Gestures (CHG) technique [30], and Barbie-Q
(BBQ) approach [31].
The ACQUA framework allows to control the order of data acquisition, the correct segments of
the data requested, the calibration of the data acquisition rates, the packet sizes and radio
characteristics, the adaptation of the dynamic changes in query selective properties, and the support
of multiple queries and heterogeneous time window semantics from all the sensors available in
mobile devices, reducing the energy consumption of the real-time data acquisition [17].
The Orchestrator framework promotes the distributed execution of data acquisition using
several mobile devices, and all devices execute a part of the data processing, avoiding to reduce the
requirements related to the processing power and energy consumption [25].
The same purpose of Orchestrator framework is achieved from ErdOS framework and
LittleRock prototype, distributing the data acquisition and processing processes by all resources
available in the devices used, and reducing the energy needed to process the data collected from all
sensors [26,27].
The Jigsaw continuous sensing engine implements a method to control the different sample
rates, adapting the data acquisition and processing for the different capabilities of the sensors [28].
The SociableSense framework has a mechanism to adapt the different sample rates of all sensors
used and it is a cloud-based framework, reducing the local data processing, but restricting the use of
the framework to the availability of the Internet connection [29].
The authors of [30] implemented a CHG technique for the data acquisition with Windows
Phone-based smartphones and low processing capabilities, capturing accelerometer and gyroscope
data, storing the sensory data in the smartphone memory.
The BBQ framework applies a multi-dimensional Gaussian probability density function from all
the sensors, inferring the order of the data acquisition with conditional probabilities [31].
The data acquisition process implemented in mobile devices may be performed without the use
of frameworks, improving only the data processing according to the different resource capabilities.
The authors of [32–35] implement the data acquisition process from accelerometer data in Apple
iPhone and Android-based smartphones for the identification of several activities, including driving,
walking, sitting, standing, running, and jumping activities. The authors of [36] implemented a Cursor
Movement Algorithm to detect several activities, capturing the real-time data from the accelerometer
and storing them into a local database in the mobile device.
Table 2 presents a summary of the data acquisition methods and their main characteristics for
further implementation in the proposed new approach.
Table 2. Summary of the data acquisition methods.
Methods:
ACQUA framework [17]
Orchestrator framework [25]
ErdOS framework [26]
84
Advantages:
Controls of the order of the data acquisition;
Controls the correct segments of the data requested;
Controls the calibration of the data acquisition rates;
Controls the packet sizes and radio characteristics;
Controls the adaptation of the dynamic changes in query selective
properties;
Controls the support of multiple queries and heterogeneous time
window semantics;
Adapted for low processing, memory, and energy capabilities.
Distributed execution of the data acquisition using several mobile
devices;
Adapted for low processing, memory, and energy capabilities.
Distributed execution of the data acquisition using several mobile
devices;
Adapted for low processing, memory, and energy capabilities.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
LittleRock prototype [27]
Jigsaw continuous sensing
engine [28]
SociableSense framework [29]
CHG technique [30]
BBQ framework [31]
Cursor movement algorithm
[36]
No framework
5 of 22
Adapted for low processing, memory, and energy capabilities.
Controls the different sample rates;
Adapted for low processing, memory, and energy capabilities.
Cloud-based framework;
Needs a constant Internet connection;
Adapted for low processing, memory, and energy capabilities.
Stores the sensory data in the smartphone memory;
Adapted for low processing, and energy capabilities.
Uses a multi-dimensional Gaussian probability density function
from all sensors;
Adapted for low processing, memory, and energy capabilities.
Stores the sensory data in the smartphone memory;
Adapted for low processing, and energy capabilities.
Adapted for low processing, memory, and energy capabilities.
2.3. Data Processing
After the data acquisition process, the sensors’ data should be processed in order to prepare the
data for the fusion from the chosen set of sensors, and, consequently, the application of the methods
for ADL recognition. First, data processing should validate the integrity and quality of the data, and,
then, applying data cleaning and/or data imputation techniques [37], in order to make this data
available for the next stage in the processing pipeline of the framework. However, data processing
depends on the environmental conditions, the types of sensors and data, the events of sensor failures,
and the capabilities of the mobile devices [38]. Several techniques have been developed to reduce the
memory and energy consumption of the data processing techniques. Other issues related to sensor
drifting and generic noise are not specifically addressed in this paper, despite recognizing that
sensors’ calibration and drift compensation may improve the outcomes of automatic recognition
algorithms. Nevertheless, the application of data cleaning techniques mentioned in Section 2.3.1, and
data imputation techniques mentioned in Section 2.3.2 may reduce the impact of drift and noise.
Additionally, both the limited acquisition time used in the proposed framework and the fusion of
data from different sensors, as discussed in [39], help in reducing the aforementioned effects. For each
sensor data capture, we show that the use of only 5 s of sensors’ data is sufficient for the recognition
of ADL and the environment. As a consequence the risk of failure in data acquisition or data
corruption over such a short time may be assumed negligible.
The ACQUA framework is also used to optimize the data processing, by automated storage and
retrieval system (ASRS) algorithms [17]. Other studies have presented approaches to adapt the data
processing methods to the low capabilities of the mobile devices, processing the data after splitting
or using methods with limited resources needed [24,40–42].
The use of data cleaning methods, presented in Section 2.3.1, is important to decrease the
influence of the environmental conditions noise or systems failures. In order to improve the results,
when the data acquisition fails, Section 2.3.2 presents the possible data imputation methods to correct
the data acquired. However, these methods are not addressed by the proposed framework for the
identification of ADL and their environments, assuming that the data acquired is sufficient for the
extraction of several features from the, presenting the feature extraction methods and possible
features to extract, in Section 2.3.3.
2.3.1. Data Cleaning
Data cleaning consists in the identification of the incorrect values, removing outlier values and
smoothing and filtering the invalid values obtained during the data acquisition process, commonly
considered as noisy values [43–45]. Using data cleaning methods, the influence of the environmental
conditions, the mobile device position, and system failures occurred during the data acquisition
process is reduced. The efficiency of these methods depends on the type of data acquired and
spatiotemporal characteristics of the data acquired.
85
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
6 of 22
The authors of [46] proposed a weighted moving average (WMA) algorithm that collects the
sensors’ data and computes the weighted moving average, applying the WMA filter for the
normalization and cleaning of the sensors’ data.
Three types of filters are used for the motion and magnetic/mechanical sensors: the low-pass
filter (LPF), the high pass filter (HPF), and the KALMAN filter [47,48]. The WMA filter and the
different types of Fourier transforms, such as Discrete Fourier Transform (DFT), Inverse Discrete
Fourier Transform (IDFT), and Fast Fourier Transform (FFT) are also used to filter the acoustic data
[49,50].
Table 3 presents a summary of the data cleaning methods related to the different types of
sensors, discussed in section 2.1. Concerning the implementation in the development of a framework
for the identification of ADL and their environments, it can be seen that the LPF is commonly used
in motion and magnetic sensors, the most used technique for acoustic sensors is the FFT and that the
filtering techniques are not important for location, force and imaging sensors because of the nature
of the values these sensors return.
Table 3. Relation between the types of sensors and the data cleaning techniques allowed.
Types of Sensors:
Motion sensors;
Magnetic/mechanical sensors.
Location sensors
Acoustic sensors
Force sensors
Imaging/video sensors
Data Cleaning Techniques:
Low-Pass Filter; High-Pass Filter; KALMAN Filter; Weighted
moving average (WMA) algorithm; Moving average filter.
The data cleaning technique is not important for this type of data
acquired.
Moving average filter; Discrete Fourier Transform (DFT); Inverse
Discrete Fourier Transform (IDFT); Fast Fourier Transform (FFT).
The data cleaning technique is not important for this type of data
acquired.
2.3.2. Data Imputation
During the data processing, the verification of the existence of faulty data is performed to flag
that some values are missing in some instants of the acquired data time series. The data imputation
methods are mainly used for motion sensors and magnetic/mechanical sensors. However, for the
development of the new approach of the framework for the identification of ADL and their
environments, the data imputation techniques were not considered, assuming that data acquired by
the sensors is complete. Thus, in this section, the best methods for data imputation will be presented
based on a literature review.
Faulty data may have different types that can be classified as Missing Completely At Random
(MCAR), Missing At Random (MAR) and Missing Not At Random (MNAR) [51]. When the faulty
data is randomly distributed during the time interval for the data acquisition, the classification of this
data is MCAR. The other types of faulty data are MAR, verified when the faulty data is randomly
distributed by different subsets of the data acquired, and MNAR, and verified when the faulty data
is distributed by defined instants of the data acquisition.
The K-Nearest Neighbor (k-NN) method is one of the most used methods for data imputation
of data acquired from motion, and magnetic/mechanical sensors [52–55]. The k-NN method has
several variants that can be used for data imputation, such as MKNNimpute (K-nearest neighbor
imputation method based on Mahalanobis distance), SKNNimpute (sequential K-nearest neighbor
method-based imputation), and KNNimpute (K-nearest neighbor imputation) [52,53].
The clustering techniques are also used for the data imputation, including K-means clustering,
K-means-based imputation, and fuzzy C-means clustering imputation [51,56,57], which are
implement in the Imputation Tree (ITree) method presented in [51].
There are other methods related to data imputation, including multiple imputation [58], hot/cold
imputation [59], maximum likelihood [60], Bayesian estimation [60], expectation maximization
[54,61,62], discarding instances [18], pairwise deletion [18], unconditional mean imputation [18],
conditional mean imputation [18], hot deck imputation [18], cold deck imputation [18], substitution
method [18], linear regression [18], logistic regression [18], expectation-maximization (EM) algorithm
86
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
7 of 22
[18], probabilistic neural networks [18], fuzzy min–max neural networks [18], general regression auto
associative neural network [18], tree-based methods [18], multi-matrices factorization model (MMF)
[63], mean imputation (MEI) [54,62], Multivariate Imputation by Chained Equations (MICE) [54,62],
Fourier method [62], and Fourier and lagged k-NN combined system (FLk-NN) [54,62,64].
In general, these methods can be applied to data collection from motion and
magnetic/mechanical sensors. Data imputation methods can also be applied to the acoustic data,
being the more common the k-NN methods and singular value decomposition (SVD) algorithms [65].
As the data imputation methods should be able to impute the empty instances of the data
acquired by motion and magnetic/mechanical sensors, the methods that are able to be used with this
purpose are MEI, EM, MICE, and FLk-NN [54]. However, k-NN can be applied with the comparison
between the history of the data acquisition that is similar to the data acquired in the stream with
faulty values [54]. It emerges from the reviewed literature that data imputation may be avoided for
acoustic and location sensors, because of the slow variability of their signals.
2.3.3. Feature Extraction
The correct definition of the features extracted from the sensors’ data increases the accuracy of
the identification of ADL and their environments. This definition depends on the types of sensors
and the data acquired, but also on the purpose of their final use.
For the correct extraction of the features for the motion and magnetic/mechanical sensors’ data,
the Euclidean norm for each instant of outputs from the sensors defined as magnitude of vector (MV).
Thus, the features that should be extracted from the motion and magnetic/mechanical sensors are the
mean for each axis [66–69], variance of MV [70,71], mean of MV [67,70–75], median of MV [70,74],
maximum of MV [66,70,71,73], minimum of MV [66,70,71,73], standard deviation of MV [66,67,
70–75], Root Mean Square (RMS) of MV [66,70], average of peak frequency (APF) of each axis [66],
maximum of each axis [66,69,74], minimum of each axis [66,69,74], standard deviation of each axis
[66,68,69], RMS of each axis [66], cross-axis signals correlation [66,67,69,73,76], Fast Fourier Transform
(FFT) spectral energy [70,76], frequency domain entropy [76], FFT coefficients [70,73], logarithm of
FFT [76], skewness of each axis [67], kurtosis of each axis [67], average absolute deviation of each axis
[67], time between peaks [72], Interquartile range of MV [71,73], skewness of MV [71], kurtosis of MV
[71], wavelet energy of MV [73], average of peak values [77], average of peak rising time [77], average
of peak fall time [77], average time per sample [77], average time between peaks [77], slope for each
axis [74], binned distribution for each axis [68], percentiles of MV [75], and zero crossing rate for each
axis [69].
Related to the motion and magnetic/mechanical sensors’ data, the most used features are mean,
standard deviation, maximum, minimum, median, correlation, variance, and FFT spectral energy of
MV.
For the correct extraction of the features for the acoustic sensors’ data, the features that should
be extracted are average [78], thresholding [78], minimum [78], maximum [78], distance [78], and
MFCC (Mel-frequency cepstrum coefficients) [79,80].
For the location sensors, the feature that should be extracted is the distance travelled between a
time interval, in order to identify ADL with high distance travelled. The distance between two points
captured by a GPS receiver is the ellipsoidal distance, because the two points are acquired in the
geodetic coordinate system, where the calculation of this distance is measured with the Vincenty
formula [81–83].
Table 4 presents a summary of the features extracted for each type of sensors presented in the
Section 2.1, for further implementation the in new approach for the development of a framework for
the identification of ADL and their environments.
87
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
8 of 22
Table 4. Relation between sensors and extracted features.
Types of Sensors:
Motion sensors;
Magnetic/mechanical
sensors.
Location sensors
Acoustic sensors
Force sensors;
Imaging/video sensors.
Features:
Mean [67,70–75], average of peak frequency (APF) [66],
maximum [66,70,71,73], minimum [66,70,71,73], standard deviation [66,67,
70–75], Root Mean Square (RMS) [66,70], cross-axis signals correlation
[66,67,69,73,76], skewness [67], kurtosis [67], average absolute deviation [67],
slope [74], binned distribution [68], and zero crossing rate for each axis [69];
Mean [67,70–75], median [70,74], variance [70,71], maximum [66,70,71,73],
minimum [66,70,71,73], standard deviation [66,67,70–75], Root Mean Square
(RMS) [66,70], Fast Fourier Transform (FFT) spectral energy [70,76], frequency
domain entropy [76], FFT coefficients [70,73], logarithm of FFT [76],
Interquartile range [71,73], skewness [67], kurtosis [67], wavelet energy [73],
and percentiles of MV [75]; Time between peaks [72], average of peak values
[77], average of peak rising time [77], average of peak fall time [77], average
time between peaks [77].
Distance between two points.
Average [78], Thresholding [78], Minimum [78], Maximum [78], Distance [78],
MFCC (Mel-frequency cepstrum coefficients) [79,80].
These sensors are not useful for the development of the framework for the
Identification of ADL and their environments.
2.4. Data Fusion
After the extraction of the features, the data acquired from all sensors should be fused to improve
the accuracy of the ADL identification and their environments in the new approach for the framework
proposed in this study [11]. The data fusion methods implemented should be related with the final
purpose of the framework presented in Section 2.6.
Based on the literature studies presented by several authors [12,20,84,85], the data fusion
methods are grouped in four categories [12,84,85]. These are: probabilistic methods, statistical
methods, knowledge base theory methods and evidence reasoning methods.
The probabilistic methods [12,20,84,85] include Bayesian analysis methods, maximum likelihood
methods, state-space models, evidential reasoning, possibility theory, Kalman Filter [86,87], Particle
filtering, k-Nearest Neighbor (k-NN), k-Means, optimal theory, uncertainty ellipsoids, Gaussian
mixture model (GMM), weighted averages, and regularization.
The statistical methods [12,84,85] for data fusion include covariance intersection, crosscovariance, and other robust statistics. However, other statistical methods used for data fusion are
dynamic time warping (DTW) [88], which measures the similarity between two temporal sequences,
based on the raw data or the features extracted.
The knowledge base theory methods [12,20,84,85,89] for data fusion include Artificial Neural
Networks (ANN), Support Vector Machines (SVM), Decision Trees, Deep Learning, Long Short Term
Memory (LSTM) Recurrent Neural Networks (RNN), Fuzzy Logic, Topic models, and Genetics
Algorithms.
The evidence reasoning methods [12,84,85] for data fusion include evidence theory, Bayesian
network, Dempster-Shafer, and recursive operators.
Based on these categories of data fusion methods, several implementations have been performed
and presented in several studies for the identification of a plethora of a real-life activities and
environments. The Rao-Blackwellization unscented Kalman filter (RBUKF) [90] was implemented to
fuse the data acquired from a compass, a gyroscope, and a GPS receiver. The Kalman filter was used
to fuse the data acquired from the GPS receiver and the gyroscope in order to support a navigation
system [91]. The Naïve Bayes classifier is used to fuse the data acquired from acoustic, accelerometer
and GPS sensors to recognize different situations during daily life [92]. The AutoregressiveCorrelated Gaussian Model was implemented in the KNOWME system [93]. Bayesian analysis and
Kalman filter where used to data acquired from the several sensors available in mobile devices for
the identification of the ADL [94]. The CHRONIOUS system implements several methods to
88
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
9 of 22
recognize several ADL, such as Support Vector Machine (SVM), random forests, Artificial Neural
Networks (ANN), decision trees, decision tables, and Naïve Bayes classifier, in order to fuse the data
collection from several sensors available in mobile devices [95]. In [96], the authors used the empirical
mode decomposition (EMD) applied to the inertial sensors available in a mobile device, including
accelerometer, gyroscope, and magnetometer, for the identification of several ADL. The authors of
[97] implements several methods for data fusion, including SVM, random forest, hidden Markov
models (HMMs), conditional random fields (CRFs), Fisher kernel learning (FKL), and ANN for
several sensors, such as Accelerometer, RFID, and Vital monitoring sensors for the correct
identification of ADL.
Table 5 presents a summary of the data fusion methods that can be applied for each type of
sensors presented in Section 2.1, for further implementation in a new approach for the development
of a framework for the identification of ADL and their environments.
Table 5. Relation between the different types of sensors and some data fusion methods.
Types of sensors:
Motion sensors;
Magnetic/mechanical
sensors;
Location sensors;
Acoustic sensors.
Force sensors;
Imaging/video sensors.
Data fusion methods:
Autoregressive-Correlated Gaussian Model;
Fuzzy Logic;
Dempster-Shafer;
Evidence Theory;
Recursive Operators;
Support Vector Machine (SVM);
Random Forests;
Artificial Neural Networks (ANN);
Decision Trees;
Naïve Bayes classifier;
Bayesian analysis;
Kalman Filter;
k-Nearest Neighbor (k-NN);
Least squares-based estimation methods;
Optimal Theory;
Long Short Term Memory (LSTM) Recurrent Neural Networks (RNN);
Uncertainty Ellipsoids.
These sensors are not useful for the development of the framework for the
Identification of ADL and their environments.
2.5. Identification of Activities of Daily Living
The definition of the methods for ADL identification represents the final module of the new
proposed framework, presented in Figure 1. The identification of the ADL and their environments
depends on the sensors’ data used, therefore, if a method uses the data acquired from motion and/or
magnetic/mechanical sensors, it will probably be used to identify the ADL. If a method uses the data
acquired from acoustic sensors, it will probably be used to identify the external environments.
Finally, if the implemented method uses the location sensors, it is probably identifying activities with
fast movement, e.g., driving, or it is probably trying to identify the place where the ADL is performed.
In general, the identification of ADL is performed at the same time of the data fusion, because the
methods use the same techniques.
The machine learning is a set of several techniques for artificial intelligence, including the
techniques for the identification of ADL and their environments. The concept of machine learning
will be presented in the Section 2.5.1. In Section 2.5.2, the pattern recognition methods are presented,
which consists in a subset of the machine learning techniques.
89
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
10 of 22
Figure 1. Schema for the framework for the recognition of Activities of Daily Living (ADL).
2.5.1. Machine Learning
Artificial Intelligence (AI) is one of the main areas for the development of computer science
systems, and machine learning is composed by a subset of AI methods, where the computers have
the ability to learn and perform some tasks, taking into account the external conditions of the system
in order to change the execution of some methods for obtaining of better results [98].
Machine learning methods are based on the creation and implementation of algorithms for the
recognition and prediction of several situations based on the data acquired, and these methods are
commonly classified in four categories [99,100], such as Supervised learning, Unsupervised learning,
Reinforcement learning, and Semi-supervised Learning and Active Learning.
Supervised learning methods are based on the automatic adjustment of the network parameters,
comparing the actual network output with the desired output previously defined in the data
acquired, where the error obtained is the mean squared error (MSE) [100]. The input data involved
in the supervised leaning should be labeled, in order to perform the comparisons.
Unsupervised learning methods consist on the correction of the results obtained based on the
input data, attempting to obtain the signification patterns or features in the unlabeled input data,
automatically learning with intuitive primitives like neural competition and cooperation [100].
Reinforcement learning methods are similar to supervised learning methods, but the exact
desired output is a priori unknown [100]. Thus, these methods are learning based on the feedback
provided during the execution of the algorithm by an artificial agent in order to maximize the total
expected reward [100].
Semi-supervised Learning and Active Learning methods are methods that should be applied to
datasets with a large collection of unlabeled input data and a few labeled examples to generalize the
results and performance of the method, based on assumptions related to the probability of occurrence
of some output.
For the development of a new approach for the development of a framework for the
identification of ADL and their environments, the machine learning may be used, as it can be adapted
to bioinformatics and human-related systems [101–104]. Pattern recognition methods, described in
Section 2.5.2, consist on a subset of machine learning methods for the recognition of patterns [105],
which are very useful in the development of the framework for the identification of ADL and their
environments.
2.5.2. Pattern Recognition
The use of pattern recognition methods is the final part of research for the creation of a new
approach for a framework for the identification of ADL and their environments. Several sensors,
presented in Section 2.1, may be used with pattern recognition methods, which should be applied to
the features extracted from the input data.
90
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
11 of 22
The methods implemented during the pattern recognition step are similar to the methods
implemented for the data fusion, presented in Section 2.4. As reported early in this paper, the data
fusion and pattern recognition may be confused, and the pattern recognition is performed at the same
time of the data fusion. The categorization of the methods is similar to the methods applied for data
fusion, and they are separated in four categories [12,84,85], these are the probabilistic methods, the
statistical methods, the knowledge base theory methods and the evidence reasoning methods.
Several ADL may be recognized with pattern recognition methods, as example for the
recognition of standing, and walking activities may be used ANN [106]. Several authors [13–16,
66–69,71–76,89,107–120] proposed the use of the ANN, probabilistic neural networks (PNN), deep
neural networks (DNN), Long Short Term Memory (LSTM) Recurrent Neural Networks (RNN),
SVM, Random Forest, Bayesian Network, Sequential Minimal Optimization (SMO), Logistic
Regression, Naïve Bayes, C4.5 Decision Tree, Logistic Model Trees (LMT), J48 Decision tree,
K-Nearest Neighbor (KNN), and Simple Logistic Logit Boost methods for the recognition of walking,
running, jogging, jumping, dancing, driving, cycling, sitting, standing, lying, walking on stairs, going
up on an escalator, laying down, walking on a ramp activities, cleaning, cooking, medication,
sweeping, washing hands, and watering plants.
The Hidden Markov Model (HMM) and their variants are also a pattern recognition
implemented in several studies related with the identification of ADL and their environments, such
as the Hidden Markov Model (HMM) [71], the Hidden Markov Model Ensemble (HMME) [121], the
Sliding-Window-based Hidden Markov Model (SW-HMM) [113]. The ADLs commonly identified by
the HMM method are walking, walking on stairs, standing, running, sitting, and laying.
Table 6 presents a summary of the pattern recognition methods that can be applied for each type
of sensors presented in Section 2.1, for further implementation in the proposed approach for the
identification of ADL and their environments. As shown in the Table, the HMM method is commonly
used for the recognition of walking, walking on stairs, standing, running, sitting and laying activities,
whereas the SVM, ANN and their variants, HMM and Random Forest methods, are useful for the
recognition of complex activities (e.g., cleaning, cooking, medication, sweeping, washing hands and
watering plants). However, all of the described methods in this study may be used for the recognition
of simple activities (e.g., walking, running, jogging, jumping, dancing, driving, cycling, sitting,
standing, lying, walking on stairs, going up on an escalator, laying down and walking on a ramp)
with reliable accuracy.
Table 6. Relation between the different types of sensors and some pattern recognition methods.
Types of sensors:
Motion sensors;
Magnetic/mechanical
sensors;
Location sensors;
Acoustic sensors.
Pattern recognition
methods:
Support Vector Machines
(SVM);
Decision trees (J48, C4.5);
Artificial Neural Networks
(ANN);
Probabilistic Neural
Networks (PNN);
Deep Neural Networks
(DNN);
Long Short Term Memory
(LSTM) Recurrent Neural
Networks (RNN);
k-Nearest Neighbour (KNN);
Naïve Bayes;
Random Forest;
Logistic Regression;
Bayesian network;
Sequential minimal
optimization (SMO);
ADL recognized:
Walking; running; jogging; jumping; dancing; driving,
cycling; sitting; standing; lying; walking on stairs; going
up on an escalator; laying down; walking on a ramp.
91
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
Force sensors;
Imaging/video sensors.
12 of 22
Logistic Model Trees (LMT);
Simple Logistic Logit Boost.
Support Vector Machines
(SVM);
Artificial Neural Networks
(ANN);
Probabilistic Neural
Networks (PNN);
Deep Neural Networks
Cleaning; cooking; medication; sweeping; washing
(DNN);
hands; watering plants.
Long Short Term Memory
(LSTM) Recurrent Neural
Networks (RNN);
Hidden Markov model
(HMM);
Random Forest.
Hidden Markov model
Walking; walking on stairs; standing; running; sitting;
(HMM).
laying.
These sensors are not useful for the development of the framework for the Identification
of ADL and their environments.
2.6. Relation between the Identification of Activities of Daily Living and User Agenda
After the identification of the ADL and their environments with machine learning methods, the
results obtained should be compared with the users’ agenda for the validation of the scheduled
activities performed during the daily life. By comparing the identified ADL with the user’s agenda,
it will be possible to monitor the lifestyle [122] and provide feedback regarding planned activities
and executed activities. However, the inputs from agenda can also be used to validate the accuracy
of the framework developed [123].
3. Methods and Expected Results
The new approach proposed for the creation of the framework for the identification of ADL
(Figure 1) is based on [6,8,9], and it is composed by several stages. They are: the selection of the
sensors, the data and processing, including data cleaning, imputation, and feature extraction, data
fusion, the identification of ADL with artificial intelligence, including pattern recognition, and other
machine learning techniques, and, at the end, the combination of the results obtained with the data
available in the users’ agenda.
In order to create a new approach for a framework for the identification of ADL and their
environments, the architecture, presented in Figure 1, and set of methods presented in Section 2 are
proposed for obtaining results with reliable accuracy.
Following the list of sensors available in off-the-shelf mobile devices, presented in Section 2.1,
the sensors that will be used in the framework should be dynamically selected, according to the
sensors available in the mobile device. Thus, the types of sensors selected to use in the framework
will be motion sensors, magnetic/mechanical sensors, acoustic sensors, and location sensors. The
accelerometer is available in all mobile devices, but the gyroscope is only available on some devices,
therefore, to cover the execution of the framework in all devices, two different methods should be
implemented, one considering the data from the accelerometer and the gyroscope, and another
considering only the data from the accelerometer. The magnetometer is only available on some
devices, therefore this sensor should be managed similarly. Related to the acoustic sensors, the
microphone is available in all mobile devices. As to the location sensors, the GPS is available in most
of the mobile devices and its data should be used in the framework whenever possible.
The data acquisition methods are not directly related to the development of the framework,
because the different manufacturers of the mobile operating systems have different methodologies
to acquire the different types of sensors’ data. Thus, the data acquisition methods, presented in
Section 2.2, should take in account the limitations of the mobile devices. Based on previous research
92
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
13 of 22
studies and preliminary experiments, acquiring only 5 seconds of data from the selected sensors
every 5 min is sufficient for the identification of the ADL and environments.
Following the creation of the new approach for a framework for the identification of ADL and
their environments, the selection of data processing methods, presented in Section 2.3, should contain
the data cleaning, data imputation, and feature extraction methods.
The data cleaning methods adapted for the framework depends on the types of sensors. On the
one hand, for the accelerometer, gyroscope, and magnetometer sensors, the data cleaning method
that should be applied is a low pass filter to remove the noise and the value of the gravity acquired
during the data acquisition process. On the other hand, for the acoustic sensors, the data cleaning
method that should be applied is the FFT in order to extract the frequencies of the audio. As the
location sensors return values that are in nature already a result (e.g., GPS coordinates), data cleaning
methods are not significant. Nevertheless, and as future work, it may be necessary to implement
algorithms that increase the accuracy of these sensors as to better contribute to a quality data fusion
process.
The data imputation methods is not important to implement in the development of a new
approach for a framework for the identification of ADL and their environments, assuming that the
data acquired from all sensors is always filled.
Related to the feature extraction, the features needed to recognize the ADL and their
environments should be selected based on the type of sensors and on the selected features already
reported in the literature and presented in Section 2.3.3. Firstly, the features selected for the
accelerometer, gyroscope, and magnetometer sensors are the five greater distances between the
maximum peaks, the average of the maximum peaks, the standard deviation of the maximum peaks,
the variance of the maximum peaks, the median of the maximum peaks, the standard deviation of
the raw signal, the average of the raw signal, the maximum value of the raw signal, the minimum
value of the raw signal, the variance of the of the raw signal, and the median of the raw signal.
Secondly, the features selected for the microphone are the standard deviation of the raw signal, the
average of the raw signal, the maximum value of the raw signal, the minimum value of the raw signal,
the variance of the of the raw signal, the median of the raw signal, and 26 MFCC coefficients. Finally,
the features selected for the GPS receiver are the distance travelled during the acquisition time.
Before the presentation of the data fusion and pattern recognition methods that should be used
for in the framework, the ADL and environments to recognize should be defined. This process should
be executed with several sensors, that will be combined as presented in the Figure 2 and Table 7,
being these the necessary stages:
1.
2.
3.
Firstly, the ADL are recognized with motion and magnetic/mechanical sensors;
Secondly, the identification of the environments is performed with acoustic sensors;
Finally, there are two options, being these:
o
o
The identification of standing activities with the fusion of the data acquired from motion
and magnetic/mechanical sensors, and the environment recognized, where the number
of ADL recognized depends on the number of sensors available;
The identification of standing activities with the fusion of the data acquired from
motion, magnetic/mechanical and location sensors, and the environment recognized,
where the number of ADL recognized depends on the number of sensors available.
In identifying the environments, what is intended is to identify the associated activity, i.e., the
sound generated in a classroom is not only the sound of the room itself, but rather the sound of a
class who is having a lesson in a classroom. This is to say that an environment is to be considered as
a place where some activity occurs in a given time of the day or the week, so there will be the need
to consider different types of “Street” environments as they will have different audio signatures at
different times of the day or week and of course, in different streets. All the proposed environments
shown in Figure 2 are expected to be plural.
93
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
14 of 22
Figure 2. Sensors used for the recognition of Activities of Daily Living (ADL) and environments for
each phase of development.
Environments
Activities
Table 7. Sensors, Activities of Daily Living (ADL), and environments for recognition with the
framework proposed.
Going Downstairs
Going Upstairs
Running
Walking
Standing
Sleeping
Driving
Bar
Classroom
Gym
Library
Kitchen
Street
Hall
Watching tv
Bedroom
Accelerometer
✓
✓
✓
✓
✓
✓
✓
Gyroscope
✓
✓
✓
✓
✓
✓
✓
Magnetometer
✓
✓
✓
✓
✓
✓
✓
Microphone
GPS
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Firstly, the ADL to be identified with the framework will be going downstairs, going upstairs,
running, walking, and standing, because they are part of the most recognized ADL in previous
studies with reliable accuracy [7]. Secondly, the proposed environments to identify with the
framework will be bar, classroom, gym, kitchen, library, street, hall, watching TV, and bedroom,
because the existence of previous studies related to the recognition of environments is very limited,
the proposed framework will take in account the most common environments and some of the
environments previously recognized [7]. Thirdly, the proposed ADL to distinct with the framework
94
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
15 of 22
will be sleeping, and standing, because the ADL may be confused as standing ADL and the inclusion
of the environment recognized as an input for the classification method will help in the accurate
recognition of these ADL. Finally, the proposed ADL to distinct with the framework are sleeping,
standing, and driving, because the driving may also confused as standing ADL and, in order to
accurately distinct these ADL, the environment recognized and the features extracted from the GPS
receiver should be included. As the data for the creation of the methods for the recognition of ADL
and environments was acquired will several conditions and different people, the generated method
with ANN will be generic and the calibration of sensor is not needed.
Based on the list of data fusion methods and pattern recognition methods, defined in Sections
2.4 and 2.5, the method selected for the implementation in the new approach for a framework for the
identification of ADL and their environments will be based in ANN methods, because, based on the
literature, it is one of the methods that reports the best accuracies. However, the selection of the best
type of ANN will be done with the comparison of the results obtained with three types of ANN
selected. The types of ANN that will be tested to the acquired data are:
•
•
•
MLP with Backpropagation;
FNN with Backpropagation;
DNN.
Regarding the data acquired from GPS receiver, it can be useful to increase the accuracy of the
identification of the ADL and their environments, but it can also be used for the identification of the
location where the ADL are executed, in order to improve the comparison with the users’ agenda
presented in Section 2.6.
4. Discussion and Conclusions
This paper presents the architecture of a new approach for a framework for the identification of
ADL and their environments, using methods with a reported good accuracy. The development of the
new approach for the development of a framework for the identification of ADL and their
environments, based on the system presented in [6,8,9], is one of the steps for the creation of a
personal digital life coach [3] using mobile devices.
The framework will be composed by several modules several, such as data acquisition, data
processing, data fusion, and a module to implement artificial intelligence techniques for the
identification of the ADL and their environments.
The sensors used in the framework will be accelerometer, gyroscope, magnetometer,
microphone, and GPS receiver, in order to recognize several ADL, including going downstairs, going
upstairs, running, walking, standing, sleeping, and driving, and their environments, including bar,
classroom, gym, kitchen, library, street, hall, watching TV, and bedroom.
The sensors’ data should be acquired and, before the extraction of the features of the sensors’
data, filters such as low pass filter and FFT, should be applied. Afterwards, the data fusion and
pattern recognition methods should be applied for the recognition of ADL and environments.
This paper consists on a conceptual definition of the framework for the recognition of the ADL
and their environments, proposing three possible methods for this purpose, based on the use of the
ANN methods. In order to define the best method, the future implementation of the proposed
methods will compare the differences between them, including the accuracy, performance, and
adaptability for the development of a local processing framework for mobile devices. It will include
the acquisition of a large set of sensors’ data related to the ADL and environments proposed for the
creation of training and testing sets and further validation of the developed methods. Additionally,
and also as future work, the framework will allow each user to validate the ADL identified by the
framework when this is not the real performed activity.
Due to the inexistence of previous studies that review the use of all sensors available in current
off-the-shelf mobile devices, our proposed framework is a function of the number of sensors available
in the mobile device used, proving a reliable feedback in almost real-time.
95
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
16 of 22
Acknowledgments: This work was supported by FCT project UID/EEA/50008/2013. The authors would also like
to acknowledge the contribution of the COST Action IC1303–AAPELE–Architectures, Algorithms and Protocols
for Enhanced Living Environments.
Author Contributions: All the authors have contributed with the structure, content, and writing of the paper.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
96
Salazar, L.H.A.; Lacerda, T.; Nunes, J.V.; von Wangenheim, C.G. A Systematic Literature Review on
Usability Heuristics for Mobile Phones. Int. J. Mob. Hum. Comput. Interact. 2013, 5, 50–61.
Foti, D.; Koketsu, J.S. Activities of daily living. In Pedretti’s Occupational Therapy: Practical Skills for Physical
Dysfunction; Elsevier Health Sciences: Amsterdam, Netherlands, 2013; Vol. 7, pp. 157–232.
Garcia, N.M. A Roadmap to the Design of a Personal Digital Life Coach. In Advances in Intelligent Systems
and Computing; Springer: Cham, Switzerland, 2016; Vol. 399.
Kleinberger, T.; Becker, M.; Ras, E.; Holzinger, A.; Müller, P. Ambient intelligence in assisted living: Enable
elderly people to handle future interfaces. In Lecture Notes in Computer Science; Springer: Berlin, Germany,
2007; Vol. 4555.
Singh, D.; Kropf, J.; Hanke, S.; Holzinger, A. Ambient Assisted Living Technologies from the Perspectives
of Older People and Professionals. In Lecture Notes in Computer Science; Springer: Cham, Switzerland, 2017;
Vol.10410.
Pires, I.M.; Garcia, N.M.; Pombo, N.; Flórez-Revuelta, F. From Data Acquisition to Data Fusion: A
Comprehensive Review and a Roadmap for the Identification of Activities of Daily Living Using Mobile
Devices. Sensors 2016, 16, 184.
Pires, I.M.; Garcia, N.M.; Pombo, N.; Flórez-Revuelta, F.; Zdravevski, E.; Spinsante, S. Machine Learning
Algorithms for the Identification of Activities of Daily Living Using Mobile Devices: A Comprehensive
Review. engrXiv, engrxiv.org/k6rxa, 2018. (In Review)
Pires, I.M.; Garcia, N.M.; Flórez-Revuelta, F. Multi-sensor data fusion techniques for the identification of
activities of daily living using mobile devices. In Proceedings of the ECMLPKDD 2015 ECML PKDD—
European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in
Databases, Porto, Portugal, 7–11 September 2015.
Pires, I.M.; Garcia, N.M.; Pombo, N.; Flórez-Revuelta, F. Identification of Activities of Daily Living Using
Sensors Available in off-the-shelf Mobile Devices: Research and Hypothesis. In Advances in Intelligent
Systems and Computing; Springer: Cham, Switzerland, 2016; Vol. 476.
Smartphones: BQ Aquaris and BQ Portugal. Available online: https://www.bq.com/pt/smartphones
(accessed on 2 September 2017).
Banos, O.; Damas, M.; Pomares, H.; Rojas, I. On the use of sensor fusion to reduce the impact of rotational
and additive noise in human activity recognition. Sensors 2012, 12, 8039–8054.
Akhoundi, M.A.A.; Valavi, E.; Multi-Sensor Fuzzy Data Fusion Using Sensors with Different
Characteristics. arXiv preprint, arXiv:1010.6096, 2010.
Paul, P.; George, T. An Effective Approach for Human Activity Recognition on Smartphone. In Proceedings
of the 2015 IEEE International Conference on Engineering and Technology (ICETECH), Coimbatore, India,
20–20 March 2015; pp. 45–47.
Hsu, Y.W.; Chen, K.H.; Yang, J.J.; Jaw, F.S. Smartphone-based fall detection algorithm using feature
extraction. In Proceedings of the 9th International Congress on Image and Signal Processing, BioMedical
Engineering and Informatics (CISP-BMEI), Datong, China, 15–17 October 2016.
Dernbach, S.; Das, B.; Krishnan, N.C.; Thomas, B.L.; Cook, D.J. Simple and Complex Activity Recognition
through Smart Phones. In Proceedings of the 8th International Conference on Intelligent Environments (IE)
Guanajuato, Mexico, 26–29 June 2012.
Shen, C.; Chen, Y.F.; Yang, G.S. On Motion-Sensor Behavior Analysis for Human-Activity Recognition via
Smartphones. In Proceedings of the 2016 Ieee International Conference on Identity, Security and Behavior
Analysis (ISBA), Sendai, Japan, 29 February–2 March 2016.
Misra, A.; Lim, L. Optimizing Sensor Data Acquisition for Energy-Efficient Smartphone-Based Continuous
Event Processing. In Proceedings of the 12th IEEE International Conference on Mobile Data Management
(MDM), Lulea, Sweden, 6–9 June 2011; pp. 88–97.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
17 of 22
D’Ambrosio, A.; Aria, M.; Siciliano, R. Accurate Tree-based Missing Data Imputation and Data Fusion
within the Statistical Learning Paradigm. J. Classif. 2012, 29, 227–258.
Dong, J.; Zhuang, D.; Huang, Y.; Fu, J. Advances in multi-sensor data fusion: algorithms and applications.
Sensors 2009, 9, 7771–7784.
King, R.C.; Villeneuve, E.; White, R.J.; Sherratt, R.S.; Holderbaum, W.; Harwin, W.S. Application of data
fusion techniques and technologies for wearable health monitoring. Med. Eng. Phys. 2017, 42, 1–12.
White, R.M. A Sensor Classification Scheme. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 1987, 34, 124–
126.
Bojinov, H.; Michalevsky, Y.; Nakibly, G.; Boneh, D. Mobile device identification via sensor fingerprinting.
arXiv preprint, arXiv:1408.1416, 2014.
Katevas, K.; Haddadi, H.; Tokarchuk, L. Sensingkit: Evaluating the sensor power consumption in ios
devices. In Proceedings of the 12th International Conference on Intelligent Environments (IE), London, UK,
14–16 September 2016.
Bersch, S.D.; Azzi, D.; Khusainov, R.; Achumba, I.E.; Ries, J. Sensor data acquisition and processing
parameters for human activity classification. Sensors 2014, 14, 4239–70.
Kang, S.; Lee, Y.; Min, C.; Ju, Y.; Park, T.; Lee, J.; Rhee, Y.; Song, J. Orchestrator: An active resource
orchestration framework for mobile context monitoring in sensor-rich mobile environments. In
Proceedings of the 2010 IEEE International Conference on Pervasive Computing and Communications
(PerCom), Mannheim, Germany, 29 March–2 April 2010.
Vallina-Rodriguez, N.; Crowcroft, J. ErdOS: Achieving energy savings in mobile OS. In Proceedings of the
sixth international workshop on MobiArch. Bethesda, MD, USA, 28 June 2011; pp. 37–42.
Priyantha, B.; Lymberopoulos, D.; Jie, L. LittleRock: Enabling Energy-Efficient Continuous Sensing on
Mobile Phones. IEEE Pervasive Comput. 2011, 10, 12–15.
Lu, H.; Yang, J.; Liu, Z.; Lane, N.D.; Choudhury, T.; Campbell, A.T. The Jigsaw continuous sensing engine
for mobile phone applications. In Proceedings of the 8th ACM Conference on Embedded Networked
Sensor Systems, Zürich, Switzerland, 3–5 November 2010; pp. 71–84.
Rachuri, K.K.; Mascolo, C.; Musolesi, M.; Rentfrow, P.J. SociableSense: exploring the trade-offs of adaptive
sampling and computation offloading for social sensing. In Proceedings of the 17th annual international
conference on Mobile computing and networking, Las Vegas, NV, USA, 19–23 September 2011; pp. 73–84.
Gupta, H.P.; Chudgar, H.S.; Mukherjee, S.; Dutta, T.; Sharma, K. A continuous hand gestures recognition
technique for human-machine interaction using accelerometer and gyroscope sensors. IEEE Sens. J. 2016,
16, 6425–6432.
Deshpande, A.; Guestrin, C.; Madden, S.R.; Hellerstein, J.M.; Hong, W. Model-driven data acquisition in
sensor networks. In Proceedings of the Thirtieth international conference on Very large data bases—2004,
VLDB Endowment: Toronto, Canada, 31 August–3 September 2004; Vol. 30, pp. 588–599.
Kubota, H.; Kyokane, M.; Imai, Y.; Ando, K.; Masuda, S.I. A Study of Data Acquisition and Analysis for
Driver’s Behavior and Characteristics through Application of Smart Devices and Data Mining. In
Proceedings of the Third International Conference on Computer Science, Computer Engineering, and
Education Technologies, Lodz, Poland, 19–21 September 2016.
Ayu, M.A.; Mantoro, T.; Matin, A.F. A.; Basamh, S.S. Recognizing user activity based on accelerometer data
from a mobile phone. In Proceedings of the 2011 IEEE Symposium on Computers & Informatics (ISCI),
Kuala Lumpur, Malaysia, 20–23 March 2011.
Banos, O.; Garcia, R.; Holgado-Terriza, J.A.; Damas, M.; Pomares, H.; Rojas, I.; Saez, A.; Villalonga, C.
mHealthDroid: A novel framework for agile development of mobile health applications. In Lecture Notes in
Comput. Sci. 2014. Springer: Cham, Switzerland, 2014; Vol. 8868.
Chavan, V.B.; Mhala, N. Development of Hand Gesture Recognition Framework Using Surface EMG and
Accelerometer
Sensor
for
Mobile
Devices.
2015.
Available
online:
https://www.irjet.net/archives/V2/i5/IRJET-V2I542.pdf (accessed on 23rd December 2017).
Sarkar, M.; Haider, M.Z.; Chowdhury, D.; Rabbi, G. An Android based human computer interactive system
with motion recognition and voice command activation. In Proceedings of the 5th International Conference
on Informatics, Electronics and Vision (ICIEV), Dhaka, Bangladesh, 13–14 May 2016.
Pires, I.M.; Garcia, N.M.; Pombo, N.; Flórez-Revuelta, F.; Rodríguez, N.D. Validation Techniques for Sensor
Data in Mobile Health Applications. J. Sens. 2016, 2016, 1687–725.
97
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
98
18 of 22
Lane, N.D.; Miluzzo, E.; Lu, H.; Peebles, D.; Choudhury, T.; Campbell, A.T. A survey of mobile phone
sensing. IEEE Commun. Mag. 2010, 48, doi: 10.1109/MCOM.2010.5560598.
Banos, O.; Toth, M.A.; Damas, M.; Pomares, H.; Rojas, I Dealing with the effects of sensor displacement in
wearable activity recognition. Sensors 2014, 14, 9995–10023.
Pejovic, V.; Musolesi, M. Anticipatory Mobile Computing. ACM Comput. Surv. 2015, 47, 1–29.
Lin, F.X.; Rahmati, A.; Zhong, L. Dandelion: A framework for transparently programming phone-centered
wireless body sensor applications for health. In Proceedings of the 10th Wireless Health, San Diego, CA,
USA, 5–7 October 2010.
Postolache, O.; Girão, P.S.; Ribeiro, M.; Guerra, M.; Pincho, J.; Santiago, F.; Pena, A. Enabling telecare
assessment with pervasive sensing and Android OS smartphone. In Proceedings of the 2011 IEEE
International Workshop on Medical Measurements and Applications Proceedings (MeMeA), Bari, Italy,
30–31 May 2011.
Jeffery, S.R.; Alonso, G.; Franklin, M.J.; Hong, W.; Widom, J. Declarative Support for Sensor Data Cleaning.
In Lecture Notes in Computer Science; Springer: Berlin, Germany, 2006; Vol. 2006.
Tomar, D.; Agarwal, S. A survey on pre-processing and post-processing techniques in data mining. Int.
J.Database Theory Appl. 2014, 7, 99–128.
Park, K.; Becker, E.; Vinjumur, J.K.; Le, Z.; Makedon, F. Human behavioral detection and data cleaning in
assisted living environment using wireless sensor networks. In Proceedings of the 2nd International
Conference on PErvasive Technologies Related to Assistive Environments, Corfu, Greece, 9–13 June 2009.
Zhuang, Y.; Chen, L.; Wang, X.S.; Lian, J. A weighted moving average-based approach for cleaning sensor
data. in Distributed Computing Systems, In Proceedings of the 27th International Conference on
Distributed Computing Systems (ICDCS'07), Toronto, ON, Canada, 25–27 June 2007.
Li, Z.; Wang, J.; Gao, J.; Li, B.; Zhou, F. A vondrak low pass filter for IMU sensor initial alignment on a
disturbed base. Sensors 2014, 14, 23803–23821.
Graizer, V. Effect of low‐pass filtering and re‐sampling on spectral and peak ground acceleration in strong‐
motion records. In Proceedings of the 15th World Conference of Earthquake Engineering, Lisbon, Portugal,
24–28 September 2012.
UiO:
Fourier
analysis
and
applications
to
sound
processing.
Available
online:
http://www.uio.no/studier/emner/matnat/math/MAT.../v12/part1.pdf (accessed on 27 August 2017)
Ninness,
B.
Spectral
Analysis
Using
the
FFT.
Available
online:
https://pdfs.semanticscholar.org/dd74/4c224d569bd9ae907b7527e7f2a92fafa19c.pdf (accessed on 27
August 2017)
Vateekul, P.; Sarinnapakorn, K. Tree-Based Approach to Missing Data Imputation. In Proceedings of IEEE
International Conference on 2009 Data Mining Workshops (ICDMW '09), Miami, FL, USA, 6 December
2009; pp. 70–75.
Ling, W.; Dong, M. Estimation of Missing Values Using a Weighted K-Nearest Neighbors Algorithm. In
Proceedings of the International Conference on 2009 Environmental Science and Information Application
Technology (ESIAT 2009), Wuhan, China; pp. 660–663.
García-Laencina, P.J.; Sancho-Gómez, J. L.; Figueiras-Vidal, A. R.; Verleysen, M. K nearest neighbours with
mutual information for simultaneous classification and missing data imputation. Neurocomputing 2009, 72,
1483–1493.
Rahman, S.A.; Rahman, S.A.; Huang, Y.; Claassen, J.; Kleinberg, S. Imputation of Missing Values in Time
Series with Lagged Correlations. In Proceedings of the 2014 IEEE International Conference on Data Mining
Workshop (ICDMW), Shenzhen, China, 14 December 2014.
Batista, G.E.; Monard, M.C. A Study of K-Nearest Neighbour as an Imputation Method. HIS 2002, 87, 251–
260.
Hruschka, E.R.; Hruschka, E.R.; Ebecken, N.F.F. Towards Efficient Imputation by Nearest-Neighbors: A
Clustering-Based Approach. In AI 2004: Advances in Artificial Intelligence; Springer: Berlin/Heidelberg,
Germany, 2004; pp. 513–525.
Luo, J.W.; Yang, T.; Wang, Y. Missing value estimation for microarray data based on fuzzy C-means
clustering. In Proceedings of Eighth International Conference on High-Performance Computing in AsiaPacific Region, Beijing, China, 30 November–3 December 2005.
Ni, D.; Leonard, J.D.; Guin, A.; Feng, C. Multiple Imputation Scheme for Overcoming the Missing Values
and Variability Issues in ITS Data. J. Trans. Eng. 2005, 131, 931–938.
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
19 of 22
Smith, B.; Scherer, W.; Conklin, J. Exploring Imputation Techniques for Missing Data in Transportation
Management Systems. Trans. Res. Rec. 2003, 1836(1), 132–142.
Qu, L.; Zhang, Y.; Hu, J.; Jia, L.; Li, L. A BPCA based missing value imputing method for traffic flow volume
data. In Proceedings of 2008 IEEE Intelligent Vehicles Symposium, Eindhoven, Netherlands, 4–6 June 2008;
pp. 985–990.
Jiang, N.; Gruenwald, L. Estimating Missing Data in Data Streams. In Proceedings of the 12th International
Conference on Database Systems for Advanced Applications (DASFAA'07), Bangkok, Thailand, 9–12 April
2007; pp. 981–987.
Rahman, S.A.; Huang, Y.; Claassen, J.; Heintzman, N.; Kleinberg, S. Combining Fourier and lagged knearest neighbor imputation for biomedical time series data. J. Biomed. Inform. 2015, 58, 198–207.
Huang, X.-Y.; Li, W.; Chen, K.; Xiang, X.-H.; Pan, R.; Li, L.; Cai, W.-X. Multi-matrices factorization with
application to missing sensor data imputation. Sensors 2013, 13, 15172–15186.
Rahman, S.A.; Huang, Y.; Claassen, J.; Kleinberg, S. Imputation of missing values in time series with lagged
correlations. In Proceedings of the 2014 IEEE International Conference on Data Mining Workshop
(ICDMW), Shenzhen, China, 14 December 2014; pp. 753–762.
Smaragdis, P.; Raj, B.; Shashanka, M. Missing Data Imputation for Time-Frequency Representations of
Audio Signals. J. Signal Process. Syst. 2010, 65, 361–370.
Bayat, A.; Pomplun, M.; Tran, D.A. A Study on Human Activity Recognition Using Accelerometer Data
from Smartphones. In Proceedings of 9th International Conference on Future Networks and
Communications (Fnc'14) / the 11th International Conference on Mobile Systems and Pervasive Computing
(Mobispc'14)/Affiliated Workshops, Ontario, Canada, 17–20 August 2014; pp. 450–457.
Khalifa, S.; Hassan, M.; Seneviratne, A. Feature selection for floor-changing activity recognition in multifloor pedestrian navigation. In Proceedings of 2014 Seventh International Conference on Mobile
Computing and Ubiquitous Networking (ICMU), Singapore, 6–8 January 2014.
Zhao, K.L.; Du, J.; Li, C.; Zhang, C.; Liu, H.; Xu, C. Healthy: A Diary System Based on Activity Recognition
Using Smartphone. In Proceedings of 2013 IEEE 10th International Conference on Mobile Ad-Hoc and
Sensor Systems (Mass 2013), Hangzhou, China, 14–16 October 2013; pp. 290–294.
Zainudin, M.N.S.; Sulaiman, M.N.; Mustapha, N.; Perumal, T. Activity Recognition based on
Accelerometer Sensor using Combinational Classifiers. In Proceedings of 2015 IEEE Conference on Open
Systems (ICOS), Bandar Melaka, Malaysia, 24–26 August 2015; pp. 68–73.
Fan, L.; Wang, Z.M.; Wang, H. Human activity recognition model based on decision tree. In Proceedings
of 2013 International Conference on Advanced Cloud and Big Data (CBD), Nanjing, China, 13–15 December
2013; pp. 64–68.
Liu, Y.Y.; Fang, Z.; Wenhua, S.; Haiyong, Z. An Hidden Markov Model based Complex Walking Pattern
Recognition Algorithm. In Proceedings of 2016 Fourth International Conference on Ubiquitous Positioning,
Indoor Navigation and Location Based Services (IEEE UPINLBS 2016), Shanghai, China, 2–4 November
2016; pp. 223–229.
Piyare, R.; Lee, S.R. Mobile Sensing Platform for Personal Health Management. In Proceedings of 18th IEEE
International Symposium on Consumer Electronics (ISCE 2014), JeJu Island, South Korea, 22–25 June 2014;
pp. 1–2.
Chen, Y.F.; Shen, C. Performance Analysis of Smartphone-Sensor Behavior for Human Activity
Recognition. IEEE Access 2017, 5, 3095–3110.
Vavoulas, G.; Chatzaki, C.; Malliotakis, T.; Pediaditis, M.; Tsiknakis, M. The MobiAct Dataset: Recognition
of Activities of Daily Living using Smartphones. In Proceedings of the International Conference on
Information and Communication Technologies for Ageing Well and E-Health (ICT4AWE), Rome, Italy, 21–
22 April 2016; pp. 143–151.
Torres-Huitzil, C.; Nuno-Maganda, M. Robust smartphone-based human activity recognition using a triaxial accelerometer. In Proceedings of 2015 IEEE 6th Latin American Symposium on Circuits & Systems
(Lascas), Montevideo, Uruguay, 24–27 February 2015; pp. 1–4.
Anjum, A.; Ilyas, M.U. Activity Recognition Using Smartphone Sensors. In Proceedings of 2013 IEEE
Consumer Communications and Networking Conference (CCNC), Las Vegas, USA, 11–14 January 2013;
pp. 914–919.
Kumar, A.; Gupta, S. Human Activity Recognition through Smartphone’s Tri-Axial Accelerometer using
Time Domain Wave Analysis and Machine Learning. Int. Comput. Appl. 2015, 127, 22–26.
99
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
78.
20 of 22
Hon, T.K.; Wang, L.; Reiss, J.D.; Cavallaro, A. Audio Fingerprinting for Multi-Device Self-Localization.
IEEE/ACM Trans. Audio, Speech Lang. Process. 2015, 23, 1623–1636.
79. Sert, M.; Baykal, B.; Yazici, A. A Robust and Time-Efficient Fingerprinting Model for Musical Audio. In
Proceedings of 2006 IEEE International Symposium on Consumer Electronics, St. Petersburg, Russia, 28
June–1 July 2007.
80. Ramalingam, A.; Krishnan, S. Gaussian Mixture Modeling of Short-Time Fourier Transform Features for
Audio Fingerprinting. IEEE Trans. Inform. Forens. Secu. 2006, 1, 457–463.
81. Vincenty, T. Direct and Inverse Solutions of Geodesics on the Ellipsoid with Application of Nested
equations. Surv. Rev. 1975, 22, 88–93.
82. Karney, C.F.F. Algorithms for Geodesics. J. Geodesy 2013, 87, 43–55.
83. Karney, C.F.F.; Deakin, R.E. The calculation of longitude and latitude from geodesic measurements. Astron.
Nachr. 2010, 331, 852–861.
84. Khaleghi, B.; Khamisa, A.; Karraya, F.O.; Razavib, S.N. Multisensor data fusion: A review of the state-ofthe-art. Inf. Fusion 2013, 14, 28–44.
85. Pombo, N.; Bousson, K.; Araújo, P.; Viana J. Medical decision-making inspired from aerospace multisensor
data fusion concepts. Inform Health Soc. Care 2015, 40, 185–97.
86. Durrant-Whyte, H.; Stevens, M.; Nettleton, E. Data fusion in decentralised sensing networks. In
Proceedings of the 4th International Conference on Information Fusion, Montreal, Canada, 7–10 August
2001.
87. Tanveer, F.; Waheed, O.T.; Atiq-ur-Rehman. Design and Development of a Sensor Fusion based Low Cost
Attitude Estimator. J.Space Technol. 2011, 1, 45–50.
88. Ko, M.H.; Westa, G.; Venkatesha, S.; Kumarb, M. Using dynamic time warping for online temporal fusion
in multisensor systems. Inf. Fusion 2008, 9, 370–388.
89. Singh, D.; Merdivan, E., Psychoula, I.; Kropf, J.; Hanke, S.; Geist, M.; Holzinger, A. Human activity
recognition using recurrent neural networks. In Proceedings of the International Cross-Domain Conference
for Machine Learning and Knowledge Extraction, Reggio, Italy, 29 August–1 September, 2017.
90. Zhao, L.; Wu, P.; Cao, H. RBUKF Sensor Data Fusion for Localization of Unmanned Mobile Platform. Res.
J. Appl. Sci. Eng. Technol. 2013, 6, 3462–3468.
91. Walter, O.; Schmalenstroeer, J.; Engler, A.; Haeb-Umbach, R. Smartphone-based sensor fusion for
improved vehicular navigation. In Proceedings of 2013 10th Workshop on Positioning Navigation and
Communication (WPNC), Dresden, Germany, 20–21 March 2013.
92. Grunerbl, A.; Muaremi A.; Osmani V.; Bahle G.; Ohler S.; Tröster G.; Mayora O.; Haring C.; Lukowicz P.
Smart-Phone Based Recognition of States and State Changes in Bipolar Disorder Patients. IEEE J Biomed.
Health Inform 2015, 15, 140–148.
93. Thatte, G.; Li, M.; Lee, S.; Emken, B.A.; Annavaram, M.; Narayanan, S.; Narayanan, D.; Mitra, U. Optimal
Time-Resource Allocation for Energy-Efficient Physical Activity Detection. IEEE Trans. Signal Process 2011,
59, 1843–1857.
94. Bhuiyan, M.Z.H.; Kuusniemi, H.; Chen, L.; Pei, L.; Ruotsalainen, L.; Guinness, R.; Chen, R. Performance
Evaluation of Multi-Sensor Fusion Models in Indoor Navigation. Eur. J. Navig. 2013, 11, 21–28.
95. Bellos, C.; Papadopoulos, A.; Rosso, R.; Fotiadis, D.I. Heterogeneous data fusion and intelligent techniques
embedded in a mobile application for real-time chronic disease management. Conf. Proc. IEEE Eng. Med.
Biol. Soc. 2011, 2011, 8303–8306.
96. Ayachi, F.S.; Nguyen, H.; Goubault, E.; Boissy, P.; Duval, C. The use of empirical mode decompositionbased algorithm and inertial measurement units to auto-detect daily living activities of healthy adults. IEEE
Trans. Neural Syst. Rehabilit. Eng. 2016, 24, 1060–1070.
97. Debes, C.; Merentitis, A.; Sukhanov, S.; Niessen, M.; Frangiadakis, N.; Bauer, A. Monitoring activities of
daily living in smart homes: Understanding human behavior. IEEE Signal Process. Mag. 2016, 33, 81–94.
98. Koza, J.R.; Bennett, F.H.; Andre, D.; Keane, M.A. Automated design of both the topology and sizing of
analog electrical circuits using genetic programming. In Artificial Intelligence in Design’96; Springer: Berlin,
Germany, 1996, pp. 151–170.
99. Russell, S.; P. Norvig, P.; Canny, C.F.; Malik, J.M. A. Artificial Intelligence: A Modern Approach; Upper Saddle
River: Bergen, NJ, USA, 1995.
100. Du, K.-L.; Swamy, M.N.S. Fundamentals of Machine Learning. In Neural Networks and Statistical Learning.
Springer: Berlin, Germany, 2014; pp. 15–65.
100
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
21 of 22
101. Zhang, Y.; Rajapakse, J.C. Machine Learning in Bioinformatics. John Wiley & Sons: Hoboken, NJ, USA, 2009.
102. Witten, I.H.; Frank, E.; Hall, M.A.; Data Mining: Practical Machine Learning Tools and Techniques; Morgan
Kaufmann: Burlington, MA, USA, 2016.
103. Schapire, R.E. The boosting approach to machine learning: An overview. In Nonlinear Estimation and
Classification, Springer: Berlin, Germany, 2003; pp. 149–171.
104. Michalski, R.S.; Carbonell, J.G.; Mitchell, T.M.X. Machine Learning: An Artificial Intelligence Approach.
Springer Science & Business Media: Berlin, Germany, 2013
105. Bishop, C.M. Pattern Recognition and Machine Learning. Springer: Berlin, Germany, 2006.
106. Lorenzi, P.; Rao, R.; Romano, G.; Kita, A.; Irrera, F.; Mobile Devices for the Real-Time Detection of Specific
Human Motion Disorders. IEEE Sens. J. 2016, 16, 8220–8227.
107. Lau, S.L.; König, I.; David, K.; Parandian, B.; Carius-Düssel, C.; Schultz, M. Supporting patient monitoring
using activity recognition with a smartphone. In Proceedings of 2010 7th International Symposium on
Wireless Communication Systems (ISWCS), York, UK, 19–22 September 2010.
108. Lau, S.L. Comparison of orientation-independent-based-independent-based movement recognition system
using classification algorithms. In Proceedings of 2013 IEEE Symposium on Wireless Technology and
Applications (ISWTA), Kuching, Malaysia, 22–25 September 2013.
109. Duarte, F.; Lourenco, A.; Abrantes, A. Activity classification using a smartphone. In Proceedings of 2013
IEEE 15th International Conference on e-Health Networking, Applications & Services (Healthcom), Lisbon,
Portugal, 9–12 October 2013
110. Fahim, M.; Lee, S.; Yoon, Y. SUPAR: Smartphone as a ubiquitous physical activity recognizer for uhealthcare services. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2014, 2014, 3666–3669.
111. Bajpai, A.; Jilla, V.; Tiwari, V.N.; Venkatesan, S.M.; Narayanan, R. Quantifiable fitness tracking using
wearable devices. In Proceedings of the 37th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), Milan, Italy, 25–29 August 2015.
112. Nguyen, P.; Akiyama, T.; Ohashi, H.; Nakahara, G.; Yamasaki, K.; Hikaru, S. User-friendly Activity
Recognition Using SVM Classifier and Informative Features. In Proceedings of the 2015 International
Conference on Indoor Positioning and Indoor Navigation (Ipin), Banff, AB, Canada, 13–16 October 2015;
pp. 1–8.
113. Wang, C.; Xu, Y.; Zhang, J.; Yu, W. SW-HMM: A Method for Evaluating Confidence of Smartphone-Based
Activity Recognition. In Proceedings of the 2016 IEEE Trustcom/BigDataSE/ISPA, Tianjin, China, 23–26
August 2016.
114. Lau, S.L.; David, K. Movement recognition using the accelerometer in smartphones. In Proceedings of the
Future Network and Mobile Summit, Florence, Italy, 16–18 June 2010.
115. Zhang, L.; Wu, X.; Luo, D. Real-Time Activity Recognition on Smartphones Using Deep Neural Networks.
In Proceedings of the Ubiquitous Intelligence and Computing and 2015 IEEE 12th Intl Conf on Autonomic
and Trusted Computing and 2015 IEEE 15th Intl Conf on Scalable Computing and Communications and
Its Associated Workshops (UIC-ATC-ScalCom), Beijing, China, 10–14 August 2015.
116. Cardoso, N.; Madureira, J.; Pereira, N. Smartphone-based Transport Mode Detection for Elderly Care. 2016
In Proceedings of the IEEE 18th International Conference on E-Health Networking, Applications and
Services (Healthcom), Munich, Germany, 14–16 September 2016; pp. 261–266.
117. Vallabh, P.; Malekian, R.; Ye, N.; Bogatinoska, D.C. Fall Detection Using Machine Learning Algorithms. In
Proceedings of the 24th International Conference on Software, Telecommunications and Computer
Networks (SoftCOM), Split, Croatia, 22–24 September 2016; pp. 51–59.
118. Filios, G.; Nikoletseas, S.; Pavlopoulou, C.; Rapti, M.; Ziegler, S. Hierarchical Algorithm for Daily Activity
Recognition via Smartphone Sensors. In Proceedings of the 2015 IEEE 2nd World Forum on Internet of
Things (WF-IOT), Milan, Italy, 14–16 December 2015; pp. 381–386.
119. Tang, C.X.; Phoha, V.V. An Empirical Evaluation of Activities and Classifiers for User Identification on
Smartphones. In Proceedings of the 2016 IEEE 8th International Conference on Biometrics Theory,
Applications and Systems (BTAS), Niagara Falls, NY, USA, 6–9 September 2016; pp. 1–8.
120. Li, P.; Wang, Y.; Tian, Y.; Zhou, T.S.; Li, J.S. An Automatic User-Adapted Physical Activity Classification
Method Using Smartphones. IEEE Trans. Biomed. Eng. 2017, 64, 706–714.
121. Kim, Y.J.; Kang, B.N.; Kim, D. Hidden Markov Model Ensemble for Activity Recognition using Tri-axis
Accelerometer. In Proceedings of the 2015 IEEE International Conference on Systems, Man, and
101
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Sensors 2018, 18, 640
22 of 22
Cybernetics (Smc 2015): Big Data Analytics for Human-Centric Systems, Kowloon, China, 9–12 October
2015; pp. 3036–3041.
122. Brdiczka, O.; Bellotti, V. Identifying routine and telltale activity patterns in knowledge work. In
Proceedings of the Fifth IEEE International Conference on Semantic Computing (ICSC), Palo Alto, CA,
USA, 18–21 September 2011.
123. Costa, Â.; Castillo, J.C.; Novais, P.; Fernández-Caballero, A.; Simoes, R. Sensor-driven agenda for intelligent
home care of the elderly. Exp. Syst. Appl. 2012, 39, 12192–12204.
© 2018 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article
distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license
(http://creativecommons.org/licenses/by/4.0/).
102
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 4
Data fusion for the Identification of Activities of
Daily Living
This chapter is related to the research and development of the framework for the recognition
of ADL and environments, and it is composed by two articles, each presented in its section.
1. Identification of Activities of Daily Living through
Data Fusion on Motion and Magnetic Sensors
embedded on Mobile Devices
The following article is the first part of the chapter 4.
Identification of Activities of Daily Living through Data Fusion on Motion and Magnetic Sensors
embedded on Mobile Devices
Ivan Miguel Pires, Nuno M. Garcia, Nuno Pombo, Francisco Flórez-Revuelta, Susanna Spinsante
and Maria Canavarro Teixeira
Pervasive and Mobile Computing (Elsevier), published, 2018.
According to 2016 Journal Citation Reports published by Thomson Reuters in 2017, this journal
has the following performance metrics:
ISI Impact Factor (2016): 2.349
ISI Article Influence Score (2016): 0.7
Journal Ranking (2016): 6/32 (Computer Science (miscellaneous))
103
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
104
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Pervasive and Mobile Computing 47 (2018) 78–93
Contents lists available at ScienceDirect
Pervasive and Mobile Computing
journal homepage: www.elsevier.com/locate/pmc
Identification of activities of daily living through data fusion
on motion and magnetic sensors embedded on mobile devices
Ivan Miguel Pires a,b,c , , Nuno M. Garcia a , Nuno Pombo a ,
Francisco Flórez-Revuelta d , Susanna Spinsante e , Maria Canavarro Teixeira f,g
∗
a
Instituto de Telecomunicações, Universidade da Beira Interior, Covilhã, Portugal
Altranportugal, Lisbon, Portugal
ALLab — Assisted Living Computing and Telecommunications Laboratory, Computer Science Department, Universidade da Beira Interior,
Covilhã, Portugal
d
Department of Computer Technology, Universidad de Alicante, Spain
e
Università Politecnica delle Marche, Ancona, Italy
f
UTC de Recursos Naturais e Desenvolvimento Sustentável, Polytechnique Institute of Castelo Branco, Castelo Branco, Portugal
g
CERNAS — Research Centre for Natural Resources, Environment and Society, Polytechnique Institute of Castelo Branco, Castelo Branco,
Portugal
b
c
article
info
Article history:
Received 23 January 2018
Received in revised form 6 May 2018
Accepted 18 May 2018
Available online xxxx
Keywords:
Mobile devices sensors
Sensor data fusion
Artificial neural networks
Identification of activities of daily living
a b s t r a c t
Several types of sensors have been available in off-the-shelf mobile devices, including
motion, magnetic, vision, acoustic, and location sensors. This paper focuses on the fusion
of the data acquired from motion and magnetic sensors, i.e., accelerometer, gyroscope
and magnetometer sensors, for the recognition of Activities of Daily Living (ADL). Based
on pattern recognition techniques, the system developed in this study includes data
acquisition, data processing, data fusion, and classification methods like Artificial Neural
Networks (ANN). Multiple settings of the ANN were implemented and evaluated in which
the best accuracy obtained, with Deep Neural Networks (DNN), was 89.51%. This novel
approach applies L2 regularization and normalization techniques on the sensors’ data
proved it suitability and reliability for the ADL recognition.
© 2018 Elsevier B.V. All rights reserved.
1. Introduction
Off-the-shelf mobile devices have several sensors available, which are capable for the acquisition of several physical and
physiological parameters [1], including the accelerometer, magnetometer, and gyroscope sensors, allowing the recognition
of Activities of Daily Living (ADL) [2]. The correct identification of ADL is one of the stages for the development of a personal
digital life coach [3], which can be used in several areas, including sports and geriatrics, among others.
For the use of several sensors in the development of a method for the recognition of ADL, data fusion techniques
should be used before the application of the classification methods. This paper focuses on the use of motion and magnetic
sensors available on mobile devices, where the most commonly available are the accelerometer, the magnetometer, and the
gyroscope, proposing the recognition of ADL with movement, including running, walking, walking on stairs, and standing.
The architecture for the method for the recognition of ADL was proposed in [4–6], which is composed by data acquisition,
data processing, data fusion, and classification methods. Taking in account that the data acquired from the sensors is fulfilled,
∗ Correspondence to: Computer Science Department, Universidade da Beira Interior, Rua Marques d’Avila e Bolama, 6200-001 Covilhã, Portugal.
E-mail addresses: impires@it.ubi.pt (I.M. Pires), ngarcia@di.ubi.pt (N.M. Garcia), ngpombo@di.ubi.pt (N. Pombo), francisco.florez@ua.es
(F. Flórez-Revuelta), s.spinsante@univpm.it (S. Spinsante), ccanavarro@ipcb.pt (M.C. Teixeira).
https://doi.org/10.1016/j.pmcj.2018.05.005
1574-1192/© 2018 Elsevier B.V. All rights reserved.
105
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
79
the data processing methods are forked in two types of methods, namely data cleaning, and features extraction methods.
After the feature extraction, the data will be fused and the classification methods will be applied.
Following the use of the data fusion techniques with the accelerometer, the gyroscope and the magnetometer sensors,
this study proposes a creation of a new method that combine multiple sensors suitable for the off-the-shelf mobile devices
such as smartphones.
Recently, several studies have been done regarding the recognition of ADL using several sensors [7–12]. Artificial Neural
Networks (ANN) are widely used, proving their reliability for the automatic learning and recognition of several patterns of
sensors’ data [13,14]. Due to the limitation of number of sensors available in the off-the-shelf mobile devices, and based
on the previous study [15] that uses only the accelerometer sensor, this study proposes the creation of two different
methods for the recognition of ADL using different number of sensors in order to adapt the method according to the number
of sensors available. Firstly, it proposes the fusion of the data acquired from the accelerometer, and the magnetometer
sensors. Secondly, it proposes the fusion of the data acquired from the accelerometer, gyroscope, and magnetometer sensors.
The ADL proposed for recognition are running, walking, going upstairs, going downstairs, and standing, consisting this
research on the analysis of the performance of three implementations of ANN, namely, Multilayer Perception (MLP) with
Backpropagation, Feedforward Neural Network (FNN) with Backpropagation, and Deep Neural Networks (DNN). A dataset
used for this research is composed by the sensors’ data acquired in several experiments with users aged between 16 and
60 years old, having distinct lifestyles, and a mobile device in the front pocket of their trousers, performing the proposed ADL.
This research was conducted with the use of three Java libraries, Neuroph [16], Encog [17], and DeepLearning4j [18], and
different datasets of features, in order to identify the best dataset of features and implementation of ANN for the recognition
of ADL, verifying that the best accuracy for the recognition of ADL with the two different methods proposed was achieved
with Deep Learning methods.
This paragraph concludes Section 1, and this paper is organized as follows: Section 2 summarizes the literature review
for the use of data fusion techniques with accelerometer, gyroscope, and magnetometer sensors; Section 3 presents the
methods used on each stage of the architecture proposed. The results obtained are presented in Section 4, presenting the
discussion about these results in Section 5. The conclusions of this study are presented in Section 6.
2. Related work
Data fusion techniques may be used with the data acquired from motion and magnetic sensors available in the off-theshelf mobile devices, i.e., accelerometer, gyroscope, and magnetometer, in order to improve the reliability of the methods
for the recognition of Activities of Daily Living (ADL) [2].
Following the main focus of this paper, the accelerometer, the gyroscope, and the magnetometer are used by the authors
of [19] with the Random Forest classifier for the recognition of standing, going downstairs, going upstairs, sitting, walking,
and running activities. Based on a myriad of features such as the variance, the mean, the frequency of the point with
maximum amplitude, the energy of the extremum value, the value of the point with maximum amplitude, the mean and the
period of the extremum value, the sum of the difference between extremum values, the maximum, the minimum and the
mean value around the midpoint, reporting an average accuracy of 99.7%.
In addition, Shoaib et al. [20] presented a method that also uses the Global Positioning System (GPS) receiver, implementing ANN, i.e., Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Naïve Bayes, Logistic regression, decision tree,
K-Nearest Neighbor (KNN), and rule based classifiers. The features extracted included the mean and the standard deviation
of the raw signal from the accelerometer, gyroscope, and magnetometer, and the distance from the GPS data, in order to
recognize running, walking, standing, sitting, going downstairs, and going upstairs with a reported accuracy between 69%
and 99%.
For the recognition of going upstairs and downstairs, standing, walking on an escalator and taking an elevator, the
authors of [21] extracted several features from the accelerometer, magnetometer and gyroscope sensors, including mean,
median, variance, standard deviation, 75th percentile, inter-quartile range, average absolute difference, binned distribution,
energy, Signal Magnitude Area (SMA), Zero-Crossing Rate, Number of Peaks, Absolute Value of short-time Fourier Transform,
Power of short-time Fourier Transform and Power Spectral Centroid. They used these features with a decision tree method,
reporting an accuracy between 80% and 90% [21].
The majority of the studies in the literature only fuses the accelerometer and gyroscope data. For example the authors
of [22] do not present the features extracted, but they implemented the Random Forests (RF) variable importance in order
to help in the selection of the best features for the recognition of walking, going upstairs and downstairs, sitting, standing
and laying activities. After the feature extraction, the implementation of the Two-stage continuous Hidden Markov Model
(HMM) reported an accuracy of 91.76%. The Hidden Markov Model (HMM) was also implemented in [23], which also
implemented the decision tree and Random Forest methods, with accelerometer and gyroscope data for the recognition
of going downstairs and upstairs, and walking. The features included: variance, mean, standard deviation, maximum,
minimum, median, interquartile range, skewness, kurtosis, and spectrum peak position of the accelerometer and gyroscope
data, reporting an accuracy of 93.8%.
In [24], the authors recognized walking, standing, running, laying activities, going downstairs and upstairs, and with
accelerometer and gyroscope data, extracting the mean, the energy, the standard deviation, the correlation, and the entropy
106
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
80
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
of the sensors’ data. Several models were implemented such as the J48 decision tree, the logistic regression, the MLP, and
the SVM reporting an accuracy between 89.3% and 100%.
The authors of [25] implemented the Signal Magnitude Vector (SMV) algorithm with a Threshold based algorithm for
feature extraction in order to recognize some ADL, such as walking, standing, sitting, and running with a reported accuracy
around 90%.
According to [26], the Gaussian mixture model (GMM) and the time series shapelets, applied to the accelerometer and
gyroscope data, allow the recognition of sitting, standing, walking, and running activities with mean and standard deviation
as features, reporting an accuracy of 88.64%. The authors of [27] also used the mean and standard deviation as features for the
application of KNN and SVM methods, in order to recognize walking, resting, running, going downstairs, and going upstairs
with a reported accuracy higher than 90%.
The standard deviation, maximum, minimum, correlation coefficients, interquartile range, mean, Dynamic time warping
distance (DTW), Fast Fourier Transform (FFT) coefficients, and wavelet energy were extracted as features from accelerometer
and gyroscope sensors. In order to recognize walking, jumping, running, going downstairs and upstairs several methods
were implemented, such as SVM, KNN, MLP, and Random Forest, reporting an accuracy between 84.97% and 90.65% [12].
The authors of [28] extracted the same features for the recognition of walking, going upstairs and downstairs, jumping, and
jogging activities, The KNN, the Random Forests and the SVM methods were implemented, reporting an accuracy of 95%.
The authors of [29] extracted the variance, mean, minimum and maximum along the Y axis of the accelerometer, and
the variance and mean along the X axis of the gyroscope, and implemented the SVM method for the recognition of running,
walking, going downstairs and upstairs, standing, cycling and sitting, which reports an accuracy of 96%.
In [30], the authors extracted the skewness, mean, minimum, maximum, standard deviation, kurtosis, median, and
interquartile range from the accelerometer and gyroscope data, implementing the MLP, the SVM, the Least Squares Method
(LSM), and the Naïve Bayes classifiers for the recognition of falling activities with a reported accuracy of 87.5%.
The SVM, Random Forest, J48 decision tree, Naïve Bayes, MLP, Rpart, JRip, Bagging, and KNN were implemented in [31] for
the recognition of going downstairs and upstairs, lying, standing, and walking with the mean and standard deviation along
the X , Y and Z axis of the accelerometer and the gyroscope signal as features, reporting an accuracy higher than 90%.
The Root Mean Square (RMS), minimum, maximum, and zero crossing rate for X , Y , and Z axis were extracted from the
accelerometer and gyroscope data, and the ANOVA method was applied for the correct recognition of sitting, resting, turning,
and walking with a maximum reported accuracy of 100% [32].
The driving, walking, running, cycling, resting, and jogging were recognized by ANN with mean, minimum, maximum,
standard deviation, difference between maximum and minimum, Parseval’s Energy either in the frequency range 0–2.5 Hz,
or in the frequencies greater than 2.5 Hz. In addition, RMS, kurtosis, correlation between axis, ratio of the maximum and
minimum values in the FFT, skewness, difference between the maximum and minimum values in the FFT, median of troughs,
median of peaks, number of troughs, number of peaks, average distance between two consecutive troughs, average distance
between two consecutive peaks, indices of the 8 highest peaks after the application of the FFT, and ratio of the average values
of peaks were also considered. The observed accuracy varies between 57.53% and 97.58% [33].
The Threshold Based Algorithm (TBA) was applied to the values of the acceleration, and the difference between adjacent
elements of the heading, extracted from the accelerometer and gyroscope sensors, in order to recognize going downstairs
and upstairs, running, walking, and jumping with a reported accuracy of 83% [34].
The median absolute deviation, minimum, maximum, absolute mean, interquartile range, Signal Magnitude Range,
skewness, and kurtosis were extracted from accelerometer and gyroscope signal for the application of KNN, SVM, Sparse
Representation Classifier, and Kernel-Extreme Learning Machine, in order to recognize standing, running, going upstairs,
walking, and going downstairs, reporting an average accuracy of 94.5% [35].
The jogging and walking activities are recognized with mean, variance, minimum, and maximum of the X , Y and Z axis
of the accelerometer and gyroscope sensors as features applied to the SVM method, reporting an accuracy of 95.5% [36].
The authors of [37] implemented sparse approximation, KNN, SVM, Spearman correlation, Fuzzy c-means, MLP, and linear
regression classifiers for the recognition of running, cycling, sitting, walking, and standing, using the standard deviation,
mean, median, power ratio of the frequency bands, peak acceleration, and energy extracted from the accelerometer and
gyroscope signal, reporting an accuracy of 98%.
In [38], the implementation of SVM and Random Forest methods was used for the recognition of standing, sitting, laying,
walking, going downstairs and upstairs, with the extraction of the angle, the minimum, the maximum, and the mean values
of the accelerometer and gyroscope signal, reporting an accuracy around 100%.
The authors of [39] used the accelerometer and gyroscope sensors for the recognition of the movements related to up and
down buses, implementing the C4.5 decision tree, Naïve Bayes, KNN, logistic regression, SVM, and MLP with mean, standard
deviation, energy, correlation between axis, and magnitude of FFT components as features, reporting an accuracy of 95.3%.
The accelerometer, gyroscope, barometer, and GPS were used for the recognition of standing, sitting, washing dishes,
going downstairs and upstairs, walking, running, and cycling with standard deviation, mean, interquartile range, mean
squared, altitude difference in meters, and speed as features applied to the SVM method, whose the authors reported an
accuracy around 90% [40].
For the recognition of walking, lying, running, cycling, jogging, washing dishes, vacuum cleaning, playing piano, playing
cello, playing tennis, brushing teeth, wiping cupboard, driving, taking an elevator, doing laundry, working on a computer,
eating, reading a book, going downstairs, going upstairs, and folding laundry, the authors of [41] used the features extracted
107
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
81
Table 1
Distribution of the ADL extracted in the studies analyzed.
from the accelerometer, the gyroscope, and the camera, including the variance and mean for each axis, the movement
intensity, the energy, the energy consumption, and the periodicity, applying them to the HMM, the SVM, the Naïve Bayes
methods, and obtaining a reported accuracy of 81.5%.
The J48 decision tree, IBk, MLP, and Logistic regression methods were implemented with the median, the mean, the
standard deviation, the kurtosis, the skewness, the maximum, the minimum, the slope, difference between maximum and
minimum, the spectral centroid, the entropy of the energy in 10 equal sized blocks, the short time energy, the spectral roll off,
the zero crossing rate, the spectral flux, and the spectral centroid for each axis and the absolute value of the accelerometer,
gyroscope, and orientation sensors [42], in order to recognize walking, standing, jogging, going downstairs, going upstairs,
jumping, and sitting activities, reporting an accuracy of 94%.
According to the analysis previously presented, Table 1 shows the ADL recognized, in more than one study, with the use of
the accelerometer, gyroscope and/or magnetometer sensors, verifying that the walking, standing/resting, going downstairs
and upstairs, running, and sitting are the most recognized ADL. The lines in Table 1 are sorted in decreasing manner regarding
the number of studies found for each activity highlighting the activities reported in at least 10 papers.
Based on the literature review, the features used for the recognition of the ADL, in more than 1 studies, are presented
in Table 2, showing that the mean, standard deviation, maximum, minimum, energy, inter-quartile range, correlation
coefficients, median, and variance are the most used features, with more relevance for mean, standard deviation, maximum,
and minimum. The Table 2 is sorted in decreasing order of the number of studies that reportedly used a specific feature
highlighting that ones used in 6 or more papers.
Finally, the methods implemented in the literature that report the achieved accuracy, are presented in Table 3, concluding
that the methods with an accuracy higher than 90% are MLP, logistic regression, random forest and decision tree methods,
verifying that the method that reports the best average accuracy in the recognition of ADL is the MLP, with an average
accuracy equal to 93.86%.
As the ANN are the methods that report the best accuracies in the literature for the recognition of ADL, this paper focuses
on the research on what implementation of ANN reports the best results in the recognition of a set of 5 ADL (i.e., walking,
going downstairs, going upstairs, standing and running) with a dataset previously acquired. Moreover, the comparison of
our results with the results available in the literature is not possible, because the authors did not make their datasets publicly
available and the efficiency of the methods depends on the number of ADL each tries to recognize. The following sections of
this paper focus on the comparison between three implementations of ANN, such as the MLP method with backpropagation,
the FNN method with backpropagation and the DNN method. Their implementation parameters will be presented in the
next section.
3. Methods
Based on the related work presented in the previous section and the proposed architecture of a framework for the
recognition of ADL previously presented in [4–6], there are several modules for the creation of the final method, such as
data acquisition, data processing, data fusion, and classification methods. Assuming that the data acquired from all sensors
is fulfilled, the data processing module is composed by data cleaning and feature extraction methods. After that, the data
fusion method will be applied for further classification of the data. In Fig. 1, a simplified schema for the development of a
framework for the identification of ADL is presented.
Section 3.1 presents the methodology for the data acquisition. Data processing methods are presented in Section 3.2 and,
finally, in Section 3.3, the data fusion and classification methods are presented.
108
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
82
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
Table 2
Distribution of the features extracted in the studies analyzed.
Table 3
Distribution of the classification methods used in the studies analyzed.
3.1. Data acquisition
A mobile application developed for Android devices [43,44] was installed in the BQ Aquaris 5.7 smartphone, which it
has a Quad Core CPU and 16 GB of internal memory [45] for the acquisition of the sensors’ data, saving the data captured
from the accelerometer, magnetometer, and gyroscope sensors into text files. The mobile application captures the data in
5 s slots every 5 min, where the frequency of the data acquisition is around 10 ms per sample. For the definition of the
experiments, 25 individuals aged between 16 and 60 years old were selected. The individuals selected had distinct lifestyles,
being 10 individuals mainly active and the remaining 15 individuals, mainly sedentary. During the data acquisition, the
mobile device was in the front pocket of the user’s trousers, and using the data collection application, the user manually
defined a label of the ADL performed in each of the 5 s of data captured. Based on the ADL that are the most identified in
the previous research studies [4–6,15], the selected ADL for this study in the mobile application are: running, walking, going
upstairs, going downstairs, and standing.
After the selection of the individuals for the experimental study and the definition of the main rules previously discussed,
the environment of the data acquisition needs to be characterized. The environment is strictly defined for the different types
of ADL performed, where the walking and running activities are performed in outdoor environments (e.g. street) at different
109
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
83
times of the day, the standing activity is performed in the living room at different times of the day, and the going upstairs
and going downstairs activities are performed in indoor environments at different times of the day. In total, there are around
36 h of captures for each activity, resulting in 2000 captures with 5 s of raw sensors’ data.
By default, without user selection, the mobile application will label the acquired data as unknown activity, therefore,
before starting the data acquisition, the user must select in the mobile application the ADL that will be performed during
a certain time for completion of the classification method. After this selection the user must put the mobile device is front
pocket of the trousers. The unlabeled data is not stored in the final dataset for the creation of the classification method.
Nevertheless, all the data used in this research is publicly available at the ALLab MediaWiki [46].
One of the problems related to the data acquisition is the battery lifetime and power processing capabilities of the mobile
device, already described in [47], verifying that the acquisition of 5 s of raw data every 5 min using the oldest tested devices
has possible for a minimum of 16 h with a normal use of the mobile device. As the mobile devices usually require a daily
recharge, it is possible to consider that the acquisition of the sensors’ data with this method can be implemented in a real-life
scenario. The performance of the mobile device is also not significantly affected during the data acquisition. In conclusion,
the acquisition of small datasets sampled in a defined time interval allows the use of this method, but it only characterizes the
data acquired on each interval, discarding the activities performed during the time where the data acquisition is in standby
(the 5 min every 5 s). It will be probably sufficient for the characterization of lifestyles, but it may miss some important
events that may not be identified, including other activities e.g. users’ falls, that are out of the scope of this research. The
development of a method that implements a more significant sampling strategy without decreasing the performance and
the availability of resources at the mobile device requires additional research.
3.2. Data processing
The data processing is the second step of the method for the recognition of ADL, which is executed after the data
acquisition. Data cleaning methods are executed for noise reduction, as presented in Section 3.2.1. After cleaning the data,
the features were extracted for further analysis, as discussed in Section 3.2.2.
3.2.1. Data cleaning
Data cleaning is a process to filter the data acquired from the accelerometer, magnetometer, and gyroscope sensors, in
order to remove the noise. The data cleaning method should be selected according to the types of sensors used, nevertheless,
the low-pass filter is the best method for the data acquired from the sensors used in this study [48], removing the noise and
allowing the correct extraction of the selected features.
3.2.2. Feature extraction
After the data cleaning, and based on the features most commonly extracted in previous research studies [4–6,15] (see
Table 2), several features were extracted from the accelerometer, magnetometer, and gyroscope sensors. These are the five
greatest distances between the maximum peaks, the average, standard deviation, variance and median of the maximum
peaks, and the standard deviation, average, maximum value, minimum value, variance and median of the raw signal.
The feature extraction process starts with the calculation of the maximum peaks of the 5 s of data acquired related to
each sensor used. After the calculation of the maximum peaks, the distance between the different peaks (in milliseconds)
is calculated, using the five highest distances between the highest peaks. After that, the remaining features related to
the maximum peaks are calculated, including the average, standard deviation, variance and median. Finally, the standard
deviation, average, maximum value, minimum value, variance and median are calculated from the original acquired data.
These features will be used in the different datasets presented in Section 3.3.
3.3. Identification of activities of daily living with data fusion
Extending a previous study [15] that used only the accelerometer sensor, this study fuses the features extracted from
the accelerometer and magnetometer sensors (Section 3.3.1), and the features extracted from the accelerometer, gyroscope
and magnetometer sensors (Section 3.3.2). Finally, the classification methods for the identification of ADL are presented
in Section 3.3.3.
3.3.1. Data fusion with accelerometer and magnetometer sensors
Regarding the features extracted from each ADL, five datasets have been constructed with features extracted from the
data acquired from the accelerometer and magnetometer during the performance of the five ADL, resulting in 2000 records
from each ADL. The datasets defined are:
• Dataset 5: Composed by the average, and the standard deviation of the raw signal extracted from both accelerometer
and magnetometer sensors;
• Dataset 4: Composed by the features of the Dataset 5 plus the variance and the median of the raw signal extracted from
both accelerometer and magnetometer sensors;
110
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
84
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
Fig. 1. Simplified diagram for the framework for the identification of ADL.
• Dataset 3: Composed by the features of the Dataset 4 plus the maximum and the minimum values of the raw signal
extracted from both accelerometer and magnetometer sensors;
• Dataset 2: Composed by the features of the Dataset 3 plus the average, the standard deviation, the variance and the
median of the maximum peaks obtained from the raw signal extracted from both accelerometer and magnetometer
sensors;
• Dataset 1: Composed by the features of the Dataset 2 plus the five greatest distances between the maximum peaks
obtained from the raw signal extracted from both accelerometer and magnetometer sensor.
3.3.2. Data fusion with accelerometer, magnetometer and gyroscope sensors
Regarding the features extracted from each ADL, five datasets have been similarly constructed with features extracted
from the data captured from the accelerometer, magnetometer and gyroscope, again resulting in 2000 records from each
ADL. The datasets defined are:
• Dataset 5: Composed by the average, and the standard deviation of the raw signal extracted from both accelerometer,
magnetometer and gyroscope sensors;
• Dataset 4: Composed by the features of the Dataset 5 plus the variance and the median of the raw signal extracted from
both accelerometer, magnetometer and gyroscope sensors;
• Dataset 3: Composed by the features of the Dataset 4 plus the maximum and the minimum values of the raw signal
extracted from both accelerometer, magnetometer and gyroscope sensors;
• Dataset 2: Composed by the features of the Dataset 3 plus the average, the standard deviation, the variance and the
median of the maximum peaks obtained from the raw signal extracted from both accelerometer, magnetometer and
gyroscope sensors;
• Dataset 1: Composed by the features of the Dataset 2 plus the five greatest distances between the maximum peaks
obtained from the raw signal extracted from both accelerometer, magnetometer and gyroscope sensors.
3.3.3. Classification
Based on the results reported by the literature review presented in Section 2, one of the most used methods for the
recognition of ADL based on the use of the mobiles’ sensors is the ANN, and this method reports a better accuracy than SVM,
KNN, Random Forest, and Naïve Bayes.
Following the datasets defined in Sections 3.3.1 and 3.3.2, this study implements three implementations of ANN, such as
MLP, FNN, and DNN, in order to identify the best ANN for the recognition of ADL, these are:
• MLP method with Backpropagation, applied with the Neuroph framework [16];
• FNN method with Backpropagation, applied with the Encog framework [17];
• DNN method, applied with the DeepLearning4j framework [18].
111
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
85
Table 4
Configurations of the ANN methods implemented.
Parameters
MLP
FNN
DNN
Activation function
Learning rate
Momentum
Maximum number of training iterations
Number of hidden layers
Weight function
Seed value
Backpropagation
Regularization
Sigmoid
0.6
0.4
4 × 106
0
N/A
N/A
Yes
N/A
Sigmoid
0.6
0.4
4 × 106
0
N/A
N/A
Yes
N/A
Sigmoid
0.1
N/A
4 × 106
3
Xavier
6
Yes
L2
The configurations between the three implementations of ANN methods used are presented in Table 4, where it is visible
that the Sigmoid as activation function and backpropagation parameters are implemented in all methods. The used of more
efficient activation functions, resulting in a better energy efficiency of the proposed method, will be subject to additional
research.
Based on the parameters of the three implementations of ANN presented in Table 4, the architecture of the MLP and the
FNN methods is presented in Fig. 2a and the architecture of the DNN method is presented in Fig. 2b.
In order to improve the results obtained by the ANN, the MIN/MAX normalizer [49] was applied to the defined datasets,
implementing the MLP method with Backpropagation, and the FNN method with Backpropagation with normalized and
non-normalized at different stages.
Before the application of the DNN method, the L2 regularization [50] was applied to the defined datasets. After the
application of the L2 regularization, the normalization with mean and standard deviation [51] was applied to the datasets,
implementing the DNN method with normalized and non-normalized data at different stages.
The number of training iterations may influence the results of the ANN, defining the maximum number of 106 , 2 × 106
and 4 × 106 iterations, in order to identify the best number of training iterations with best results.
After this research, the methods that should be implemented in the framework for the recognition of ADL defined in [4–6]
are a function of the number of sensors available in the off-the-shelf mobile device. According to the results available in [15],
if the mobile device has only the accelerometer sensor, the method that should be implemented is the DNN method, verifying
with this research the best methods for the use of the datasets defined in Sections 3.3.1 and 3.3.2.
4. Results
This research consists in the creation of two different methods for the recognition of ADL with different number of sensors.
Firstly, the results of the creation of a method with accelerometer and magnetometer sensors are presented in Section 4.1.
Finally, the results of the creation of a method with accelerometer, magnetometer, and gyroscope sensors are presented
in Section 4.2.
4.1. Identification of activities of daily living with accelerometer and magnetometer sensors
Based on the datasets defined in Section 3.3.1, the three implementations of ANN proposed in Section 3.3.3 are
implemented with the frameworks proposed, these are the MLP method with Backpropagation, the FNN method with
Backpropagation, and the DNN method. The defined training dataset has 10 000 records, where each ADL has 2000
records.
Firstly, the results of the implementation of the MLP method with Backpropagation using the Neuroph framework are
presented in Fig. 3, verifying that the results have very low accuracy with all datasets, achieving values between 20% and
40% with non-normalized data (Fig. 3a), and values between 20% and 30% with normalized data (Fig. 3b).
Secondly, the results of the implementation of the FNN method with Backpropagation using the Encog framework are
presented in Fig. 4. In general, this implementation of ANN achieves low accuracy results with both non-normalized and
normalized data, reporting the maximum results around 40%. With non-normalized data (Fig. 4a), the ANN reports results
above 30% with the dataset 1 trained over 106 and 4 × 106 iterations, the dataset 2 trained over 106 iterations, the dataset
3 trained over 2 × 106 iterations, the dataset 4 trained over 106 , 2 × 106 and 4 × 106 iterations, and the dataset 5 trained
over 106 and 4 × 106 iterations. With normalized data (Fig. 4b), the results reported are lower than 40%, with an exception
for the ANN trained over 2 × 106 iterations with the dataset 5 that reports results higher than 60%.
Finally, the results of the implementation of the DNN method with the DeepLearning4j framework are presented in Fig. 5.
With non-normalized data (Fig. 5a), the results obtained are below the expectations (around 20%) for the datasets 2, 3 and
4, and the results obtained with dataset 5 are around 70%. On the other hand, with normalized data (Fig. 5b), the results
reported are always higher than 70%, achieving better results with the dataset 1, decreasing with the reduction of the number
of features in the dataset.
112
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
86
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
Fig. 2. Schema of the architecture of the different implementations of ANN used in this study. The Figure (a) shows the architecture related to the MLP
and FNN methods. The Figure (b) shows the architecture of the DNN method.
In Table 5, the maximum accuracies achieved with the different implementations of ANN are presented with the relation
of the different datasets used for accelerometer and magnetometer data, and the computational complexity [52] and the
maximum number of iterations, verifying that the use of the DNN method with normalized data reports better results than
others. The computational complexity taken in account in this study consists on the absolute value of the lower bounds of
time complexity calculated with Big-Oh [52].
Regarding the results obtained, in the case of the use of accelerometer and magnetometer sensors in the framework for
the identification of ADL, the implementation of ANN that should be used is the DNN method with normalized data, because
the results obtained are always higher than 80%.
4.2. Identification of activities of daily living with accelerometer, magnetometer and gyroscope sensors
Based on the datasets defined in Section 3.3.2, the three implementations of ANN proposed in Section 3.3.3 are
implemented with the frameworks proposed, these are the MLP method with Backpropagation, the FNN method with
Backpropagation, and the DNN method. The defined training dataset has 10 000 records, where each ADL has 2000 records.
Firstly, the results of the implementation of the MLP method with Backpropagation using the Neuroph framework are
presented in Fig. 6, verifying that the results have very low accuracy with all datasets. With non-normalized data (Fig. 6a), the
113
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
87
Fig. 3. Results obtained with the Neuroph framework for the different datasets of accelerometer and magnetometer sensors (horizontal axis) and
different maximum number of iterations (series), obtaining the accuracy in percentage (vertical axis). The Figure (a) shows the results with data without
normalization. The Figure (b) shows the results with normalized data.
Fig. 4. Results obtained with the Encog framework for the different datasets of accelerometer and magnetometer sensors (horizontal axis) and different
maximum number of iterations (series), obtaining the accuracy in percentage (vertical axis). The Figure (a) shows the results with data without
normalization. The Figure (b) shows the results with normalized data.
Fig. 5. Results obtained with DeepLearning4j framework for the different datasets of accelerometer and magnetometer sensors (horizontal axis) and
different maximum number of iterations (series), obtaining the accuracy in percentage (vertical axis). The Figure (a) shows the results with data without
normalization. The Figure (b) shows the results with normalized data.
results achieved are between 20% and 40%, where the better accuracy was achieved with the dataset 2. And, with normalized
data (Fig. 6b), the results obtained are between 30% and 40%, with lower results for the dataset 5.
Secondly, the results of the implementation of the FNN method with Backpropagation using the Encog framework
are presented in Fig. 7. In general, this implementation of ANN achieves low accuracy results with non-normalized and
normalized data, reporting the maximum results around 40%. With non-normalized data (Fig. 7a), the ANN reports results
above 30% with the dataset 2 trained over 2 × 106 iterations, the dataset 3 trained over 4 × 106 iterations, and the dataset 4
trained over 4 × 106 iterations, reporting an accuracy higher than 70% with the dataset 2 trained over 2 × 106 iterations. With
normalized data (Fig. 7b), the results reported are lower than 20%, with an exception for the ANN with the dataset 3 trained
over 4 × 106 iterations, the dataset 4 trained over 2 × 106 iterations, and the dataset 5 trained over 106 and 2 × 106 iterations.
Finally, the results of the implementation of the DNN method with the DeepLearning4j framework are presented in Fig. 8.
With non-normalized data (Fig. 8a), the results obtained are below the expectations (around 40%) for the datasets 2, 3 and
4, and the results obtained with datasets 1 and 5 are around 70%. On the other hand, with normalized data (Fig. 8b), the
results reported are always higher than 80%, achieving better results with the dataset 1, decreasing with the reduction of
the number of features in the dataset.
114
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
88
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
Fig. 6. Results obtained with the Neuroph framework for the different datasets of accelerometer, magnetometer and gyroscope sensors (horizontal axis)
and different maximum number of iterations (series), obtaining the accuracy in percentage (vertical axis). The Figure (a) shows the results with data without
normalization. The Figure (b) shows the results with normalized data.
Fig. 7. Results obtained with the Encog framework for the different datasets of accelerometer, magnetometer and gyroscope sensors (horizontal axis) and
different maximum number of iterations (series), obtaining the accuracy in percentage (vertical axis). The Figure (a) shows the results with data without
normalization. The Figure (b) shows the results with normalized data.
Fig. 8. Results obtained with the DeepLearning4j framework for the different datasets of accelerometer, magnetometer and gyroscope sensors (horizontal
axis) and different maximum number of iterations (series), obtaining the accuracy in percentage (vertical axis). The Figure (a) shows the results with data
without normalization. The Figure (b) shows the results with normalized data.
In Table 6, the maximum accuracies achieved with the different implementations of ANN are presented with the
relation of the different datasets used for the accelerometer, magnetometer and gyroscope data, and the computational
complexity [52] and the maximum number of iterations, verifying that the use of the DNN method with normalized data
reports better results than others. The computational complexity taken in account in this study consists on the absolute
value of the lower bounds of time complexity calculated with Big-Oh [52].
Regarding the results obtained, in the case of the use of accelerometer, magnetometer and gyroscope sensors in the
framework for the identification of ADL, the implementation of ANN that should be used is the DNN method with normalized
data, because the results obtained are always higher than 80%, and the best result was achieved with dataset 1 that was equals
to 89.51%.
5. Discussion
Based on the previous studies available in the literature, there are only 3 studies in the literature that focus on the
use of the accelerometer, the gyroscope and the magnetometer sensors, but datasets and detailed specifications of the
115
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
89
Table 5
Best accuracies obtained with the different frameworks, datasets and number of iterations.
Framework
Not normalized data
Normalized data
Absolute value of the lower
bound of the time complexity
Dataset
Iterations needed for training
6
Best accuracy achieved (%)
Neuroph
Encog
Deep learning
9.53674 × 10
0.00390625
4
−27
2
5
5
10
106
4 × 106
35.15
42.75
70.43
Neuroph
Encog
Deep learning
4.85694 × 10−45
0.00390625
4
1
5
1
4 × 106
2 × 106
4 × 106
24.93
64.94
86.49
Iterations needed for training
Best accuracy achieved (%)
Table 6
Best accuracies obtained with the different frameworks, datasets and number of iterations.
Framework
Absolute value of the lower
bound of the time complexity
−45
Datasets
6
Not normalized data
Neuroph
Encog
Deep learning
4.85694 × 10
4.85694 × 10−45
4
2
2
1
10
2 × 106
2 × 106
38.32
76.13
74.47
Normalized data
Neuroph
Encog
Deep learning
4.03122 × 10−75
1.12157 × 10−13
4
1
4
1
2 × 106
2 × 106
4 × 106
37.13
29.54
89.51
Table 7
Best accuracies achieved by the method using only the accelerometer sensor.
Framework
Best accuracy achieved (%)
Not normalized data
Neuroph
Encog
Deep learning
34.76
74.45
80.35
Normalized data
Neuroph
Encog
Deep learning
24.03
37.07
85.89
methods implemented are not available. In line with this, and due to the absence of previous studies focused on the
proposed topology, we compared the use of three Java frameworks with different implementations of ANN, i.e., MLP method
with Backpropagation, FNN method with Backpropagation and DNN method. The main goal is to research on the most
efficient implementation for the recognition of the ADL, in order to may provide both the sequence of the steps used for
the recognition of ADL and the comparison based on the accuracy of the recognition of ADL.
Following the research for the development of a framework for the identification of the ADL using motion and magnetic
sensors, presented in [4–6], there are several modules, such as data acquisition, data cleaning, feature extraction, data fusion,
and classification methods. The choice of the methods for data fusion, and classification modules, depends on the number
of sensors available on the mobile device, aiming to use the maximum number of sensors available on the mobile device, in
order to increase the reliability of the method.
According to the previous study based only in the use of the accelerometer sensor for the recognition of ADL, presented
in [15], the best results achieved for each implementation of ANN are presented in Table 7, verifying that the best method
is the DNN method with normalized data, reporting an accuracy of 85.89%. In the case of the mobile device only has
the accelerometer sensor available, DNN method with normalized data should be implemented in the framework for the
recognition of ADL, removing the data fusion, as presented in Fig. 1.
Based on results obtained with the use of accelerometer and magnetometer sensors, presented in Section 4.1, the
comparison of the results between the use of the accelerometer sensor, and the use of accelerometer and magnetometer
sensors is presented in Table 8. In general, the accuracy increases with the use of normalized data, and decreases with the
use of non-normalized data, where the highest difference was verified with the use of the accelerometer and magnetometer
sensors with the implementation of FNN method with Backpropagation using the Encog framework, reporting a difference
of 27.87%. However, the DNN method continues to report the best results with an accuracy of 86.49%. In the case of the
mobile device only having the accelerometer and magnetometer sensors available, the DNN method with normalized data
should be implemented in the framework for the recognition of ADL.
Based on results obtained with the use of accelerometer, magnetometer and gyroscope sensors, presented in Section 4.2,
the comparison of the results between the use of the accelerometer sensor, and the use of accelerometer, magnetometer
and gyroscope sensors is presented in Table 9. In general, the accuracy increases, except in the cases of the use of the
DNN method with non-normalized data and the FNN method with Backpropagation using the Encog framework with
normalized data. The highest difference in the accuracy is verified with the use of MLP method with Backpropagation
using the Neuroph framework, where the accuracy results increased 13.1% with normalized data, but the DNN method
achieves better results with an accuracy of 89.51%. However, the computational complexity of the DNN method is higher
116
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
90
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
Table 8
Comparison between the best results achieved only using the accelerometer sensor, and using the accelerometer and
magnetometer sensors.
Framework
Best accuracy achieved (%)
Accelerometer
Accelerometer
Magnetometer
Difference (%)
Not normalized data
Neuroph
Encog
Deep learning
34.76
74.45
80.35
35.15
42.75
70.43
+0.39
−31.70
−9.92
Normalized data
Neuroph
Encog
Deep learning
24.03
37.07
85.89
24.93
64.94
86.49
+0.90
+27.87
+0.60
Table 9
Comparison between the best results achieved only using the accelerometer sensor, and using the accelerometer, magnetometer and gyroscope sensors.
Framework
Best accuracy achieved (%)
Accelerometer
Accelerometer
Magnetometer
Gyroscope
Difference (%)
Not normalized data
Neuroph
Encog
Deep learning
34.76
74.45
80.35
38.32
76.13
74.47
+3.56
+1.46
−5.88
Normalized data
Neuroph
Encog
Deep learning
24.03
37.07
85.89
37.13
29.54
89.51
+13.10
−7.53
+3.62
Table 10
Comparison between the best results achieved only using the accelerometer and magnetometer sensors, and using the
accelerometer, magnetometer and gyroscope sensors.
Framework
Best accuracy achieved (%)
Accelerometer
Accelerometer
Magnetometer
Magnetometer
Gyroscope
Difference (%)
Not normalized data
Neuroph
Encog
Deep learning
35.15
42.75
70.43
38.32
76.13
74.47
+3.17
+33.38
+4.04
Normalized data
Neuroph
Encog
Deep learning
24.93
64.94
86.49
37.13
29.54
89.51
+12.20
−35.40
+3.02
than the MLP and FNN methods, but the results obtained are significantly higher than the others in the recognition of the
proposed ADL.
Based on results obtained with the use of accelerometer, magnetometer and gyroscope sensors, presented in Section 4.2,
and the results obtained with the use of accelerometer and magnetometer sensors, presented in Section 4.1, the comparison
between these results is presented in Table 10. In general, the accuracy increases, except in the case of the use of the FNN
method with Backpropagation using the Encog framework with normalized data. The highest difference in the accuracy
is verified with the use of the FNN method with Backpropagation using the Encog framework with non-normalized data,
where the accuracy results increased 33.38% with non-normalized data, but the DNN method continues achieving the better
results with an accuracy of 89.51%. Thus, in the case of the mobile device has the accelerometer, magnetometer and gyroscope
sensors available, the DNN method with normalized data should be implemented in the framework for the recognition of
ADL.
In conclusion, when compared with MLP method with Backpropagation using the Neuroph framework and FNN method
with Backpropagation using the Encog framework, the DNN method with normalized data achieves better results for the
recognition of the ADL with accuracies between 85% and 90% with the different datasets.
The fusion of the data acquired from the sensors available in the mobile device requires the acquisition, processing,
fusion and classification of the different sensors’ data as defined in the schema of the framework proposed (see Fig. 1).
The energy consumptions for the different sensors is very different, being the accelerometer consumption range of one
order of magnitude, the magnetometer in the range of two orders of magnitude, and the gyroscope in the range of three
orders of magnitude, unbalancing and increasing the resources (e.g., battery and power processing capabilities) needed for
the execution of the framework. However, as presented in [47], with the method proposed in this study the effects are
minimized and only a daily recharge is needed for the execution of the framework with the three sensors. The devices used
in [47] were used to study different combinations of sensors both in number and type, including a configuration where the
oldest devices had only the accelerometer sensor, and the remaining devices were used for other configurations. Regarding
117
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
91
the use of only the accelerometer and the use of the fusion of one accelerometer and one magnetometer, the accuracy has a
marginal improvement (+0.60%), but it increases the reliability of the framework and reduces the problems with the possible
failures of one of the sensors during data acquisition. Finally, the use of the three sensors improves the accuracy of the
framework in 3.02%, the main problem consisting in the fact that some commercially available devices do not integrate the
set comprising the gyroscope, magnetometer and accelerometer sensors. In conclusion, data fusion increases the reliability
of the framework, proving the viability of the use of the proposed framework.
6. Conclusions
The sensors that are available in the mobile devices, including accelerometer, gyroscope, and magnetometer sensors,
allow the capture of data that can be used to the recognition of ADL [2]. This study focused on the architecture defined
in [4–6], composed by data acquisition, data cleaning, feature extraction, data fusion and classification methods. Based on
the literature review, the proposed ADL for the recognition with motion and magnetic sensors are running, walking, going
upstairs, going downstairs, and standing.
Based on the data acquired, several features must be extracted from the sensors’ signal, such as the five greatest distances
between the maximum peaks, the average, standard deviation, variance, and median of the maximum peaks, plus the
average, the standard deviation, the variance, the median, and the minimum, the maximum of the raw signal have been
extracted from the accelerometer, magnetometer and/or gyroscope sensors available on the off-the-shelf mobile device.
The fusion of the features implemented in the framework for the recognition of ADL should be a function of the number
of sensors available. The method implemented in the framework for the recognition of ADL should also adapted with the
software and hardware limitations of the mobile devices, including the number of sensors available and the low memory
and power processing capabilities.
For the development of the framework for the recognition of the ADL, three implementations of ANN were created in
order to identify the best framework and implementation of ANN for the development of each step of the framework for the
recognition of ADL, such as the MLP method with Backpropagation using the Neuroph framework [16], the FNN method with
Backpropagation using the Encog framework [17], and the DNN method using DeepLearning4j framework [18], verifying that
the DNN method achieves better results than others.
Due to the limitations of mobile devices and regarding the results obtained with the method for the recognition of ADL
with the accelerometer previously performed, presented in [15], it was verified that the best results were obtained with the
DNN method with L2 regularization and normalized data with an accuracy of 85.89%.
Related to the development of a method for the recognition of ADL, this study proves that the best accuracy are always
achieved with the DNN method with L2 regularization and normalized data, and the data fusion increases the accuracy of
the method, reporting an accuracy of 86.49% with the fusion of the data acquired from two sensors (i.e., accelerometer and
magnetometer), and 89.51% with the fusion of the data acquired from three sensors (i.e., accelerometer, magnetometer and
gyroscope).
On the other hand, the MLP method with Backpropagation and the FNN method with Backpropagation achieve low
accuracies, because the networks are overfitting, and this problem may be solved with several strategies, these being the
stopping of the training when the network error increases for several iterations, the application of dropout regularization,
the application of L2 regularization, the application of the batch normalization, or the reduction of the number of features in
the ANN implemented.
As future work, the methods for the recognition of ADL presented in this study should be implemented during the
development of the framework for the identification of ADL, adapting the method to the number of sensors available on
the mobile device, but the method that should be implemented is the DNN method. The data related to this research is
available in a free repository [46].
Acknowledgments
This work was supported by FCT project UID/EEA/50008/2013 (Este trabalho foi suportado pelo projecto
FCT UID/EEA/50008/2013).
The authors would also like to acknowledge the contribution of the COST Action IC1303 – AAPELE – Architectures,
Algorithms and Protocols for Enhanced Living Environments.
References
[1] L.H.A. Salazar, T. Lacerda, J.V. Nunes, C. Gresse von Wangenheim, A systematic literature review on usability heuristics for mobile phones, Int. J. Mob.
Hum. Comput. Interact. 5 (2013) 50–61. http://dx.doi.org/10.4018/jmhci.2013040103.
[2] D. Foti, J.S. Koketsu, Activities of daily living, in: Pedretti’s Occupational Therapy: Practical Skills for Physical Dysfunction, vol. 7 2013, pp. 157–232.
[3] N.M. Garcia, A roadmap to the design of a personal digital life coach, in: ICT Innovations 2015, Springer, 2016.
[4] I. Pires, N. Garcia, N. Pombo, F. Flórez-Revuelta, From data acquisition to data fusion: A comprehensive review and a roadmap for the identification of
activities of daily living using mobile devices, Sensors 16 (2016) 184.
[5] I.M. Pires, N.M. Garcia, F. Flórez-Revuelta, Multi-sensor data fusion techniques for the identification of activities of daily living using mobile devices,
in: Proceedings of the ECMLPKDD 2015 Doctoral Consortium, European Conference on Machine Learning and Principles and Practice of Knowledge
Discovery in Databases, Porto, Portugal, 2015.
118
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
92
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
[6] I.M. Pires, N.M. Garcia, N. Pombo, F. Flórez-Revuelta, Identification of activities of daily living using sensors available in off-the-shelf mobile devices:
Research and hypothesis, in: Ambient Intelligence-Software and Applications–7th International Symposium on Ambient Intelligence, ISAmI 2016,
2016, pp. 121–130.
[7] O. Banos, M. Damas, H. Pomares, I. Rojas, On the use of sensor fusion to reduce the impact of rotational and additive noise in human activity recognition,
Sensors (Basel) 12 (2012) 8039–8054. http://dx.doi.org/10.3390/s120608039.
[8] M.A.A. Akhoundi, E. Valavi, Multi-sensor fuzzy data fusion using sensors with different characteristics, 2010, arXiv preprint arXiv:1010.6096.
[9] P. Paul, T. George, An effective approach for human activity recognition on smartphone, in: 2015 IEEE International Conference on Engineering and
Technology, ICETech, 2015, pp. 45–47. http://dx.doi.org/10.1109/icetech.2015.7275024.
[10] Y.-W. Hsu, K.-H. Chen, J.-J. Yang, F.-S. Jaw, Smartphone-based fall detection algorithm using feature extraction, in: 2016 9th International Congress on
Image and Signal Processing, BioMedical Engineering and Informatics, CISP-BMEI, Datong, China, 2016, pp. 1535–1540.
[11] S. Dernbach, B. Das, N.C. Krishnan, B.L. Thomas, D.J. Cook, Simple and complex activity recognition through smart phones, in: 2012 8th International
Conference on Intelligent Environments, IE, Guanajuato, Mexico, 2012, pp. 214–221.
[12] C. Shen, Y.F. Chen, G.S. Yang, On motion-sensor behavior analysis for human-activity recognition via smartphones, in: 2016 IEEE International
Conference on Identity, Security and Behavior Analysis, Isba, Sendai, Japan, 2016, pp. 1–6.
[13] D. Wang, Pattern recognition: neural networks in perspective, IEEE Expert 8 (1993) 52–60. http://dx.doi.org/10.1109/64.223991.
[14] K. Doya, D. Wang, Exciting time for neural networks, Neural Netw. 61 (2015) xv–xvi. http://dx.doi.org/10.1016/s0893-6080(14)00260-3.
[15] I.M. Pires, N.M. Garcia, N. Pombo, F. Flórez-Revuelta, S. Spinsante, Pattern recognition techniques for the identification of activities of daily living using
mobile device accelerometer, 2017, submitted for publication, ed. arXiv:1711.00096.
[16] Neuroph, Java Neural Network Framework Neuroph, 2 Sep. 2017. Available: http://neuroph.sourceforge.net/.
[17] H. Research, Encog Machine Learning Framework, 2 Sep. 2017. Available: http://www.heatonresearch.com/encog/.
[18] A. Chris Nicholson, Deeplearning4j: Open-source, Distributed Deep Learning for the JVM, 2 Sep. 2017. Available: https://deeplearning4j.org/.
[19] Q. Guo, B. Liu, C.W. Chen, A two-layer and multi-strategy framework for human activity recognition using smartphone, in: 2016 IEEE International
Conference on Communications, ICC, 2016, pp. 1–6, http://dx.oi.org/d10.1109/icc.2016.7511487.
[20] M. Shoaib, H. Scholten, P.J.M. Havinga, Towards physical activity recognition using smartphone sensors, in: 2013 IEEE 10th International Conference
on and 10th International Conference on Autonomic and Trusted Computing (UIC/ATC) Ubiquitous Intelligence and Computing, 2013, pp. 80–87.
http://dx.doi.org/10.1109/Uic-Atc.2013.43.
[21] M. Elhoushi, J. Georgy, A. Wahdan, M. Korenberg, A. Noureldin, Using portable device sensors to recognize height changing modes of motion, in: 2014
IEEE International Instrumentation and Measurement Technology Conference, I2MTC, Proceedings, 2014, pp. 477–481. http://dx.doi.org/10.1109/
i2mtc.2014.6860791.
[22] C.A. Ronao, S.B. Cho, Human activity recognition using smartphone sensors with two-stage continuous hidden Markov models, in: 2014 10th
International Conference on Natural Computation, ICNC, 2014, pp. 681–686. http://dx.doi.org/10.1109/icnc.2014.6975918.
[23] Y.Y. Liu, F. Zhao, W.H. Shao, H.Y. Luo, An hidden Markov model based complex walking pattern recognition algorithm, in: Proceedings of 2016
Fourth International Conference on Ubiquitous Positioning, Indoor Navigation and Location Based Services, IEEE UPINLBS 2016, 2016, pp. 223–229.
http://dx.doi.org/10.1109/upinlbs.2016.7809976.
[24] W.C. Hung, F. Shen, Y.L. Wu, M.K. Hor, C.Y. Tang, Activity recognition with sensors on mobile devices, in: Proceedings of 2014 International Conference
on Machine Learning and Cybernetics, ICMLC, Vol. 2, 2014, pp. 449–454. http://dx.doi.org/10.1109/icmlc.2014.7009650.
[25] M.B. Rasheed, N. Javaid, T.A. Alghamdi, S. Mukhtar, U. Qasim, Z.A. Khan, et al., Evaluation of human activity recognition and fall detection using android
phone, in: 2015 IEEE 29th International Conference on Advanced Information Networking and Applications, IEEE AINA 2015, 2015, pp. 163–170.
http://dx.doi.org/10.1109/Aina.2015.181.
[26] F. Kawsar, M.K. Hasan, R. Love, S.I. Ahamed, A novel activity detection system using plantar pressure sensors and smartphone, in: 39th Annual IEEE
Computers, Software and Applications Conference, COMPSAC 2015, Vol. 1, 2015, pp. 44–49. http://dx.doi.org/10.1109/Compsac.2015.201.
[27] V. Simonetti, W. Baccinelli, M. Bulgheroni, E. d’Amico, Free context smartphone based application for motor activity levels recognition, in: 2016
IEEE 2nd International Forum on Research and Technologies for Society and Industry Leveraging a Better Tomorrow, RTSI, 2016, pp. 292–295.
http://dx.doi.org/10.1109/rtsi.2016.7740601.
[28] Y.F. Chen, C. Shen, Performance analysis of smartphone-sensor behavior for human activity recognition, IEEE Access 5 (2017) 3095–3110. http:
//dx.doi.org/10.1109/access.2017.2676168.
[29] Y. Tian, W.J. Chen, MEMS-based human activity recognition using smartphone, in: Proceedings of the 35th Chinese Control Conference 2016, 2016.
pp. 3984–3989. http://dx.doi.org/10.1109/ChiCC.2016.7553975.
[30] P. Vallabh, R. Malekian, N. Ye, D.C. Bogatinoska, Fall detection using machine learning algorithms, in: 2016 24th International Conference on Software,
Telecommunications and Computer Networks, Softcom, 2016, pp. 51–59. http://dx.doi.org/10.1109/softcom.2016.7772142.
[31] C.X. Tang, V.V. Phoha, An empirical evaluation of activities and classifiers for user identification on smartphones, in: 2016 IEEE 8th International
Conference on Biometrics Theory, Applications and Systems, BTAS, 2016, pp. 1–8. http://dx.doi.org/10.1109/btas.2016.7791159.
[32] T.H. Sanders, M.A. Clements, Multimodal monitoring for neurological disorders, in: 2014 40th Annual Northeast Bioengineering Conference, NEBEC,
2014, pp. 1-2. http://dx.doi.org/10.1109/nebec.2014.6972928.
[33] A. Bajpai, V. Jilla, V.N. Tiwari, S.M. Venkatesan, R. Narayanan, Quantifiable fitness tracking using wearable devices, in: Conf Proc IEEE Eng Med Biol Soc,
Vol. 2015, Aug. 2015, pp. 1633–1637. http://dx.doi.org/10.1109/EMBC.2015.7318688.
[34] F. De Cillisy, F. De Simioy, F. Guidoy, R.A. Incalzi, R. Setolay, Fall-detection solution for mobile platforms using accelerometer and gyroscope data, in:
Conf Proc IEEE Eng Med Biol Soc, Vol. 2015, Aug. 2015, pp. 3727–3730. http://dx.doi.org/10.1109/EMBC.2015.7319203.
[35] Q.C. Zhu, Z.H. Chen, Y.C. Soh, Smartphone-based human activity recognition in buildings using locality-constrained linear coding, in: Proceedings of
the 2015 10th IEEE Conference on Industrial Electronics and Applications, 2015, pp. 214–219. http://dx.doi.org/10.1109/iciea.2015.7334113.
[36] Y.Y. Fan, L. Xie, Y.F. Yin, S.L. Lu, A context aware energy-saving scheme for smart camera phones based on activity sensing, in: 2015 IEEE 12th
International Conference on Mobile Ad Hoc and Sensor Systems, MASS, 2015, pp. 64–72. http://dx.doi.org/10.1109/Mass.2015.17.
[37] P. Li, Y. Wang, Y. Tian, T.S. Zhou, J.S. Li, An automatic user-adapted physical activity classification method using smartphones, IEEE Trans. Biomed. Eng.
64 (2017) 706–714. http://dx.doi.org/10.1109/TBME.2016.2573045.
[38] M.T. Uddin, M.M. Billah, M.F. Hossain, Random forests based recognition of human activities and postural transitions on smartphone, in: 2016 5th
International Conference on Informatics, Electronics and Vision, ICIEV, 2016, pp. 250–255. http://dx.doi.org/10.1109/iciev.2016.7760005.
[39] L. Fang, S. Yishui, C. Wei, Up and down buses activity recognition using smartphone accelerometer, in: 2016 IEEE Information Technology, Networking,
Electronic and Automation Control Conference, 2016, pp. 761–765.
[40] M. Altini, R. Vullers, C. Van Hoof, M. van Dort, O. Amft, Self-calibration of walking speed estimations using smartphone sensors, in: 2014 IEEE
International Conference on Pervasive Computing and Communications Workshops, Percom Workshops, 2014, pp. 10–18. http://dx.doi.org/10.1109/
PerComW.2014.6815158.
[41] J. Windau, L. Itti, Situation awareness via sensor-equipped eyeglasses, in: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems,
IROS, 2013, pp. 5674–5679. http://dx.doi.org/10.1109/iros.2013.6697178.
119
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
I.M. Pires et al. / Pervasive and Mobile Computing 47 (2018) 78–93
93
[42] G. Vavoulas, C. Chatzaki, T. Malliotakis, M. Pediaditis, M. Tsiknakis, The MobiAct dataset: Recognition of activities of daily living using smartphones,
in: Proceedings of the International Conference on Information and Communication Technologies for Ageing Well and E-Health, ICT4AWE, 2016,
pp. 143–151. http://dx.doi.org/10.5220/0005792401430151.
[43] H. Bojinov, Y. Michalevsky, G. Nakibly, D. Boneh, Mobile device identification via sensor fingerprinting, 2014, arXiv preprint arXiv:1408.1416.
[44] K. Katevas, H. Haddadi, L. Tokarchuk, Sensingkit: Evaluating the sensor power consumption in IOS devices, in: Intelligent Environments, IE, 2016 12th
International Conference on, 2016, pp. 222–225.
[45] Bq.com., Smartphones BQ Aquaris / BQ Portugal, 2 Sep. 2017. Available: https://www.bq.com/pt/smartphones.
[46] ALLab, August 2017- Multi-sensor data fusion in mobile devices for the identification of activities of daily living - ALLab Signals, September
2nd. Available: https://allab.di.ubi.pt/mediawiki/index.php/August_2017-_Multi-sensor_data_fusion_in_mobile_devices_for_the_identification_of_
activities_of_daily_living.
[47] I.M. Pires, N.M. Garcia, N. Pombo, F. Flórez-Revuelta, Limitations of the use of mobile devices and smart environments for the monitoring of ageing
people, in: ICT4AWE 2018 4th International Conference on Information and Communication Technologies for Ageing Well and e-Health, Madeira,
Portugal, 2018.
[48] V. Graizer, Effect of low-pass filtering and re-sampling on spectral and peak ground acceleration in strong-motion records, in: Proc. 15th World
Conference of Earthquake Engineering, Lisbon, Portugal, 2012, pp. 24–28.
[49] A. Jain, K. Nandakumar, A. Ross, Score normalization in multimodal biometric systems, Pattern Recognit. 38 (2005) 2270–2285. http://dx.doi.org/10.
1016/j.patcog.2005.01.012.
[50] A.Y. Ng, Feature selection, L 1 vs. L 2 regularization, and rotational invariance, in: Proceedings Of the Twenty-First International Conference on Machine
Learning, 2004, p. 78.
[51] L. Brocca, F. Melone, T. Moramarco, W. Wagner, V. Naeimi, Z. Bartalis, et al., Improving runoff prediction through the assimilation of the ASCAT soil
moisture product, Hydrol. Earth Syst. Sci. 14 (2010) 1881–1893. http://dx.doi.org/10.5194/hess-14-1881-2010.
[52] M. Bianchini, F. Scarselli, On the complexity of shallow and deep neural network classifiers, in: ESANN, 2014.
120
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
2. Android library for recognition of activities of daily
living: implementation considerations, challenges,
and solutions
The following article is the second part of the chapter 4.
Android library for recognition of activities of daily living: implementation considerations,
challenges, and solutions
Ivan Miguel Pires, Maria Canavarro Teixeira, Nuno Pombo, Nuno M. Garcia, Francisco FlórezRevuelta, Susanna Spinsante, Rossitza Goleva and Eftim Zdravevski
The Open Bioinformatics Journal (Bentham Science Publishers B.V.), published, 2018.
According to 2016 Journal Citation Reports published by SCImago in 2017, this journal has the
following performance metrics:
CiteScore (2016): 4.86
Journal Ranking (2016): 2/32 (Computer Science (miscellaneous))
121
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
122
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
6HQG2UGHUVIRU5HSULQWVWRUHSULQWV#EHQWKDPVFLHQFHDH
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO
&RQWHQWOLVWDYDLODEOHDWZZZEHQWKDPRSHQFRP72%,2,-
'2,
5(6($5&+$57,&/(
$QGURLG /LEUDU\ IRU 5HFRJQLWLRQ RI $FWLYLWLHV RI 'DLO\ /LYLQJ
,PSOHPHQWDWLRQ&RQVLGHUDWLRQV&KDOOHQJHVDQG6ROXWLRQV
0DULD &DQDYDUUR 7HL[HLUD 1XQR 3RPER 1XQR 0 *DUFLD )UDQFLVFR
,YDQ 0LJXHO 3LUHV
)OhUH]5HYXHOWD6XVDQQD6SLQVDQWH5RVVLW]D*ROHYDDQG(IWLP=GUDYHYVNL
,QVWLWXWRGH7HOHFRPXQLFDdeHV8QLYHUVLGDGHGD%HLUD,QWHULRU&RYLOKc3RUWXJDO
$OWUDQSRUWXJDO/LVERQ3RUWXJDO
$//DE $VVLVWHG /LYLQJ &RPSXWLQJ DQG 7HOHFRPPXQLFDWLRQV /DERUDWRU\ &RPSXWLQJ 6FLHQFH 'HSDUWPHQW
8QLYHUVLGDGH GD %HLUD ,QWHULRU &RYLOKc 3RUWXJDO
87&GH5HFXUVRV1DWXUDLVH'HVHQYROYLPHQWR6XVWHQWbYHO3RO\WHFKQLTXH,QVWLWXWHRI&DVWHOR%UDQFR&DVWHOR%UDQFR
3RUWXJDO
&(51$6 5HVHDUFK &HQWUH IRU 1DWXUDO 5HVRXUFHV (QYLURQPHQW DQG 6RFLHW\ 3RO\WHFKQLTXH ,QVWLWXWH RI &DVWHOR
%UDQFR&DVWHOR%UDQFR3RUWXJDO
'HSDUWPHQWRI&RPSXWHU7HFKQRORJ\8QLYHUVLGDGGH$OLFDQWH$OLFDQWH6SDLQ
'HSDUWPHQWRI,QIRUPDWLRQ(QJLQHHULQJ0DUFKH3RO\WHFKQLF8QLYHUVLW\$QFRQD,WDO\
'HSDUWPHQWRI,QIRUPDWLFV1HZ%XOJDULDQ8QLYHUVLW\6RILD%XOJDULD
)DFXOW\RI&RPSXWHU6FLHQFHDQG(QJLQHHULQJ8QLYHUVLW\6V&\ULODQG0HWKRGLXV6NRSMH0DFHGRQLD
5HFHLYHG)HEUXDU\
5HYLVHG$SULO
$FFHSWHG$SULO
$EVWUDFW
%DFNJURXQG
2IIWKHVKHOIPRELOHGHYLFHVKDYHVHYHUDOVHQVRUVDYDLODEOHRQERDUGWKDWPD\EHXVHGIRUWKHUHFRJQLWLRQRI$FWLYLWLHVRI'DLO\
/LYLQJ $'/ DQGWKHHQYLURQPHQWVZKHUHWKH\DUHSHUIRUPHG7KLVUHVHDUFKLVIRFXVHGRQWKHGHYHORSPHQWRI$PELHQW$VVLVWHG
/LYLQJ $$/ V\VWHPVXVLQJPRELOHGHYLFHVIRUWKHDFTXLVLWLRQRIWKHGLIIHUHQWW\SHVRIGDWDUHODWHGWRWKHSK\VLFDODQGSK\VLRORJLFDO
FRQGLWLRQVRIWKHVXEMHFWVDQGWKHHQYLURQPHQWV0RELOHGHYLFHVZLWKWKH$QGURLG2SHUDWLQJ6\VWHPVDUHWKHOHDVWH[SHQVLYHDQG
H[KLELWWKHELJJHVWPDUNHWZKLOHSURYLGLQJDYDULHW\RIPRGHOVDQGRQERDUGVHQVRUV
2EMHFWLYH
7KLVSDSHUGHVFULEHVWKHLPSOHPHQWDWLRQFRQVLGHUDWLRQVFKDOOHQJHVDQGVROXWLRQVDERXWDIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQG
WKHHQYLURQPHQWVSURYLGHGDVDQ$QGURLGOLEUDU\7KHIUDPHZRUNLVDIXQFWLRQRIWKHQXPEHURIVHQVRUVDYDLODEOHLQGLIIHUHQWPRELOH
GHYLFHVDQGXWLOL]HVDYDULHW\RIDFWLYLW\UHFRJQLWLRQDOJRULWKPVWRSURYLGHDUDSLGIHHGEDFNWRWKHXVHU
0HWKRGV
7KH$QGURLGOLEUDU\LQFOXGHVGDWDIXVLRQGDWDSURFHVVLQJIHDWXUHVHQJLQHHULQJDQGFODVVLILFDWLRQPHWKRGV7KHVHQVRUVWKDWPD\EH
XVHGDUHWKHDFFHOHURPHWHUWKHJ\URVFRSHWKHPDJQHWRPHWHUWKH*OREDO3RVLWLRQLQJ6\VWHP *36 UHFHLYHUDQGWKHPLFURSKRQH
7KHGDWDSURFHVVLQJLQFOXGHVWKHDSSOLFDWLRQRIGDWDFOHDQLQJPHWKRGVDQGWKHH[WUDFWLRQRIIHDWXUHVZKLFKDUHXVHGZLWK'HHS
1HXUDO1HWZRUNV '11 IRUWKHFODVVLILFDWLRQRI$'/DQGHQYLURQPHQW7KURXJKRXWWKLVZRUNWKHOLPLWDWLRQVRIWKHPRELOHGHYLFHV
ZHUHH[SORUHGDQGWKHLUHIIHFWVKDYHEHHQPLQLPL]HG
$GGUHVVFRUUHVSRQGHQFHWRWKLVDXWKRUDWWKH&RPSXWHU6FLHQFH'HSDUWPHQW8QLYHUVLGDGHGD%HLUD,QWHULRU5XD0DUTXeVGlmYLODH%RODPD
&RYLOKb3RUWXJDO7HO(PDLOLPSLUHV#LWXELSW
%HQWKDP2SHQ
123
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
5HVXOWV
7KHLPSOHPHQWDWLRQRIWKH$QGURLGOLEUDU\UHSRUWHGDQRYHUDOODFFXUDF\EHWZHHQDQGGHSHQGLQJRQWKHQXPEHURI
VHQVRUV XVHG DQG WKH QXPEHU RI $'/ DQG HQYLURQPHQWV UHFRJQL]HG &RPSDUHG ZLWK WKH UHVXOWV DYDLODEOH LQ WKH OLWHUDWXUH WKH
SHUIRUPDQFHRIWKHOLEUDU\UHSRUWHGDPHDQLPSURYHPHQWRIDQGWKH\GRQRWGLIIHUDWWKHPD[LPXPIRXQGLQSULRUZRUNWKDW
EDVHGRQWKH6WXGHQWlVWWHVW
&RQFOXVLRQ
7KLVVWXG\SURYHVWKDW$'/OLNHZDONLQJJRLQJXSVWDLUVDQGGRZQVWDLUVUXQQLQJZDWFKLQJ79GULYLQJVOHHSLQJDQGVWDQGLQJ
DFWLYLWLHV DQG WKH EHGURRP FRRNLQJNLWFKHQ J\P FODVVURRP KDOO OLYLQJ URRP EDU OLEUDU\ DQG VWUHHW HQYLURQPHQWV PD\ EH
UHFRJQL]HGZLWKWKHVHQVRUVDYDLODEOHLQRIIWKHVKHOIPRELOHGHYLFHV)LQDOO\WKHVHUHVXOWVPD\DFWDVDSUHOLPLQDU\UHVHDUFKIRUWKH
GHYHORSPHQWRIDSHUVRQDOGLJLWDOOLIHFRDFKZLWKDPXOWLVHQVRUPRELOHGHYLFHFRPPRQO\XVHGGDLO\
.H\ZRUGV$FWLYLWLHVRIGDLO\OLYLQJ6HQVRUV0RELOHGHYLFHV3DWWHUQUHFRJQLWLRQ'DWDIXVLRQ$QGURLGOLEUDU\$UWLILFLDOQHXUDO
QHWZRUNV5HFRJQLWLRQ
,1752'8&7,21
%DFNJURXQG
0RELOHGHYLFHVOLNHVPDUWSKRQHVFRPPRQO\XVHGLQGDLO\OLIHFDQEHFODVVLILHGDVPXOWLVHQVRUGHYLFHV>@EHFDXVH
WKH\KRVWVHYHUDOVHQVRUVHJDFFHOHURPHWHUJ\URVFRSHPDJQHWRPHWHUPLFURSKRQH*OREDO3RVLWLRQLQJ6\VWHP *36
UHFHLYHU DQG FDPHUD WKDW DOORZ WKH DFTXLVLWLRQ RI GLIIHUHQW W\SHV RI GDWD LQFOXGLQJ VRPH RI WKH SK\VLFDO DQG
SK\VLRORJLFDOGDWDSHUWDLQLQJWRWKHXVHUWKXVDOORZLQJIRUWKHUHFRJQLWLRQRIWKH$FWLYLWLHVRI'DLO\/LYLQJ $'/ >
@
7KH GHYHORSPHQW RI $PELHQW $VVLVWHG /LYLQJ $$/ V\VWHPV KDV PDQ\ FKDOOHQJHV > @ VXFK DV WLPHO\ DQG
DFFXUDWHUHFRJQLWLRQRI$'/>@GHVLJQLQJDUFKLWHFWXUHVWKDWDUHDSSOLFDEOHIRU(QKDQFHG/LYLQJ(QYLURQPHQWV>
@SURYLGLQJUHOLDEOHV\VWHPVDQGLQWHJUDWLRQZLWKRWKHUV\VWHPV>@$'/UHFRJQLWLRQLVDOVRLPSRUWDQWWRGHVLJQD
SHUVRQDOGLJLWDOOLIHFRDFK>@7KHGHYHORSPHQWRIWKHVHV\VWHPVLVVSHFLDOO\LPSRUWDQWWRVXSSRUWWKHDXWRQRP\RI
ROGHUXVHUVSDWLHQWVZLWKFKURQLFGLVHDVHVDQGXVHUVWKDWPD\KDYHVRPHW\SHRIGLVDELOLW\>@+RZHYHUWKHVH
V\VWHPVPD\EHXVHIXOIRUHYHU\RQHLQFOXGLQJDWKOHWHVDQG\RXQJXVHUVDVWKHSURSRVHGIUDPHZRUNFDQEHLQWHJUDWHG
LQWRDWRROIRUWKHPRQLWRULQJDQGWUDLQLQJRIOLIHVW\OHV>@
0RWLYDWLRQ
7KHUHFRJQLWLRQRI$'/DQGWKHHQYLURQPHQWZKHUHWKHVXEMHFWLVDWDJLYHQPRPHQWLVZLGHO\LPSRUWDQWIRUWKH
FUHDWLRQ RI V\VWHPV DLPHG DW VHYHUDO SXUSRVHV VXFK DV WKH SHUVRQDO FRDFKLQJ WKH KHDOWK PRQLWRULQJ WKH OLIHVW\OH
PRQLWRULQJDQGRWKHUV1RZDGD\VVXFKUHFRJQLWLRQLVFRPPRQO\SHUIRUPHGE\FRPSOH[DQGH[SHQVLYHV\VWHPVZLWK
KLJK SRZHU SURFHVVLQJ DQG PHPRU\ FDSDELOLWLHV > @ EXW LW FRXOG EH SHUIRUPHG ZLWK FRPPRQO\ XVHG
HTXLSPHQWEDVHGRQORFDOSURFHVVLQJWHFKQLTXHVLQRUGHUWRSURYLGHDIHHGEDFNUHODWHGWRWKH$'/SHUIRUPHGDQG
HQYLURQPHQWIUHTXHQWHGLQDOPRVWUHDOWLPH
2XUPRWLYDWLRQLVWRVROYHWKLVSUREOHPZLWKWKHGHYHORSPHQWRIDQ$QGURLGOLEUDU\WKDWSHUIRUPVWKHUHFRJQLWLRQRI
$'/DQGWKHDVVRFLDWHGHQYLURQPHQWVXVLQJRQO\WKHGDWDDFTXLUHGIURPWKHVHQVRUVDYDLODEOHRQERDUGRIIWKHVKHOI
PRELOHGHYLFHV LHVPDUWSKRQHV DQGORFDOSURFHVVLQJ:HVHOHFWHGWKH$QGURLGSODWIRUPIRUWKHGHYHORSPHQWRIWKH
OLEUDU\EHFDXVHWKLVLVWKHPRVWXVHGRSHQSODWIRUPDYDLODEOHLQWKHPDUNHW>@DQGWKHDUFKLWHFWXUHRIDQ$QGURLG
OLEUDU\IRUWKHUHFRJQLWLRQRI$'/DQGHQYLURQPHQWVKDVEHHQDOUHDG\SURSRVHGLQ>@
7KHXVHRIWKHRIIWKHVKHOIPRELOHGHYLFHVIRUWKLVSXUSRVHUHTXLUHVWKHGHYHORSPHQWRIOLJKWZHLJKWSURFHVVLQJ
WHFKQLTXHV ,Q IDFW GHVSLWH UHFHQW GHYHORSPHQWV DQG LPSURYHPHQWV WKHVH GHYLFHV XVXDOO\ KDYH VHYHUDO OLPLWDWLRQV
LQFOXGLQJUHGXFHGSURFHVVLQJFDSDELOLWLHVORZFDSDFLW\EDWWHULHVDQGVPDOOLQWHUQDOVWRUDJHVSDFH
+RZHYHU EDVHG RQ SUHYLRXV VWXGLHV > @ ZH FRQILUP WKDW WKH XVH RI WKH VHQVRUV DYDLODEOH LQ RIIWKHVKHOI
PRELOHGHYLFHVPD\IXHOWKHGHYHORSPHQWRI$'/UHFRJQLWLRQV\VWHPVKDYLQJWKHSRWHQWLDOWRLPSURYHWKHTXDOLW\RI
OLIH RI LWV XVHUV :LWK WKLV VWXG\ ZH SUHVHQW D UHDOWLPH PRELOH DSSOLFDWLRQ WKDW UHOLDEO\ LGHQWLILHV $'/V WKH
HQYLURQPHQW DQG ORFDWLRQ SURYLGLQJ IHHGEDFN WR WKH XVHU 7KH UHFRJQLWLRQ RI WKH $'/ IHDWXULQJ WKH VXEMHFWlV
PRYHPHQWLVEDVHGRQWKHGDWDDFTXLUHGIURPWKHPRWLRQDQGRUPDJQHWLFVHQVRUV LHDFFHOHURPHWHUJ\URVFRSHDQG
PDJQHWRPHWHU EXWWKHUHFRJQLWLRQRIWKHHQYLURQPHQWLVEDVHGRQDFRXVWLFGDWDDFTXLUHGIURPWKHPLFURSKRQH,QRUGHU
124
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
WRVXSSRUWWKHEHWWHUUHFRJQLWLRQRIDQ$'/ZKHQWKHXVHULVVWDQGLQJVWLOOWKHIXVLRQRIWKHIHDWXUHVH[WUDFWHGIURPWKH
GDWDDFTXLUHGIURPPRWLRQDQGPDJQHWLFVHQVRUVZLWKWKHHQYLURQPHQWSUHYLRXVO\UHFRJQL]HGDQGWKHIHDWXUHVH[WUDFWHG
IURPWKHGDWDDFTXLUHGIURPWKH*36UHFHLYHULVSHUIRUPHG
3ULRU:RUN
7R GDWH DQG EDVHG RQ WKH OLWHUDWXUH WKHUH DUH VHYHUDO VWXGLHV XVLQJ GLIIHUHQW VXEVHWV RI VHQVRUV DYDLODEOH RQ
VPDUWSKRQHV<HWVWXGLHVUHSRUWLQJPHWKRGVWKDWUHO\RQDODUJHUVHWRIVHQVRUVDYDLODEOHLQRIIWKHVKHOIPRELOHGHYLFHV
IRUWKHUHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWVZHUHQRWIRXQG)RUWKHSXUSRVHRIFRPSDULVRQZLWKWKLVUHVHDUFK
ZHDQDO\]HGVWXGLHVDYDLODEOHLQWKHOLWHUDWXUHWKDWXVH$UWLILFLDO1HXUDO1HWZRUNV $11 IRUWKHUHFRJQLWLRQRI$'/
DQGHQYLURQPHQWV7KHLPSOHPHQWDWLRQRI$11ZLWKPRELOHGHYLFHVLVZLGHO\VWXGLHGDQGVHYHUDOIUDPHZRUNVKDYH
EHHQ GHYHORSHG ,Q WKLV FDVH RWKHU W\SHV RI PHWKRGV KDYH EHHQ DOUHDG\ XVHG LQ WKH OLWHUDWXUH LQFOXGLQJ WKH QHXUDO
QHWZRUNVDQGWKHHQVHPEOHOHDUQLQJPHWKRGV>@EXWWKHHQVHPEOHOHDUQLQJPHWKRGVRIWHQLQFOXGHDFRPELQDWLRQRI
PHWKRGV WKXVUHTXLULQJKLJKHUFRPSXWDWLRQDOUHVRXUFHV DQGFDQEHXVHGZLWKODUJHUGDWDVHWVWKDQWKHQHXUDOQHWZRUNV
PHWKRGVZKLFKZLOOEHWKHFDVHIRUWKHGDWDVHWVDYDLODEOHLQWKHVWRUDJHRIPRELOHGHYLFHV>@
7KHXVHRIWKH$11ZLWKYDULDQWVWRUHFRJQL]HGLIIHUHQWW\SHVRI$'/LQYROYLQJWKHVXEMHFWlVPRYHPHQWVLVTXLWH
ZHOOGRFXPHQWHGLQWKHOLWHUDWXUH6HYHUDOZRUNVDVWKHIROORZLQJGLVFXVVLRQVKRZVDGGUHVVHGWKLVPHWKRGSURYLGLQJ
GLIIHUHQWDFFXUDF\SHUIRUPDQFHRQGLIIHUHQWGDWDVHWV7KHUHDUHVHYHUDOVWXGLHVXVLQJRQO\WKHDFFHOHURPHWHUVHQVRU
ZKHUHWKHDXWKRUVRI>@UHFRJQL]HGWKHZDONLQJDQGVWDQGLQJDFWLYLWLHVXVLQJ0XOWLOD\HU3HUFHSWURQ 0/3 ZLWKD
UHSRUWHG DFFXUDF\ DURXQG %D\DW HW DO >@ LPSOHPHQWHG WKH 0/3 PHWKRG XVLQJ IHDWXUHV H[WUDFWHG IURP WKH
DFFHOHURPHWHUGDWDIRUWKHUHFRJQLWLRQRIUXQQLQJZDONLQJGDQFLQJJRLQJXSVWDLUVDQGJRLQJGRZQVWDLUVUHSRUWLQJDQ
DFFXUDF\ RI ,Q >@ WKH DXWKRUV XVHG WKH DFFHOHURPHWHU GDWD DQG LPSOHPHQWHG WKH 0/3 PHWKRG IRU WKH
UHFRJQLWLRQRIJRLQJXSVWDLUVJRLQJXSRQDQHVFDODWRUDQGZDONLQJUHSRUWLQJDQDYHUDJHDFFXUDF\RI7KHVWXG\
RI>@LPSOHPHQWHGWKH0/3PHWKRGIRUWKHUHFRJQLWLRQRIVHYHUDODFWLYLWLHVVXFKDVZDONLQJMRJJLQJJRLQJXSVWDLUV
JRLQJGRZQVWDLUVVLWWLQJVWDQGLQJDQGO\LQJUHSRUWLQJDPD[LPXPDFFXUDF\RI7KHVWDQGLQJZDONLQJDQG
UXQQLQJDFWLYLWLHVUHSRUWHGDPLQLPXPDFFXUDF\RIXVLQJWKH0/3PHWKRG>@7KHDXWKRUVRI>@LPSOHPHQWHG
WKH0/3PHWKRGIRUWKHFRPSDULVRQRIWKHUHVXOWVREWDLQHGZLWKWKH6OLGLQJ:LQGRZEDVHG+LGGHQ0DUNRY0RGHO
6:+00 UHSRUWLQJDQDFFXUDF\DOZD\VKLJKHUWKDQLQWKHUHFRJQLWLRQRIZDONLQJVWDQGLQJUXQQLQJJRLQJ
GRZQVWDLUVDQGJRLQJXSVWDLUV0/3PHWKRGZDVLPSOHPHQWHGE\WKHDXWKRUVRI>@IRUWKHUHFRJQLWLRQRIVWDQGLQJ
VLWWLQJZDONLQJO\LQJDQGIDOOLQJDFWLYLWLHVUHSRUWLQJDQDFFXUDF\KLJKHUWKDQ8VLQJWKHVDPHPHWKRGLQ>@WKH
DXWKRUVUHSRUWHGWKDWWKHUHFRJQLWLRQRIWKHVWDQGLQJZDONLQJMRJJLQJMXPSLQJJRLQJXSVWDLUVJRLQJGRZQVWDLUVDQG
VLWWLQJDFWLYLWLHVKDVDQDFFXUDF\KLJKHUWKDQ,Q>@WKHZDONLQJDQGVWDQGLQJDFWLYLWLHVUHSRUWHGDUHFRJQLWLRQ
DFFXUDF\EHWZHHQDQGXVLQJWKH0/3PHWKRG7KH0/3PHWKRGZDVXVHGE\WKHDXWKRUVRI>@IRUWKH
UHFRJQLWLRQRIVWDQGLQJZDONLQJJRLQJXSVWDLUVJRLQJGRZQVWDLUVDQGUXQQLQJDFWLYLWLHVUHSRUWLQJDQDFFXUDF\RI
&\FOLQJVWDQGLQJZDONLQJMRJJLQJGULYLQJDQGUXQQLQJDUHRWKHUDFWLYLWLHVUHFRJQL]HGE\WKHDXWKRUVRI>@ZLWKWKH
0/3 PHWKRG UHSRUWLQJ DQ DFFXUDF\ EHWZHHQ DQG ,Q >@ WKH 0/3 PHWKRG ZDV XVHG IRU WKH
UHFRJQLWLRQRIZDONLQJJRLQJXSVWDLUVJRLQJGRZQVWDLUVVWDQGLQJVLWWLQJDQGMRJJLQJDFWLYLWLHVUHSRUWLQJDQDFFXUDF\
DURXQG,QRUGHUWRLGHQWLI\WKHDFWLYLWLHVSHUIRUPHGGXULQJDIRRWEDOOJDPHWKHDXWKRUVRI>@XVHGWKH0/3
PHWKRGZLWKDUHSRUWHGDFFXUDF\KLJKHUWKDQ,Q>@WKHZDONLQJJRLQJXSVWDLUVMRJJLQJJRLQJGRZQVWDLUV
VWDQGLQJDQGVLWWLQJDFWLYLWLHVZHUHUHFRJQL]HGZLWKWKH0/3PHWKRGZKLFKUHSRUWHGDQDFFXUDF\RI7KHVWXG\
>@SUHVHQWHGWKHLPSOHPHQWDWLRQRIWKH0/3PHWKRGIRUWKHUHFRJQLWLRQRIZDONLQJMRJJLQJJRLQJXSVWDLUVJRLQJ
GRZQVWDLUVVLWWLQJDQGVWDQGLQJDFWLYLWLHVUHSRUWHGDQDYHUDJHDFFXUDF\RI,Q>@WKH0/3PHWKRGZDVXVHG
IRUWKHUHFRJQLWLRQRIF\FOLQJUXQQLQJVLWWLQJVWDQGLQJDQGZDONLQJZLWKDPLQLPXPUHSRUWHGDFFXUDF\RI,Q
DGGLWLRQWRWKHXVHRIWKH0/3PHWKRGZLWKWKHDFFHOHURPHWHUGDWDDVLQSXWRWKHUPHWKRGVDUHJURZLQJVXFKDV'11
DQG3UREDELOLVWLF1HXUDO1HWZRUNV 311 PHWKRGV'11PHWKRGZDVXVHGE\WKHDXWKRUVRI>@IRUWKHUHFRJQLWLRQRI
ZDONLQJ UXQQLQJ VWDQGLQJ VLWWLQJ O\LQJ JRLQJ XSVWDLUV DQG JRLQJ GRZQVWDLUV DFWLYLWLHV ZLWK D UHSRUWHG DFFXUDF\
EHWZHHQDQG311PHWKRGZDVXVHGE\WKHDXWKRUVRI>@IRUWKHUHFRJQLWLRQRIZDONLQJMRJJLQJF\FOLQJ
JRLQJXSVWDLUVDQGJRLQJGRZQVWDLUVDFWLYLWLHVUHSRUWLQJUHVXOWVZLWKDQDYHUDJHDFFXUDF\RI
)ROORZLQJWKHVWXGLHVWKDWIXVHWKHGDWDDFTXLUHGIURPWKHDFFHOHURPHWHUDQGWKHJ\URVFRSHVHQVRUVWKHDXWKRUVRI
>@XVHGWKH0/3PHWKRGIRUWKHUHFRJQLWLRQRIVWDQGLQJZDONLQJUXQQLQJVLWWLQJO\LQJF\FOLQJGULYLQJFOHDQLQJ
FRRNLQJWDNLQJPHGLFDWLRQVZHHSLQJZDVKLQJKDQGVDQGZDWHULQJSODQWVDFWLYLWLHVUHSRUWLQJDQDYHUDJHDFFXUDF\RI
7KH0/3PHWKRGZDVXVHGLQ>@IRUWKHUHFRJQLWLRQRIZDONLQJMXPSLQJUXQQLQJJRLQJGRZQVWDLUVDQG
JRLQJXSVWDLUVUHSRUWLQJDQDYHUDJHDFFXUDF\RI)RUWKHUHFRJQLWLRQRIIDOOLQJDFWLYLWLHVWKHDXWKRUVRI>@
LPSOHPHQWHGWKH0/3PHWKRGDQGUHSRUWHGDQDFFXUDF\RI7KHZDONLQJMRJJLQJVLWWLQJJRLQJXSVWDLUVDQG
125
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
JRLQJGRZQVWDLUVZHUHUHFRJQL]HGZLWKWKHXVHRIWKH0/3PHWKRGUHSRUWLQJDQDYHUDJHDFFXUDF\RI>@,Q
>@WKHDXWKRUVLPSOHPHQWHGWKH0/3PHWKRGIRUWKHUHFRJQLWLRQRIPRYHPHQWVUHODWHGWRJHWWLQJXSDQGGRZQIURP
EXVHVUHSRUWLQJDQDYHUDJHDFFXUDF\RI
$ V\VWHP WKDW XVHV 6XSSRUW 9HFWRU 0DFKLQHV 690V ZLWK WXQHG SDUDPHWHUV DQG IHDWXUHV EDVHG RQ KLVWRJUDPV
H[WUDFWHG IURP PXOWLSOH DFFHOHURPHWHUV DQG J\URVFRSHV LV SURSRVHG LQ >@ 7KLV LV D YDOLG DSSURDFK EHFDXVH VXFK
IHDWXUHVDUHHDV\WRFRPSXWH$EDODQFHGDFFXUDF\RYHULVUHSRUWHG
$QDSSURDFKEDVHGRQDXWRPDWLFIHDWXUHHQJLQHHULQJLVXVHGLQ>@IRUWKHUHFRJQLWLRQRIDYDULHW\RI$'/VXVLQJ
VHQVRUVSODFHGRQPXOWLSOHORFDWLRQV'HSHQGLQJRQWKHGDWDVHWDQDFFXUDF\IURPWRLVUHSRUWHG,QWHUHVWLQJO\
WKHEHVWDFFXUDF\DFKLHYHGZLWK690V5DQGRP)RUHVWDQG([WUD7UHHVLVFRPSDUDEOHWRWKHDFFXUDF\RIPXFKPRUH
OLJKWZHLJKWDSSURDFKHVOLNHORJLVWLFUHJUHVVLRQDQG1DgYH%D\HV$VLPLODUDSSURDFKZDVSURYHQWRZRUNIRUMRJJLQJ
GHWHFWLRQLQ>@HYHQRQILHOGFRQGLWLRQVZKHQWKHVXEMHFWZHDUVDFFHOHURPHWHUVDOOWKHWLPH
,Q>@WKHDFFHOHURPHWHUDQGJUDYLW\VHQVRUVlGDWDZHUHIXVHGDQGWKH0/3PHWKRGXVHGIRUWKHUHFRJQLWLRQRI
ZDONLQJJRLQJXSVWDLUVJRLQJGRZQVWDLUVVWDQGLQJDQGOD\LQJDFWLYLWLHVUHSRUWLQJDQDFFXUDF\KLJKHUWKDQ
&RPELQLQJWKHGDWDDFTXLUHGIURPWKHDFFHOHURPHWHUDQGWKH*36UHFHLYHUWKH0/3PHWKRGKDQGOHVWKHUHFRJQLWLRQ
RIVWDQGLQJZDONLQJDQGWUDYHOOLQJE\FDUDQGWUDLQUHSRUWLQJDPLQLPXPDFFXUDF\RI>@
,Q>@WKHGDWDDFTXLUHGIURPWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUWKHJ\URVFRSHDQGWKH*36UHFHLYHUZHUH
IXVHG DQG WKH 0/3 PHWKRG ZDV DSSOLHG IRU WKH UHFRJQLWLRQ RI UXQQLQJ ZDONLQJ VLWWLQJ JRLQJ XSVWDLUV JRLQJ
GRZQVWDLUV DQG VLWWLQJ DFWLYLWLHV UHSRUWLQJ D UHFRJQLWLRQ DFFXUDF\ KLJKHU WKDQ
7KHDFRXVWLFGDWDPD\EHH[SORLWHGIRUWKHUHFRJQLWLRQRIWKHGLIIHUHQWHQYLURQPHQWVXVLQJ$11$XWKRUVRI>@
LPSOHPHQWHG WKH 0/3 PHWKRG IRU WKH UHFRJQLWLRQ RI VRXQGV RI HPHUJHQF\ YHKLFOHV UHSRUWLQJ DQ DFFXUDF\ DURXQG
,Q>@VHYHUDOVRXQGVKDYHEHHQUHFRJQL]HGZLWKWKH0/3PHWKRGVXFKDVEROOLPSDFWPHWDOLPSDFWZRRG
LPSDFWSODVWLFLPSDFWRSHQFORVHGRRUW\SLQJNQRFNLQJWHOHSKRQHULQJLQJJUDLQVIDOOLQJVSUD\DQGZKLVWOHUHSRUWLQJ
DQDYHUDJHDFFXUDF\LQHQYLURQPHQWUHFRJQLWLRQRI7KHVRXQGVRIVQHH]LQJGRJEDUNLQJFORFNWLFNLQJEDE\
FU\LQJFURZLQJURRVWHUUDLQLQJVHDZDYHVILUHFUDFNOLQJKHOLFRSWHUDQGFKDLQVDZZHUHUHFRJQL]HGLQ>@XVLQJWKH
0/3 PHWKRG DQG UHSRUWLQJ DQ DFFXUDF\ DURXQG 7KH DXWKRUV RI >@ UHFRJQL]HG VHYHUDO VRXQGV LQFOXGLQJ
DODUPVELUGVFODSSLQJGRJVPRWRUF\FOHVUDLQLQJVHDZDYHVDQGZLQGEDVHGRQWKHDSSOLFDWLRQRIWKH0/3PHWKRG
DQGUHSRUWHGDQDFFXUDF\DURXQG2WKHU$11PHWKRGVKDYHEHHQXVHGVXFKDV)HHGIRUZDUGQHXUDOQHWZRUNV
)11 5HFXUUHQW1HXUDO1HWZRUNV 511 DQG'11PHWKRGV7KH)11ZHUHLPSOHPHQWHGE\WKHDXWKRUVRI>@IRU
WKHUHFRJQLWLRQRIHPHUJHQF\YHKLFOHVFDUKRUQVDQGQRUPDOVWUHHWVRXQGVUHSRUWLQJDQDFFXUDF\EHWZHHQDQG
7KHDXWKRUVRI>@LPSOHPHQWHG'11DQG511PHWKRGVIRUWKHUHFRJQLWLRQRIWKHVRXQGVUHODWHGWREDE\
FU\LQJDQGVPRNLQJDODUPUHSRUWLQJDQDFFXUDF\DURXQG7KH'11PHWKRGVZDVDOVRXVHGIRUWKHUHFRJQLWLRQRI
GLIIHUHQWVRXQGVLQFOXGLQJODXJKLQJVLQJLQJFU\LQJDUJXLQJVLJKLQJYRLFHDQGPXVLFUHSRUWLQJDFFXUDFLHVDURXQG
>@
7KHDFRXVWLFGDWDPD\DOVREHXVHGWRKHOSLQWKHUHFRJQLWLRQRIVRPH$'/WKHDXWKRUVRI>@UHFRJQL]HGWKH
IDOOLQJHYHQWVZLWKWKH0/3PHWKRGUHSRUWLQJDQDFFXUDF\DURXQG
$VXPPDU\RIWKHOLWHUDWXUHRYHUYLHZGLVFXVVHGDERYHLVSURYLGHGLQ7DEOHVWRZKHUHGHWDLOVUHJDUGLQJWKH
QXPEHU RI $'/V DQDO\]HG E\ HDFK VWXG\ DQG WKH DFFXUDF\ REWDLQHG RQ DYHUDJH DQG RQ WKH VLQJOH W\SH RI $'/
FRQVLGHUHGDUHSUHVHQWHG+RZHYHUWKHDFFXUDF\YDOXHVUHSRUWHGE\WKHVWXGLHVDQDO\]HGDUHQRWGLUHFWO\FRPSDUDEOHDV
WKH\ZHUHREWDLQHGIURPGLIIHUHQWGDWDVHWV
7DEOH 6XPPDU\ RI WKH OLWHUDWXUH VWXGLHV UHODWHG WR $'/ UHFRJQLWLRQ EDVHG RQ WKH $11 PHWKRG 7KH QXPEHU RI $'/
DQDO\]HGWKHDYHUDJHDFFXUDF\DQGWKHDFFXUDF\UHODWHGWRHDFK$'/DUHUHSRUWHG
$YHUDJH
$'/ $FFXUDF\
$FFXUDF\ $FFXUDF\
SHU
*RLQJ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\
6WXG\ SHU
$FFXUDF\ $FFXUDF\ $FFXUDF\ *RLQJ
5HIV VWXG\ VWXG\
6WDQGLQJ :DONLQJ 5XQQLQJ 8SVWDLUV 'RZQVWDLUV -RJJLQJ 6LWWLQJ
/\LQJ
)DOOLQJ &\FOLQJ 'ULYLQJ
>@
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
>@
k
k
k
k
k
k
k
k
126
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
(Table 1) contd.....
$YHUDJH
$'/ $FFXUDF\
$FFXUDF\ $FFXUDF\
SHU
*RLQJ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\
6WXG\ SHU
$FFXUDF\ $FFXUDF\ $FFXUDF\ *RLQJ
5HIV VWXG\ VWXG\
6WDQGLQJ :DONLQJ 5XQQLQJ 8SVWDLUV 'RZQVWDLUV -RJJLQJ 6LWWLQJ
/\LQJ
)DOOLQJ &\FOLQJ 'ULYLQJ
>@
k
k
k
k
k
k
>@
k
k
k
k
k
k
>@
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
>@
k
k
k
k
k
>@
k
k
k
k
k
>@
k
k
k
k
>@
k
k
k
k
k
>@
k
k
k
k
k
>@
k
k
k
k
>@
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
>@
k
k
k
k
k
k
>@
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
>@
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
7DEOH 6XPPDU\ RI WKH OLWHUDWXUH VWXGLHV UHODWHG WR $'/ UHFRJQLWLRQ EDVHG RQ WKH $11 PHWKRG 7KH QXPEHU RI $'/
DQDO\]HGWKHDYHUDJHDFFXUDF\DQGWKHDFFXUDF\UHODWHGWRHDFK$'/DUHUHSRUWHG FRQW
$YHUDJH
$FFXUDF\
$FFXUDF\
$'/ $FFXUDF\
*RLQJRQ
$FFXUDF\ $FFXUDF\
$FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ *RLQJ
SHU $FFXUDF\
DQ
6WXG\ SHU
$FFXUDF\ $FFXUDF\ $FFXUDF\ 3OD\
7DNLQJ $FFXUDF\ :DVKLQJ :DWHULQJ 7UDYHOOLQJ 7UDYHOOLQJ *RLQJXS 'RZQWR
%XV
5HIV VWXG\ VWXG\
'DQFLQJ (VFDODWRU -XPSLQJ &\FOLQJ 'ULYLQJ )RRVEDOO 0HGLFDWLRQ 6ZHHSLQJ +DQGV
3ODQWV
E\&DU E\7UDLQ WR%XV
>@
k
k
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
k
7DEOH6XPPDU\RIWKHOLWHUDWXUHVWXGLHVUHODWHGWRHQYLURQPHQWUHFRJQLWLRQEDVHGRQWKH$11PHWKRG7KHQXPEHURI
HQYLURQPHQWVDQDO\]HGWKHDYHUDJHDFFXUDF\DQGWKHDFFXUDF\UHODWHGWRHDFKHQYLURQPHQWDUHUHSRUWHG
$YHUDJH
$FFXUDF\ $FFXUDF\ $FFXUDF\
$FFXUDF\ $FFXUDF\
$FFXUDF\ $FFXUDF\
%DE\ &URZLQJ $FFXUDF\
6HD
)LUH
6WXG\ (QYLURQPHQWV
$FFXUDF\ $FFXUDF\ $FFXUDF\
SHU (PHUJHQF\ $FFXUDF\ $FFXUDF\ &ORFN
SHUVWXG\
9HKLFOHV 6QHH]LQJ 'RJV
7LFNLQJ &U\LQJ 5RRVWHU 5DLQLQJ :DYHV &UDFNOLQJ +HOLFRSWHU &KDLQVDZ 6WUHHW
5HIV
VWXG\
>@
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
127
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
7DEOH6XPPDU\RIWKHOLWHUDWXUHVWXGLHVUHODWHGWRHQYLURQPHQWUHFRJQLWLRQEDVHGRQWKH$11PHWKRG7KHQXPEHURI
HQYLURQPHQWVDQDO\]HGWKHDYHUDJHDFFXUDF\DQGWKHDFFXUDF\UHODWHGWRHDFKHQYLURQPHQWDUHUHSRUWHG FRQW
$YHUDJH
$FFXUDF\
$FFXUDF\ $FFXUDF\
$FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ 2SHQRU
%ROO
0HWDO
:RRG
3ODVWLF
*UDLQV $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\
SHU
&ORVH $FFXUDF\ $FFXUDF\ 3KRQH
6WXG\ (QYLURQPHQWV
SHUVWXG\
,PSDFW ,PSDFW ,PSDFW ,PSDFW
VWXG\
'RRU
5HIV
7\SLQJ .QRFNLQJ 5LQJLQJ )DOOLQJ
6SUD\ :KLVWOH
9RLFH
0XVLF
>@
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
7DEOH6XPPDU\RIWKHOLWHUDWXUHVWXGLHVUHODWHGWRHQYLURQPHQWUHFRJQLWLRQEDVHGRQWKH$11PHWKRG7KHQXPEHURI
HQYLURQPHQWVDQDO\]HGWKHDYHUDJHDFFXUDF\DQGWKHDFFXUDF\UHODWHGWRHDFKHQYLURQPHQWDUHUHSRUWHG FRQW
$YHUDJH
$FFXUDF\
$FFXUDF\
6WXG\ (QYLURQPHQWV
$FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ 6PRNLQJ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\ $FFXUDF\
SHU
5HIV
7UDIILF
%LUGV &ODSSLQJ )RRWVWHSV 0RWRUF\FOHV 5LYHUV
:LQG
VWXG\ /DXJKLQJ 6LQJLQJ $UJXLQJ 6LJKLQJ :DWHU
SHUVWXG\
$ODUP
>@
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
k
k
k
k
k
k
>@
k
k
k
k
k
k
7DNLQJLQWRDFFRXQWWKHVHWRIVWXGLHVZHVHOHFWHGDQGDQDO\]HGIURPWKHOLWHUDWXUHDVOLVWHGLQ7DEOHVWRDQGLQ
WKH ELEOLRJUDSK\ VHFWLRQ RI WKLV ZRUN ZH FDQ FRQFOXGH WKDW WKH PDMRULW\ RI WKH VWXGLHV XVHV RQO\ WKH
DFFHOHURPHWHUIRUWKHUHFRJQLWLRQRI$'/DQGRQO\VWXGLHV XVHGDFRXVWLFGDWDIRUWKHUHFRJQLWLRQRIWKH
HQYLURQPHQWV1HYHUWKHOHVVVWXGLHV SUHVHQWPHWKRGVWKDWIXVHWKHGDWDDFTXLUHGIURPWKHDFFHOHURPHWHUDQG
J\URVFRSHVHQVRUV2WKHUFRPELQDWLRQVDUHLQYHVWLJDWHGLQWKHOLWHUDWXUHVXFKDVWKHIXVLRQRIWKHGDWDDFTXLUHGIURPWKH
DFFHOHURPHWHUDQG*36UHFHLYHU WKHIXVLRQRIWKHGDWDDFTXLUHGIURPWKHDFFHOHURPHWHUDQGWKHJUDYLW\VHQVRUV
DQGWKHIXVLRQRIWKHGDWDDFTXLUHGIURPWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUWKHJ\URVFRSHDQGWKH*36
UHFHLYHU
7KHQXPEHURI$'/UHFRJQL]HGE\HDFKVWXG\YDULHVEHWZHHQDQGEXWLQRIWKHVWXGLHVWKDWQXPEHUYDULHV
EHWZHHQDQG,QKDOIRIWKDWVWXGLHVWKHQXPEHURI$'/UHFRJQL]HGLVXSWRLQFOXGHG,QFDVHRIWKHQXPEHURI
HQYLURQPHQWVUHFRJQL]HGE\HDFKVWXG\WKH\YDU\EHWZHHQDQGEXWLWVGLVSHUVLRQLVHQRUPRXV,QRIWKDW
VWXGLHVWKHQXPEHURIHQYLURQPHQWVUHFRJQL]HGLVXSWRDQGWKHUHVW DUHDQG
7DEOHVWRSUHVHQWWKHPDSSLQJEHWZHHQWKHVWXG\DQDO\VHVDQGWKH$'/DQGHQYLURQPHQWVUHFRJQL]HGLQFOXGLQJ
WKHDFFXUDFLHVUHSRUWHGIRUIXUWKHUFRPSDULVRQZLWKRXUVWXG\7KHDFFXUDFLHVSUHVHQWHGDUHUHODWHGWRHDFK$'/DQG
HQYLURQPHQWUHFRJQLWLRQ
)LQDOO\7DEOHVWRVXPPDUL]HWKHDYHUDJHVWDQGDUGGHYLDWLRQDQGWKHFRHIILFLHQWRIYDULDWLRQRIWKHVWDQGDUG
GHYLDWLRQRIWKHDFFXUDF\IRUHDFK$'/DQGHQYLURQPHQWUHFRJQL]HGLQWKHSUHYLRXVZRUNDQDO\]HG7KHORZFRHIILFLHQW
RIYDULDWLRQRIVWDQGDUGGHYLDWLRQYDOXHV SHUPLWFRQFOXGHWKDWERWKDYHUDJHDFFXUDF\RI$'/DQGHQYLURQPHQWV
UHFRJQL]HGDUHUHSUHVHQWDWLYH
7DEOH$YHUDJHDFFXUDF\RIWKH$'/UHFRJQL]HG
128
$'/
$YHUDJH$FFXUDF\
6WDQGDUG'HYLDWLRQ
&OHDQLQJ
&RHIILFLHQWRI9DULDWLRQ
&\FOLQJ
'DQFLQJ
'ULYLQJ
)DOOLQJ
*RLQJGRZQWREXV
*RLQJGRZQVWDLUV
*RLQJRQDQHVFDODWRU
*RLQJXSWRDEXV
*RLQJXSVWDLUV
-RJJLQJ
-XPSLQJ
/\LQJ
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
(Table 6) contd.....
$'/
$YHUDJH$FFXUDF\
6WDQGDUG'HYLDWLRQ
3OD\IRRVEDOO
&RHIILFLHQWRI9DULDWLRQ
5XQQLQJ
6LWWLQJ
6WDQGLQJ
6ZHHSLQJ
7DNLQJPHGLFDWLRQ
7UDYHOOLQJE\FDU
7UDYHOOLQJE\WUDLQ
:DONLQJ
:DVKLQJKDQGV
:DWHULQJSODQWV
&RHIILFLHQWRI9DULDWLRQ
7DEOH$YHUDJHDFFXUDF\RIWKHHQYLURQPHQWVUHFRJQL]HG
(QYLURQPHQW
$YHUDJH$FFXUDF\
6WDQGDUG'HYLDWLRQ
$UJXLQJ
%DE\FU\LQJ
%LUGV
%ROOLPSDFW
&KDLQVDZ
&ODSSLQJ
&ORFNWLFNLQJ
&RRNLQJ.LWFKHQ
&URZLQJURRVWHU
'RJV
'RRURSHQLQJFORVLQJ
(PHUJHQF\YHKLFOHV
)LUHFUDFNOLQJ
)RRWVWHSV
*UDLQVIDOOLQJ
+HOLFRSWHU
.QRFNLQJ
/DXJKLQJ
0HWDOLPSDFW
0RWRUF\FOHV
0XVLF
3ODVWLFLPSDFW
5DLQLQJ
5LYHUV
6HDZDYHV
6LJKLQJ
6LQJLQJ
6PRNLQJDODUP
6QHH]LQJ
6SUD\
6WUHHW
7HOHSKRQHULQJLQJ
7UDIILF
7\SLQJ
9RLFH
:DWHU
:KLVWOH
:LQG
:RRGLPSDFW
129
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
3XUSRVHRIWKLV6WXG\
%DVHGRQRXUSUHYLRXVZRUN>@WKHSXUSRVHRIWKLVVWXG\FRQVLVWVLQWKHGHYHORSPHQWRIDQ$QGURLGOLEUDU\
WKDWGHSHQGLQJRQWKHQXPEHUDQGW\SHRIVHQVRUVDYDLODEOHRQERDUGWKHPRELOHGHYLFHLPSOHPHQWVWKHIUDPHZRUN
SUHYLRXVO\GHVLJQHGIRUWKHUHFRJQLWLRQRIWKH$'/DQGWKHLUHQYLURQPHQWVH[SORLWLQJWKHPDMRUQXPEHURIVHQVRUV
DYDLODEOHLQWKHPRELOHGHYLFHXVHG7KHLPSOHPHQWDWLRQVKRXOGXVHOLJKWZHLJKWPHWKRGVLQDPRELOHDSSOLFDWLRQLQ
RUGHUWRJLYHDUDSLGIHHGEDFNWRWKHXVHUZLWKWKHGDWDSURFHVVLQJPHWKRGVLPSOHPHQWHGORFDOO\RQWKHPRELOHGHYLFH
7KHSUHYLRXVZRUNH[SORUHGWKHXVHRIGLIIHUHQWLPSOHPHQWDWLRQVRI$11VXFKDV0/3ZLWK%DFNSURSDJDWLRQ
)11 PHWKRG ZLWK %DFNSURSDJDWLRQ DQG '11 PHWKRG ZLWK QRUPDOL]HG DQG QRQQRUPDOL]HG GDWD 7KH IUDPHZRUN
LPSOHPHQWHGLQWKLVUHVHDUFKLVFRPSRVHGRIVHYHUDOVWDJHVWKHVHEHLQJGDWDDFTXLVLWLRQGDWDSURFHVVLQJGDWDIXVLRQ
DQG FODVVLILFDWLRQ PHWKRGV 7KH GDWD DFTXLVLWLRQ DQG SURFHVVLQJ GHSHQGV RQ WKH W\SHV RI VHQVRUV ZKHUH IRU WKH
DFFHOHURPHWHUJ\URVFRSHDQGPDJQHWRPHWHUGDWDDORZSDVVILOWHULVDSSOLHGIRUWKHDFRXVWLFGDWDWKH)DVW)RXULHU
7UDQVIRUP ))7 IRUWKHH[WUDFWLRQRIWKHUHOHYDQWIUHTXHQFLHVRIWKHDXGLRVLJQDOLVDSSOLHGDQGIRUWKHORFDWLRQGDWDQR
ILOWHUVDUHDSSOLHGDVWKHUDZGDWDDFTXLUHGLVXVHGIRUWKHPHDVXUHPHQWRIWKHGLVWDQFHWUDYHOOHG>@)ROORZLQJWKH
DSSOLFDWLRQRIWKHORZSDVVILOWHUDQGWKH))7WRFOHDQWKHGDWDDFTXLUHGVRPHIHDWXUHVUHODWHGWRWKHVHQVRUVlVLJQDOV
ZHUH H[WUDFWHG DQG IXVHG )LQDOO\ WKH FODVVLILFDWLRQ PHWKRGV ZHUH DSSOLHG IRU WKH UHFRJQLWLRQ RI $'/ DQG
HQYLURQPHQWV
7KH UHFRJQLWLRQ RI $'/ DQG WKHLU HQYLURQPHQWV DUH VHSDUDWHG LQWR WKUHH VWDJHV 7KHVH DUH WKH UHFRJQLWLRQ RI
FRPPRQ$'/ LHZDONLQJJRLQJXSVWDLUVJRLQJGRZQVWDLUVUXQQLQJVWDQGLQJDQGZDWFKLQJ79 WKHUHFRJQLWLRQRI
WKHHQYLURQPHQWV LHEDUFODVVURRPEHGURRPJ\PKDOONLWFKHQOLEUDU\VWUHHWDQGOLYLQJURRP DQGWKHUHFRJQLWLRQ
RI$'/ZLWKRXWERG\PRWLRQ LHVOHHSLQJZDWFKLQJ79DQGGULYLQJ %DVHGRQWKHRXWFRPHVRISUHYLRXVZRUNV
ILUVWO\ IRU WKH UHFRJQLWLRQ RI WKH FRPPRQ $'/ WKH PHWKRG WKDW UHSRUWHG EHWWHU UHVXOWV ZDV '11 PHWKRG ZLWK
QRUPDOL]HG GDWD 6HFRQGO\ IRU WKH UHFRJQLWLRQ RI HQYLURQPHQWV WKH PHWKRG WKDW UHSRUWHG EHWWHU UHVXOWV ZDV )11
PHWKRGZLWKQRQQRUPDOL]HGGDWDEXWZKHQFRPSDUHGZLWKWKHSUHYLRXVUHVXOWVWKLVPHWKRGUHSRUWVDORZSHUIRUPDQFH
DQGDFFXUDF\GXULQJWKHLPSOHPHQWDWLRQRIWKH$QGURLGOLEUDU\ZKHUHWKHUHVXOWVREWDLQHGDUHEHWWHUZLWK'11PHWKRG
ZLWK QRUPDOL]HG GDWD )LQDOO\ IRU WKH UHFRJQLWLRQ RI $'/ ZLWKRXW ERG\ PRWLRQ WKH PHWKRG WKDW UHSRUWHG EHWWHU
DFFXUDF\ZDVDOVR'11PHWKRGZLWKQRUPDOL]HGGDWD7KHGDWDXVHGLQWKLVVWXG\LVDYDLODEOHLQWKH$//DE0HGLD:LNL
>@
6WUXFWXUHRIWKLV6WXG\
7KLVVHFWLRQSUHVHQWVRXUPRWLYDWLRQDQGUHODWHGZRUNIRUWKHGHYHORSPHQWRIDQ$QGURLGOLEUDU\IRUWKHUHFRJQLWLRQ
RI $'/ DQG WKHLU HQYLURQPHQWV 7KH IROORZLQJ VHFWLRQV SUHVHQW WKH PHWKRGV XVHG DQG UHVXOWV REWDLQHG ZLWK WKH
GHYHORSHG$QGURLGOLEUDU\LQFOXGLQJWKHLUGLVFXVVLRQDQGFRPSDULVRQZLWKUHVHDUFKVWXGLHVDYDLODEOHLQWKHOLWHUDWXUH
ZLWKDVLPLODUSXUSRVH
0(7+2'6
6WXG\'HVLJQDQG3DUWLFLSDQWV
2IIWKHVKHOI PRELOH GHYLFHV DUH HTXLSSHG ZLWK VHYHUDO VHQVRUV HJ DFFHOHURPHWHU J\URVFRSH PDJQHWRPHWHU
PLFURSKRQHDQG*36UHFHLYHUZKLFKDOORZWKHDFTXLVLWLRQRIVHYHUDOSK\VLFDODQGSK\VLRORJLFDOSDUDPHWHUV
7KLVVWXG\LVIRFXVHGRQWKHGHYHORSPHQWRIDQ$QGURLGOLEUDU\>@IRUWKHSURFHVVLQJRIWKHVHQVRUVGDWDlDQG
UHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWV7KHPRELOHGHYLFHXVHGIRUWKHWHVWVRIWKH$QGURLGOLEUDU\ZDVD%4
$TXDULXVGHYLFH>@EXWWKHVDPSOLQJIUHTXHQF\RIWKHVHQVRUVlGDWDDFTXLVLWLRQLVGLIILFXOWWRPHDVXUHEHFDXVHWKHVH
GHYLFHVKDYHVHYHUDOFRQVWUDLQWVPRVWO\EHFDXVHDVWKHV\VWHPRSHUDWHVLQDWLPHVKDULQJPDQQHULWLVLPSRVVLEOHWR
DVVXUHDVWHDG\PLFURSURFHVVRUDWWHQWLRQWRWKHVHQVRUUHODWHGSURFHVVHV
6XEMHFWV PDOHVDQGIHPDOHV DJHGEHWZHHQDQG\HDUVROGZHUHVHOHFWHGIRUWKHGDWDDFTXLVLWLRQZLWKD
PRELOHDSSOLFDWLRQWKDWXVHVWKH$QGURLGOLEUDU\GHYHORSHG,QGHSHQGHQWO\RIWKHJHQGHUZHKDYHVHOHFWHGSHRSOHZLWK
GLVWLQFWOLIHVW\OHVZKHUHVXEMHFWVVHOIDVVHVVHGWKHLUOLYLQJVW\OHDVPDLQO\DFWLYHDQGWKHRWKHUVXEMHFWVVHOI
DVVHVVHG WKHLU OLYLQJ VW\OH DV PDLQO\ VHGHQWDU\ 'XULQJ WKH XVH RI WKH PRELOH DSSOLFDWLRQ WKH PRELOH GHYLFH LV
SHUIRUPLQJ RWKHU WDVNV VXFK DV UHFHLYLQJ FDOOV XVLQJ WKH ,QWHUQHW FRQQHFWLRQ DQG XVLQJ RWKHU PRELOH DSSOLFDWLRQV
+RZHYHUWKHPDMRUSDUWRIWKHWLPHWKHPRELOHGHYLFHLVLQWKHIURQWSRFNHWRIWKHXVHUlVSDQWVIRUWKHFRUUHFWGDWD
DFTXLVLWLRQ
130
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
7KHGHYHORSHGPRELOHDSSOLFDWLRQDFTXLUHGWKHVHQVRUVlGDWDDVDEDFNJURXQGVHUYLFHDFTXLULQJDQGSURFHVVLQJVORWV
RIVHFRQGVRIGDWD:LWKRQO\VHFRQGVRIGDWDWKH$'/DQGHQYLURQPHQWVZLOOEHUHFRJQL]HG7KHGDWDDFTXLVLWLRQ
LV SHUIRUPHG HYHU\ PLQXWHV DQG FDSWXUHV IRU HDFK $'/ LH ZDONLQJ UXQQLQJ JRLQJ XSVWDLUV
VWDQGLQJZDWFKLQJ79JRLQJGRZQVWDLUVGULYLQJDQGVOHHSLQJ DQGHQYLURQPHQW LHEHGURRPEDUFODVVURRPJ\P
NLWFKHQOLYLQJURRPKDOOVWUHHWDQGOLEUDU\ ZKLFKFRUUHVSRQGVWRVHFRQGV DSSUR[LPDWHO\KRXUV IRUHDFK
$'/DQGHQYLURQPHQW'XULQJWKHSHUIRUPDQFHRIWKHH[SHULPHQWVWKHXVHUVKRXOGODEHOWKH$'/DQGHQYLURQPHQW
DFTXLUHGEHFDXVHWKHVHGDWDZHUHXVHGWRWUDLQDQGWHVWWKHPHWKRGVIRUWKHFODVVLILFDWLRQZKLFKDUHLPSOHPHQWHGLQWKH
$QGURLGOLEUDU\
'HVFULSWLRQRIWKH)UDPHZRUN
)ROORZLQJWKHSUHYLRXVZRUNWKDWH[SODLQVWKHDUFKLWHFWXUHDQGWKHPHWKRGZLWKEHVWUHVXOWVLQWKHUHFRJQLWLRQRI
$'/ DQG WKHLU HQYLURQPHQWV > @ WKH IUDPHZRUN LPSOHPHQWHG LV D FROOHFWLRQ RI WKH PHWKRGV WKDW
DFFRUGLQJWRWKHOLWHUDWXUHSURYLGHWKHEHVWSHUIRUPDQFHRQUHFRJQLWLRQRI$'/DQGHQYLURQPHQWVVXEMHFWHGWRWKH
DYDLODEOHVHQVRUVLQWKHPRELOHGHYLFHZKHUHWKHIUDPHZRUNLVH[HFXWHG7KXVLWVGHYHORSPHQWLVVHSDUDWHGLQWRVHYHUDO
PRGXOHVVXFKDVGDWDDFTXLVLWLRQGDWDSURFHVVLQJGDWDIXVLRQDQGFODVVLILFDWLRQPHWKRGV
'DWDDFTXLVLWLRQPHWKRGVDUHQRWSURYLGHGE\WKH$QGURLGOLEUDU\SHUIRUPLQJRQO\WKHGDWDSURFHVVLQJGDWDIXVLRQ
DQGFODVVLILFDWLRQPHWKRGV'DWDDFTXLVLWLRQRIWKHDFFHOHURPHWHUJ\URVFRSHPDJQHWRPHWHUPLFURSKRQHDQG*36GDWD
VKRXOGEHSHUIRUPHGZLWKWKHVWDQGDUGPHWKRGVRI$QGURLGSODWIRUPXVLQJDPRELOHDSSOLFDWLRQ)RUWKHUHPDLQLQJ
PRGXOHVWKH$QGURLGOLEUDU\VKRXOGEHDIXQFWLRQRIWKHQXPEHURIVHQVRUVDYDLODEOHLQWKHPRELOHGHYLFH
'DWDSURFHVVLQJPHWKRGVLPSOHPHQWHGLQWKH$QGURLGOLEUDU\FRQVLVWRIWKHGDWDFOHDQLQJDQGIHDWXUHH[WUDFWLRQ
PHWKRGV 7KH GDWD FOHDQLQJ PHWKRG XVHG GHSHQGV RQ WKH W\SH RI VHQVRUV LH IRU DFFHOHURPHWHU J\URVFRSH DQG
PDJQHWRPHWHUGDWDWKHORZSDVVILOWHULVDSSOLHGDQGIRUWKHDFRXVWLFGDWDWKH))7LVDSSOLHGIRUWKHH[WUDFWLRQRIWKH
UHOHYDQWIUHTXHQFLHV7KHIHDWXUHH[WUDFWLRQDOVRGHSHQGVRQWKHW\SHRIVHQVRUVLHIRUDFFHOHURPHWHUJ\URVFRSHDQG
PDJQHWRPHWHUGDWDWKHIHDWXUHVH[WUDFWHGDUHWKHILYHJUHDWHVWGLVWDQFHVEHWZHHQWKHPD[LPXPSHDNVFRPELQHGZLWK
WKH$YHUDJH6WDQGDUG'HYLDWLRQ9DULDQFHDQG0HGLDQRIWKHPD[LPXPSHDNVDQGWKH6WDQGDUG'HYLDWLRQ$YHUDJH
0D[LPXP0LQLPXP9DULDQFHDQG0HGLDQRIWKHUDZVLJQDO)RUWKHDFRXVWLFGDWDWKHIHDWXUHVH[WUDFWHGDUHWKH
0HO)UHTXHQF\&HSVWUDO&RHIILFLHQWV 0)&& FRPELQHGZLWKWKH6WDQGDUG'HYLDWLRQ$YHUDJH0D[LPXP0LQLPXP
9DULDQFHDQG0HGLDQRIWKHUDZVLJQDO)LQDOO\IRUWKHORFDWLRQGDWDWKHIHDWXUHH[WUDFWHGLVWKHGLVWDQFHWUDYHOOHG
$IWHUWKHGDWDIXVLRQRIWKHIHDWXUHVH[WUDFWHGIURPWKHVHQVRUVlGDWDWKHQRUPDOL]DWLRQDQGFODVVLILFDWLRQPHWKRGV
DUHDSSOLHGLQWKUHHVWDJHV
6WDJH
6HQVRUVDFFHOHURPHWHUJ\URVFRSHDQGRUPDJQHWRPHWHU
3XUSRVH5HFRJQLWLRQRIZDONLQJVWDQGLQJUXQQLQJJRLQJXSVWDLUVDQGJRLQJGRZQVWDLUVDFWLYLWLHVXVLQJ
'11PHWKRGZLWKQRUPDOL]HGGDWD
6WDJH
6HQVRUV0LFURSKRQH
3XUSRVH 5HFRJQLWLRQ RI EDU EHGURRP FODVVURRP J\P KDOO NLWFKHQ VWUHHW OLEUDU\ DQG ZDWFKLQJ
79OLYLQJURRPHQYLURQPHQWVXVLQJ)11PHWKRGZLWKQRQQRUPDOL]HGGDWDDQGRU'11PHWKRGZLWK
QRUPDOL]HGGDWD
6WDJH
6HQVRUVDFFHOHURPHWHUJ\URVFRSHPDJQHWRPHWHUDQGRU*36UHFHLYHU
3XUSRVH 5HFRJQLWLRQ RI ZDWFKLQJ 79 VOHHSLQJ DQG GULYLQJ DFWLYLWLHV XVLQJ '11 PHWKRG ZLWK
QRUPDOL]HGGDWD
5HODWHG WR WKH VWDJH WKH SUHYLRXV ZRUN >@ UHFRPPHQGV WKH XVH RI )11 PHWKRG ZLWK QRQQRUPDOL]HG GDWD
+RZHYHU DV LW ZLOO EH SUHVHQWHG LQ WKH IROORZLQJ VHFWLRQV ZH YHULILHG WKDW WKH DFFXUDFLHV FRXOG EH LPSURYHG E\
FRPELQLQJGLIIHUHQWPHWKRGRORJLHV
8VHRI)11PHWKRGZLWKQRQQRUPDOL]HGGDWDDQGZKHQWKHUHVXOWVDUHQRWFOHDUO\LGHQWLILHGWKHXVH'11
PHWKRGZLWKQRUPDOL]HGGDWDIRUWKHVHFDVHVPD\SHUIRUPDEHWWHUUHFRJQLWLRQRIWKHHQYLURQPHQWV )LJ$
8VHRQO\WKH'11PHWKRGZLWKQRUPDOL]HGGDWD )LJ%
131
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
!"#$%&'
3LUHVHWDO
!"#$%&'
!""#$%&'()#*+&'
,(-$./+0%)#).&.
!"#$%&(
!""#$%&'()#*+&'
,(-$./+0%)#).&.
!"#$%&(
2""#$%&'()#*+&'
,(,1,(-$./+0%)#).&.
!"#$%&)
!""#$%&'()#*+&'
,(-$./+0%)#).&.
!"#$%&)
!""#$%&'()#*+&'
,(-$./+0%)#).&.
*+&&,%"-./.0.$1&'
!""#$%&'()#*+&'
,(-$./+0%)#).&.
2+&&,%"-./.0.$1&(
)LJ 6FKHPDRIWKHFODVVLILFDWLRQVWDJHRIWKHIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWV
7KH)11PHWKRGZLWKQRQQRUPDOL]HGGDWDZDVLPSOHPHQWHGWUDLQHGDQGWHVWHGZLWKWKH(QFRJIUDPHZRUN>@LQ
D -DYD GHVNWRS DSSOLFDWLRQ FUHDWLQJ D QHXUDO QHWZRUN IRU IXUWKHU DSSOLFDWLRQ LQ WKH PRELOH DSSOLFDWLRQ 7KH '11
PHWKRGVZLWKQRUPDOL]HGGDWDZHUHLPSOHPHQWHGWUDLQHGDQGWHVWHGZLWK'HHSOHDUQLQJMIUDPHZRUN>@LQD-DYD
GHVNWRSDSSOLFDWLRQFUHDWLQJDQHXUDOQHWZRUNIRUIXUWKHUDSSOLFDWLRQLQWKHPRELOHDSSOLFDWLRQ7KHWUDLQRIWKHQHXUDO
QHWZRUNVQHHGVDVLJQLILFDQWSURFHVVLQJHIIRUWWKDWLVSHUIRUPHGSUHYLRXVO\RQDGHVNWRSFRPSXWHUXVLQJRQO\WKH
PHWKRGGHYHORSHGLQWKHPRELOHDSSOLFDWLRQ7KHUHDUHVHYHUDOSDUDPHWHUVXVHGIRUWKHFUHDWLRQDQGWUDLQLQJRIWKH
QHXUDOQHWZRUNLQFOXGLQJWKHDFWLYDWLRQIXQFWLRQDQGWKHOHDUQLQJUDWHDQGPRPHQWXPYDOXHVXVHGZKLFKDUHSUHVHQWHG
LQ7DEOH7KHVHFRQILJXUDWLRQVPD\LQIOXHQFHWKHVSHHGRIWKHWUDLQLQJVWDJHDVZHOODVWKHUHVXOWVREWDLQHGZLWKWKH
WUDLQHGQHWZRUNILOH
7DEOH&RQILJXUDWLRQVRIWKHFODVVLILFDWLRQPHWKRGV
3DUDPHWHUV
)11PHWKRGZLWK%DFNSURSDJDWLRQ
'110HWKRG
$FWLYDWLRQIXQFWLRQ
6LJPRLG
6LJPRLG
/HDUQLQJUDWH
0RPHQWXP
1$
0D[LPXPQXPEHURIWUDLQLQJLWHUDWLRQV
[
[
1XPEHURIKLGGHQOD\HUV
:HLJKWIXQFWLRQ
1$
;DYLHU
6HHGYDOXH
1$
%DFNSURSDJDWLRQ
<HV
<HV
5HJXODUL]DWLRQ
1$
/
7HVWRIWKH0HWKRG
:HFUHDWHGDVHWRIPHWKRGVWRWHVWWKHFRUUHFWLGHQWLILFDWLRQRIWKH$'/DQGHQYLURQPHQWVXVLQJWKH$QGURLGOLEUDU\
ZHGHYHORSHG)RUWKLVSXUSRVHZHDFTXLUHGWKHVHQVRUVlVLJQDOZLWKDPRELOHGHYLFHUHODWHGWRDURXQGKRXUVIRU
HDFK $'/ DQG HQYLURQPHQWV FRUUHVSRQGHQW WR UHFRUGV ZLWK D OHQJWK RI VHFRQGV ZKLFK DUH ODEHOOHG LQ WKH
PRELOH DSSOLFDWLRQ LQ RUGHU WR DOORZ WKH PHDVXUHPHQW RI WKH DFFXUDF\ RI WKH $QGURLG OLEUDU\ GHYHORSHG :H
132
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
LPSOHPHQWHGXQLWWHVWVWRHYDOXDWHWKHSUHYLRXVO\DFTXLUHGUHFRUGVUHSRUWLQJWKH$'/RUHQYLURQPHQWUHFRJQL]HGZLWK
HDFKUHFRUGLQRUGHUWRPHDVXUHWKHRYHUDOOUHFRJQLWLRQDFFXUDF\ SHUFHQWDJH RIWKH$QGURLGOLEUDU\DVZHOODVWKH
UHFRJQLWLRQDFFXUDF\ SHUFHQWDJH RIHDFK$'/DQGHQYLURQPHQWZLWKWKH$QGURLGOLEUDU\
5(68/76
3URWRW\SH
7KHPDLQUHVXOWRIWKLVVWXG\FRQVLVWVRIWKHGHYHORSPHQWRIDQ$QGURLGOLEUDU\WKDWLPSOHPHQWHGWKHIUDPHZRUNIRU
WKH UHFRJQLWLRQ RI $'/ DQG WKHLU HQYLURQPHQWV ,Q DGGLWLRQ D PRELOH DSSOLFDWLRQ ZLWK D *UDSKLF 8VHU ,QWHUIDFH
SUHVHQWHGLQ)LJ WKDWHPEHGVWKH$QGURLGOLEUDU\GHYHORSHGZDVLPSOHPHQWHGIRULWVYDOLGDWLRQ
>
>
= ? =@ -A
= ? =@ -A
678
99:67
678
23!)4.#"5*$0&"*
99:69
23!)4.#"5*$0&"*
/4BC5B/5C
/D6CD3
!"#$%&'$(")*"+$,
$?0.@)69)1&*A0./)B&0C"A0
/.*/"@/D)($0$)$#EA&/&0&"*
.1
*+,-".(,/01
2/345603,7'89("
:"#)';(<1,:"#)';(<
= ? =@ -A
I%0&1")4.5&/0"
!"#$%#
&'(#)"
>
-./0.)1"1.*0")./0$
678
99:67
K&/0"@&#"
=
>/0$J$
MP
MQ
MR
69
L
M
6
!EFG"8HEG%9E
7
N
O
P
Q
R
L9
LL
LM
L6
L7
LN
LO
LP
LQ
LR
M9
ML
MM
M6
M7
MN
MO
MP
MQ
MR
69
6L
L
M
6
7
N
O
P
!"#$%&'$(")*"+$,
2
!EFG"8HEG%9E
<F.*
2//.//1.*0
H$%.*(.@
.1
I.+,9(F,J<%.E%#(F,B+
;<=4>)-<;
>
<F.*)G%$**&*5
H$%.*($@
3
>
678
99:67
>
= ? =@ -A
S
H@.$0.)*.B
G%$*
= ? =@ -A
= ? =@ -A
678
2(&#&"*$@)F%$*.$1.*0"
G%$*.$1.*0$
<F.*)3$&%U
2//.//1.*0
>
H
99:67
M9LPVLMVP
3$0$)(.%*&#&"
3$0/)(.)S&1
MP
MQ
MR
69
L
M
6
7
N
O
P
Q
R
L9
LL
LM
L6
L7
LN
LO
LP
3.$#.@)>$#$("/
LQ
LR
M9
ML
MM
M6
M7
MN
MO
MP
MQ
MR
69
6L
L
M
6
7
N
O
P
TA$@0"
)LJ 'HVLJQRIWKHLQWHUIDFHRIWKH$QGURLGDSSOLFDWLRQGHYHORSHG)LJXUHVD DQGE SUHVHQWWKHPDLQVFUHHQ)LJXUHF SUHVHQWV
WKHFDOHQGDUZLWKWKHGDLO\SODQQLQJ)LJXUHG SUHVHQWVWKHIRUPIRUWKHFUHDWLRQRIDSODQQLQJUHFRUG)LJXUHH SUHVHQWVDFDOHQGDU
WKDWKLJKOLJKWVWKHGD\VWKDWKDYH$'/DQGHQYLURQPHQWVUHFRJQL]HG)LJXUHI SUHVHQWVDOLVWRI$'/DQGHQYLURQPHQWVSHUIRUPHG
HDFKGD\
7KHPDLQIXQFWLRQDOLW\RIWKHPRELOHDSSOLFDWLRQGHYHORSHGFRQVLVWVLQWKHSUHVHQWDWLRQRIWKHFXUUHQWRUWKHODVW
$'/HQYLURQPHQWRUORFDWLRQUHFRJQL]HG )LJ$DQG% %DVHGRQWKHFDOHQGDUWKHPRELOHDSSOLFDWLRQKDVDPRGXOH
WRPDQDJHWKH$'/DQGHQYLURQPHQWVSODQQHG )LJ&DQG' )LQDOO\WKHPRELOHDSSOLFDWLRQLQFOXGHVDPRGXOHWR
VKRZWKHOLVWRIWKH$'/DQGHQYLURQPHQWVUHFRJQL]HGE\HDFKFDOHQGDUGD\ )LJ) LQFOXGLQJWKHKLJKOLJKWRIWKH
UHVSHFWLYHFDOHQGDUGD\VWKDWLQFOXGHGWKH$'/DQGHQYLURQPHQWVUHFRJQL]HG )LJ(
,QWKHIROORZLQJVHFWLRQVWKHYDOLGDWLRQDQGUHVXOWVRIWKH$QGURLGOLEUDU\HPEHGGHGLQWKHPRELOHDSSOLFDWLRQZLOO
133
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
EHVKRZQSURYLQJWKHUHOLDELOLW\RIWKHIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWV
5HTXLUHPHQWV
7KLV VWXG\ FRQVLVWV RI WKH XVH RI WKH VHQVRUV DYDLODEOH LQ FRQVXPHU PRELOH GHYLFHV 'XH WR WKH GLIIHUHQW
PDQXIDFWXUHUV LQ WKH PDUNHW WKHVH GHYLFHV KDYH D GLIIHUHQW QXPEHU RI VHQVRUV DQG WKH $QGURLG OLEUDU\ GHYHORSHG
VKRXOGEHIOH[LEOHDQGDGDSWDEOHWRWKHQXPEHURIVHQVRUVHIIHFWLYHO\DYDLODEOHDOORZLQJWKHLUXVHZLWKDOOGHYLFHV
DYDLODEOHRQWKHPDUNHW
7RXVHRXU$QGURLGOLEUDU\WKHPRELOHGHYLFHVKRXOGKDYHDPLQLPXPQXPEHURIVHQVRUV LHDFFHOHURPHWHUDQG
PLFURSKRQH EXWGHSHQGLQJRQWKHQXPEHURIVHQVRUVLQFOXGHGLQHDFKPRELOHWKHILQDODFFXUDF\RIWKHUHFRJQLWLRQRI
$'/ DQG WKHLU HQYLURQPHQWV ZLOO EH GLIIHUHQW 7KXV LI WKH PRELOH GHYLFH LQFOXGHV WKH *36 UHFHLYHU WKH PRELOH
DSSOLFDWLRQZLOOLQFUHDVHWKHQXPEHURI$'/FRUUHFWO\UHFRJQL]HG$QGLIWKHPRELOHGHYLFHKDVDQHWZRUNFRQQHFWLRQ
WKHXVHUlVORFDWLRQZLOOEHVKRZQLQWKHPRELOHDSSOLFDWLRQ
,QFRQFOXVLRQGHSHQGLQJRQWKHQXPEHURIVHQVRUVDYDLODEOHLQWKHPRELOHGHYLFHWKH$QGURLGOLEUDU\LVDEOHWR
UHFRJQL]HEHWZHHQDQG$'/DQGHQYLURQPHQWVDVVKRZQLQWKHQH[WVHFWLRQ
9DOLGDWLRQ
2QFHWKHIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQGHQYLURQPHQWVLQWKH$QGURLGOLEUDU\ZDVLPSOHPHQWHGZH
WHVWHGERWKWKHGLIIHUHQWVWDJHVRIWKHIUDPHZRUNLQGHSHQGHQWO\DQGWKHRYHUDOOH[HFXWLRQRIWKHIUDPHZRUN)RUWKHVH
WHVWVZHWRRNLQWRDFFRXQWWKHGLIIHUHQWFRPELQDWLRQVRIVHQVRUVSRVVLEO\DYDLODEOHLQWKHPRELOHGHYLFHV7KHVHWHVWV
ZHUHSHUIRUPHGZLWKWKHGDWDVHWSUHYLRXVO\DFTXLUHGDQGDYDLODEOHDW>@
%DVHGRQWKHLPSOHPHQWDWLRQRIVWDJHIRUWKHUHFRJQLWLRQRIFRPPRQ$'/ LHVWDQGLQJZDONLQJUXQQLQJJRLQJ
XSVWDLUV DQG JRLQJ GRZQVWDLUV XVLQJ WKH '11 PHWKRG ZLWK QRUPDOL]HG GDWD ZH DFKLHYHG DQ DYHUDJH DFFXUDF\ RI
ZLWKWKHDFFHOHURPHWHUVHQVRU 7DEOH ZLWKWKHDFFHOHURPHWHUDQGPDJQHWRPHWHUVHQVRUV 7DEOH
DQGZLWKWKHDFFHOHURPHWHUPDJQHWRPHWHUDQGJ\URVFRSHVHQVRUV 7DEOH
7DEOH$QDO\VLVRIWKHUHVXOWVRIVWDJHZLWKWKHDFFHOHURPHWHUVHQVRU
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV *RLQJ8SVWDLUV 5XQQLQJ 6WDQGLQJ :DONLQJ
$FWXDO&ODVV
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ
:DONLQJ
7DEOH$QDO\VLVRIWKHUHVXOWVRIVWDJHZLWKWKHDFFHOHURPHWHUDQGPDJQHWRPHWHUVHQVRUV
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV *RLQJ8SVWDLUV 5XQQLQJ 6WDQGLQJ :DONLQJ
$FWXDO&ODVV
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ
:DONLQJ
'XULQJ WKH LPSOHPHQWDWLRQ RI VWDJH IRU WKH UHFRJQLWLRQ RI HQYLURQPHQWV LH EHGURRP EDU FODVVURRP J\P
NLWFKHQZDWFKLQJ79OLYLQJURRPKDOOVWUHHWDQGOLEUDU\ ZHYHULILHGWKDWWKH)11PHWKRGZLWK%DFNSURSDJDWLRQ
XVHGLQWKHSUHYLRXVVWXG\>@UHSRUWHGFRQIXVHGUHVXOWV,QIDFWXVLQJWKH)11PHWKRGZLWK%DFNSURSDJDWLRQWKH
IUDPHZRUNUHSRUWVDQDYHUDJHDFFXUDF\RILQWKHUHFRJQLWLRQRIWKHHQYLURQPHQWVUHFRJQL]LQJRQO\RXWRI
HQYLURQPHQWV 7DEOH $VWKHUHVXOWVZHUHQRWJRRGZHFUHDWHGDQHZPHWKRGLQWKHIUDPHZRUNWKDWILUVWWULHVWKH
UHFRJQLWLRQRIWKHHQYLURQPHQWVXVLQJDQ)11PHWKRGZLWK%DFNSURSDJDWLRQDQGZKHQWKHUHVXOWVDUHQRWFOHDUO\
LGHQWLILHGDXWRPDWLFDOO\DSSOLHVWKH'11PHWKRGZLWKQRUPDOL]HGGDWD7KLVZD\ZHREWDLQHGDQDYHUDJHDFFXUDF\RI
ZLWK D VORZLQJ GRZQ SHUIRUPDQFH RI WKH PRELOH GHYLFH UHFRJQL]LQJ RXW RI HQYLURQPHQWV 7DEOH
)LQDOO\ ZH LPSOHPHQWHG WKH '11 PHWKRG RQO\ ZLWK QRUPDOL]HG GDWD IRU WKH UHFRJQLWLRQ RI WKH HQYLURQPHQWV
134
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
LQFUHDVLQJWKHSHUIRUPDQFHRIWKH$QGURLGOLEUDU\DQGUHSRUWLQJDQDYHUDJHDFFXUDF\RIUHFRJQL]LQJDOOWKH
HQYLURQPHQWV 7DEOH ,QWKHQH[WDQDO\VLVUHODWHGWRWKHUHFRJQLWLRQRIWKHHQYLURQPHQWVWKHUHVXOWVREWDLQHGZLWK
WKH'11RQO\PHWKRGZLOOEHWDNHQLQWRDFFRXQW
7DEOH$QDO\VLVRIWKHUHVXOWVRIVWDJHZLWKWKHDFFHOHURPHWHUPDJQHWRPHWHUDQGJ\URVFRSHVHQVRUV
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV *RLQJ8SVWDLUV 5XQQLQJ 6WDQGLQJ :DONLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ
:DONLQJ
$FWXDO&ODVV
7DEOH$QDO\VLVRIWKHUHVXOWVRIVWDJHXVLQJ)11PHWKRGZLWK%DFNSURSDJDWLRQ
b
3UHGLFWHG&ODVV
%HGURRP %DU
&ODVVURRP
*\P +DOO .LWFKHQ /LEUDU\ 6WUHHW
:DWFKLQJ79OLYLQJURRP
%HGURRP
%DU
&ODVVURRP
*\P
+DOO
.LWFKHQ
/LEUDU\
6WUHHW
:DWFKLQJ79OLYLQJURRP
$FWXDO
&ODVV
7DEOH$QDO\VLVRIWKHUHVXOWVRIVWDJHXVLQJ)11PHWKRGZLWK%DFNSURSDJDWLRQDQG'11PHWKRG
b
3UHGLFWHG&ODVV
%HGURRP %DU &ODVVURRP *\P +DOO .LWFKHQ /LEUDU\ 6WUHHW
$FWXDO
&ODVV
:DWFKLQJ79OLYLQJURRP
%HGURRP
%DU
&ODVVURRP
*\P
+DOO
.LWFKHQ
/LEUDU\
6WUHHW
:DWFKLQJ79OLYLQJURRP
%DVHG RQ WKH LPSOHPHQWDWLRQ RI VWDJH IRU WKH UHFRJQLWLRQ RI $'/ ZLWKRXW ERG\ PRWLRQ LH ZDWFKLQJ 79
VOHHSLQJDQGGULYLQJ WKHUHVXOWVREWDLQHGDUHDOZD\VEXWWKHVHUHVXOWVGHSHQGRQWKHUHVXOWVREWDLQHGGXULQJWKH
SUHYLRXVVWDJHVRIWKHIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWVEHFDXVHWKHRXWFRPHVGHSHQGRQ
WKHFRUUHFWUHFRJQLWLRQRIWKHHQYLURQPHQW VWDJH WKHFRUUHFWUHFRJQLWLRQRIWKH$'/DVVWDQGLQJ VWDJH DQGWKH
FRUUHFWH[WUDFWLRQRIWKH*36IHDWXUHV
)LQDOO\ ZH WHVWHG WKH RYHUDOO UHFRJQLWLRQ RI WKH $'/ LH ZDONLQJ UXQQLQJ JRLQJ XSVWDLUV JRLQJ GRZQVWDLUV
ZDWFKLQJ79VOHHSLQJDQGGULYLQJ ZLWKWKHVHTXHQWLDOH[HFXWLRQRIWKHWKUHHVWDJHVXVLQJRQO\WKH'11PHWKRGZLWK
QRUPDOL]HG GDWD UHSRUWLQJ DQ DYHUDJH DFFXUDF\ RI ZLWK WKH XVH RI WKH DFFHOHURPHWHU DQG WKH HQYLURQPHQW
UHFRJQL]HG 7DEOH ZLWKWKHXVHRIWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUDQGWKHHQYLURQPHQWUHFRJQL]HG
7DEOH ZLWKWKHXVHRIWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUWKHJ\URVFRSHDQGWKHHQYLURQPHQWUHFRJQL]HG
7DEOH ZLWKWKHXVHRIWKHDFFHOHURPHWHUWKH*36UHFHLYHUDQGWKHHQYLURQPHQWUHFRJQL]HG 7DEOH
ZLWKWKHXVHRIWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUWKH*36UHFHLYHUDQGWKHHQYLURQPHQWUHFRJQL]HG 7DEOH
135
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
DQGILQDOO\ZLWKWKHXVHRIWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUWKHJ\URVFRSHWKH*36UHFHLYHUDQGWKH
HQYLURQPHQWUHFRJQL]HG 7DEOH
7DEOH$QDO\VLVRIWKHUHVXOWVRIVWDJHXVLQJWKH'11PHWKRG
b
3UHGLFWHG&ODVV
%HGURRP %DU &ODVVURRP *\P +DOO .LWFKHQ /LEUDU\ 6WUHHW
:DWFKLQJ79OLYLQJURRP
%HGURRP
%DU
&ODVVURRP
*\P
+DOO
.LWFKHQ
/LEUDU\
6WUHHW
:DWFKLQJ79OLYLQJURRP
$FWXDO
&ODVV
7DEOH$QDO\VLVRIWKHRYHUDOOUHFRJQLWLRQRIWKH$'/ZLWKDFFHOHURPHWHUDQGHQYLURQPHQWUHFRJQL]HG
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV *RLQJ8SVWDLUV 5XQQLQJ
$FWXDO&ODVV
6WDQGLQJ:DWFKLQJ79
:DONLQJ 6OHHSLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ79
:DONLQJ
6OHHSLQJ
7DEOH$QDO\VLVRIWKHRYHUDOOUHFRJQLWLRQRIWKH$'/ZLWKDFFHOHURPHWHUPDJQHWRPHWHUDQGHQYLURQPHQWUHFRJQL]HG
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV *RLQJ8SVWDLUV 5XQQLQJ
$FWXDO&ODVV
6WDQGLQJ:DWFKLQJ79
:DONLQJ 6OHHSLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ79
:DONLQJ
6OHHSLQJ
7DEOH$QDO\VLVRIWKHRYHUDOOUHFRJQLWLRQRIWKH$'/ZLWKDFFHOHURPHWHUPDJQHWRPHWHUJ\URVFRSHDQGHQYLURQPHQW
UHFRJQL]HG
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV *RLQJ8SVWDLUV 5XQQLQJ
$FWXDO&ODVV
136
6WDQGLQJ:DWFKLQJ79
:DONLQJ 6OHHSLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ79
:DONLQJ
6OHHSLQJ
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
7DEOH$QDO\VLVRIWKHRYHUDOOUHFRJQLWLRQRIWKH$'/ZLWKDQDFFHOHURPHWHU*36UHFHLYHUDQGHQYLURQPHQWUHFRJQL]HG
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV
$FWXDO
&ODVV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ
79
:DONLQJ 6OHHSLQJ 'ULYLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ79
:DONLQJ
6OHHSLQJ
'ULYLQJ
7DEOH $QDO\VLV RI WKH RYHUDOO UHFRJQLWLRQ RI WKH $'/ ZLWK DQ DFFHOHURPHWHU PDJQHWRPHWHU *36 UHFHLYHU DQG
HQYLURQPHQW UHFRJQL]HG
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV
$FWXDO
&ODVV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ
79
:DONLQJ 6OHHSLQJ 'ULYLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ79
:DONLQJ
6OHHSLQJ
'ULYLQJ
7DEOH$QDO\VLVRIWKHRYHUDOOUHFRJQLWLRQRIWKH$'/ZLWKDQDFFHOHURPHWHUPDJQHWRPHWHUJ\URVFRSH*36UHFHLYHUDQG
HQYLURQPHQWUHFRJQL]HG
b
3UHGLFWHG&ODVV
*RLQJ'RZQVWDLUV
$FWXDO
&ODVV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ
79
:DONLQJ 6OHHSLQJ 'ULYLQJ
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ:DWFKLQJ79
:DONLQJ
6OHHSLQJ
'ULYLQJ
',6&866,21
0DLQ)LQGLQJV
%DVHG RQ WKH SDUWLDO UHVXOWV REWDLQHG LQ WKH SUHYLRXV ZRUN > @ ZH SUHVHQW D VXPPDU\ RI WKH DFFXUDFLHV
REWDLQHGLQHDFKVWDJHRIWKHIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWV 7DEOHVWR IRU
IXUWKHUFRPSDULVRQZLWKWKHUHVXOWVREWDLQHGZLWKWKH$QGURLGOLEUDU\GHYHORSHGIRUWKLVVWXG\ 7DEOHVWR ZKLFK
FRPELQHVWKHGLIIHUHQWVWDJHVDOOWRJHWKHU
7DEOH$QDO\VLVRIWKHDFFXUDFLHVUHSRUWHGLQWKHVWXGLHV>@XVLQJWKHDFFHOHURPHWHUVHQVRU
6WDJHV
$FFHOHURPHWHU
$FFHOHURPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0LFURSKRQH
*36
6WDJH
6WDJH
6WDJH
137
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
(Table 21) contd.....
6WDJHV
$FFHOHURPHWHU
$FFHOHURPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0LFURSKRQH
*36
2YHUDOO
7DEOH$QDO\VLVRIWKHDFFXUDFLHVUHSRUWHGLQWKHVWXGLHV>@XVLQJWKHDFFHOHURPHWHUDQGWKHPDJQHWRPHWHUVHQVRUV
6WDJHV
$FFHOHURPHWHU
0DJQHWRPHWHU
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
*36
6WDJH
6WDJH
6WDJH
2YHUDOO
7DEOH$QDO\VLVRIWKHDFFXUDFLHVUHSRUWHGLQWKHVWXGLHV>@XVLQJWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUDQGWKH
J\URVFRSHVHQVRUV
6WDJHV
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
*36
6WDJH
6WDJH
6WDJH
2YHUDOO
7DEOH$QDO\VLVRIWKHDFFXUDFLHVUHSRUWHGE\WKH$QGURLGOLEUDU\XVLQJWKHDFFHOHURPHWHUVHQVRU
6WDJHV
$FFHOHURPHWHU
$FFHOHURPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0LFURSKRQH
*36
6WDJH
6WDJH
6WDJH
2YHUDOO
7DEOH$QDO\VLVRIWKHDFFXUDFLHVUHSRUWHGE\WKH$QGURLGOLEUDU\XVLQJWKHDFFHOHURPHWHUDQGWKHPDJQHWRPHWHUVHQVRUV
6WDJHV
$FFHOHURPHWHU
0DJQHWRPHWHU
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
*36
6WDJH
6WDJH
6WDJH
2YHUDOO
7DEOH$QDO\VLVRIWKHDFFXUDFLHVUHSRUWHGE\WKH$QGURLGOLEUDU\XVLQJWKHDFFHOHURPHWHUWKHPDJQHWRPHWHUDQGWKH
J\URVFRSHVHQVRUV
138
6WDJHV
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
*36
6WDJH
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
(Table 26) contd.....
6WDJHV
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
*36
6WDJH
6WDJH
2YHUDOO
)ROORZLQJ WKH UHVXOWV REWDLQHG LQ WKH VWDJH UHODWHG WR WKH UHFRJQLWLRQ RI WKH FRPPRQ $'/ ZLWK WKH
LPSOHPHQWDWLRQRIWKH'11PHWKRGZLWKQRUPDOL]HGGDWDZHYHULILHGWKDWZLWKWKH$QGURLGOLEUDU\GHYHORSHGWKH
$'/DUHUHFRJQL]HGZLWKPRUHDFFXUDF\ ZLWKWKHDFFHOHURPHWHUVHQVRUZLWKWKHVDPHDFFXUDF\ZLWKWKHMRLQW
DFFHOHURPHWHUDQGPDJQHWRPHWHUVHQVRUVDQGZLWKOHVVDFFXUDF\ ZLWKWKHDFFHOHURPHWHUPDJQHWRPHWHUDQG
J\URVFRSH VHQVRUV WKDQ WKH DFFXUDFLHV REWDLQHG LQ WKH SUHYLRXV ZRUN > @ 7DEOH VKRZV WKH YDOXHV RI WKH
DFFXUDF\ $ WKHVSHFLILFLW\ 6 WKHUHFDOO 5 DQGWKHSUHFLVLRQ 3 RIWKHUHFRJQLWLRQRIHDFK$'/DVZHOODVWKH
DYHUDJHRIWKHVHSDUDPHWHUVLQWKH$QGURLGOLEUDU\GHYHORSHGYHULI\LQJWKDWWKH$'/ZLWKKLJKHVWUHFRJQLWLRQDFFXUDF\
LVWKHVWDQGLQJZLWKDSUHFLVLRQRIDQGDUHFDOORILQFUHDVLQJWKHUHVXOWVLQWKHIUDPHZRUNEHFDXVHLW
LQIOXHQFHVWKHUHVXOWVRIVWDJH
7DEOH5HFRJQLWLRQRIHDFK$'/LQVWDJHPHDQYDOXHV
$'/
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
$FFHOHURPHWHU
0DJQHWRPHWHU
$FFHOHURPHWHU
$
6
5
3
$
6
5
3
$
6
5
3
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ
:DONLQJ
)ROORZLQJWKHUHVXOWVREWDLQHGLQVWDJHUHODWHGWRWKHUHFRJQLWLRQRIWKHHQYLURQPHQWVZHWULHGWRXVHWKHVDPH
PHWKRG LPSOHPHQWHG LQ WKH SUHYLRXV ZRUNV > @ ZKLFK LV WKH )11 PHWKRG ZLWK %DFNSURSDJDWLRQ DQG QRQ
QRUPDOL]HGGDWDEXWLQVWHDGRIWKHRIWKHUHFRJQLWLRQDFFXUDF\WKH$QGURLGOLEUDU\UHSRUWHGDQDFFXUDF\RI
DQG RQO\ HQYLURQPHQWV ZHUH UHFRJQL]HG +RZHYHU GXH WR WKH FDSDELOLWLHV RI WKH PRELOH GHYLFHV WKH
UHFRJQLWLRQREWDLQHGZDVYHU\ORZDQGZHWULHGWRFRPELQHWKHXVHRIWKH)11PHWKRGZLWK%DFNSURSDJDWLRQDQGQRQ
QRUPDOL]HG GDWD ZLWK WKH XVH RI WKH '11 PHWKRG ZLWK QRUPDOL]HG GDWD ZKLFK SURYHV WKDW WKH DFFXUDF\ REWDLQHG
LQFUHDVHVWRUHFRJQL]LQJRIWKHSURSRVHGHQYLURQPHQWV$VZHYHULILHGWKDWWKHXVHRIWKH'11PHWKRG
ZLWKQRUPDOL]HGGDWDLQFUHDVHVWKHUHFRJQLWLRQDFFXUDF\RIWKHHQYLURQPHQWVZHWULHGWRLPSOHPHQWRQO\WKH'11
PHWKRG ZLWK QRUPDOL]HG GDWD UHSRUWLQJ DQ DFFXUDF\ RI UHFRJQL]LQJ DOO RI HQYLURQPHQWV ZLWK EHWWHU
SHUIRUPDQFHEHFDXVHWKHH[HFXWLRQRIWKH'11PHWKRGUHTXLUHVOHVVUHVRXUFHVWKDQWKHH[HFXWLRQRIWKH)11PHWKRG
ZLWK %DFNSURSDJDWLRQ 7DEOH VKRZV WKH YDOXHV RI WKH DFFXUDF\ $ WKH VSHFLILFLW\ 6 WKH UHFDOO 5 DQG WKH
SUHFLVLRQ 3 RIWKHUHFRJQLWLRQRIHDFKHQYLURQPHQWDVZHOODVWKHDYHUDJHRIWKHVHSDUDPHWHUVLQWKH$QGURLGOLEUDU\
GHYHORSHG YHULI\LQJ WKDW WKH UHVXOWV ZLWK WKH KLJKHVW SUHFLVLRQ ZHUH DFKLHYHG LQ WKH VWUHHW DQG WKH EHGURRP
HQYLURQPHQWV DQG WKH KLJKHVW UHFDOO YDOXH LQ WKH UHFRJQLWLRQ RI WKH EHGURRP HQYLURQPHQW
7DEOH5HFRJQLWLRQRIHDFKHQYLURQPHQWLQVWDJHPHDQYDOXHV
(QYLURQPHQWV
0LFURSKRQH
$
6
5
3
%HGURRP
%DU
&ODVVURRP
*\P
+DOO
.LWFKHQ
/LEUDU\
6WUHHW
139
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
(Table 28) contd.....
0LFURSKRQH
(QYLURQPHQWV
:DWFKLQJ79OLYLQJURRP
$
6
5
3
)ROORZLQJWKHUHVXOWVREWDLQHGLQVWDJHUHODWHGWRWKHUHFRJQLWLRQRIWKH$'/ZLWKRXWERG\PRWLRQZLWK$QGURLG
OLEUDU\GHYHORSHGWKH$'/ZLWKRXWERG\PRWLRQLVUHFRJQL]HGZLWKRIUHFRJQLWLRQDFFXUDF\ZKLFKLVWKHVDPH
DFFXUDF\UHSRUWHGE\WKHSUHYLRXVZRUNV>@7DEOHVDQGVKRZWKHYDOXHVRIWKHDFFXUDF\ $ WKHVSHFLILFLW\
6 WKHUHFDOO 5 DQGWKHSUHFLVLRQ 3 RIWKHUHFRJQLWLRQRIHDFK$'/ZLWKRXWERG\PRWLRQDVZHOODVWKHDYHUDJHRI
WKHVHSDUDPHWHUVLQWKH$QGURLGOLEUDU\GHYHORSHGYHULI\LQJWKDWWKHYDOXHVRIWKHDFFXUDF\UHFDOOVSHFLILFLW\DQG
SUHFLVLRQDUHDOZD\V
7DEOH6WDWLVWLFDODQDO\VLVRIWKHUHFRJQLWLRQRIHDFK$'/LQVWDJH
$'/
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0LFURSKRQH
$ 6 5 3 $ 6 5 3 $ 6 5 3
:DWFKLQJ796WDQGLQJ
6OHHSLQJ
'ULYLQJ
7DEOH6WDWLVWLFDODQDO\VLVRIWKHUHFRJQLWLRQRIHDFK$'/LQVWDJH FRQW
$'/
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
*36
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
*36
$FFHOHURPHWHU
0LFURSKRQH
*36
$ 6 5 3 $ 6 5 3 $ 6 5 3
:DWFKLQJ796WDQGLQJ
6OHHSLQJ
'ULYLQJ
)ROORZLQJWKHRYHUDOOUHVXOWVREWDLQHGE\WKHIUDPHZRUNLQWKHSUHYLRXVZRUNV>@IRUWKHPHDVXUHPHQWRIWKH
RYHUDOODFFXUDF\ZHWRRNLQDFFRXQWWKDWZKHQWKH$'/ZDVUHFRJQL]HGDVDQDFWLYLW\ZLWKRXWPRWLRQLWZDVDOZD\V
FRUUHFWO\ FODVVLILHG 7KXV XVLQJ WKH VDPH GDWDVHW IRU WKH WHVWV LQ WKH $QGURLG OLEUDU\ ZH WHVWHG WKH IORZ RI WKH
IUDPHZRUN IRU WKH UHFRJQLWLRQ RI $'/ UHSRUWLQJ DFFXUDFLHV EHWZHHQ DQG ZLWK WKH DFFHOHURPHWHU
VHQVRU EHWZHHQ DQG ZLWK WKH DFFHOHURPHWHU DQG PDJQHWRPHWHU VHQVRUV DQG EHWZHHQ DQG
GHSHQGLQJRQWKHQXPEHURI$'/UHFRJQL]HGWKDWYDULHVEHWZHHQDQG$'/7DEOHVDQGVKRZWKH
YDOXHVRIWKHDFFXUDF\ $ WKHVSHFLILFLW\ 6 WKHUHFDOO 5 DQGWKHSUHFLVLRQ 3 RIWKHUHFRJQLWLRQRIHDFK$'/DV
ZHOO DV WKH DYHUDJH RI WKHVH SDUDPHWHUV LQ WKH $QGURLG OLEUDU\ GHYHORSHG YHULI\LQJ LQ JHQHUDO WKH UHVXOWV REWDLQHG
UHSRUWHG KLJKHVW SUHFLVLRQ DQG VSHFLILFLW\ ZKHUH WKH ORZHU DFFXUDFLHV ZHUH UHSRUWHG IRU WKH JRLQJ XSVWDLUV DQG
GRZQVWDLUVDFWLYLWLHV
7DEOH6WDWLVWLFDODQDO\VLVRIWKHUHFRJQLWLRQRIHDFK$'/LQWKHIUDPHZRUN
$'/
140
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0LFURSKRQH
$
6
5
3
$
6
5
3
$
6
5
3
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ
:DONLQJ
6OHHSLQJ
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
(Table 31) contd.....
$'/
'ULYLQJ
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
$FFHOHURPHWHU
0LFURSKRQH
$
6
5
3
$
6
5
3
$
6
5
3
7DEOH6WDWLVWLFDODQDO\VLVRIWKHUHFRJQLWLRQRIHDFK$'/LQWKHIUDPHZRUN FRQW
$'/
$FFHOHURPHWHU
0LFURSKRQH
*36
$FFHOHURPHWHU
0DJQHWRPHWHU
0LFURSKRQH
*36
$FFHOHURPHWHU
0DJQHWRPHWHU
*\URVFRSH
0LFURSKRQH
*36
$ 6 5 3 $ 6 5 3 $ 6 5 3
*RLQJ'RZQVWDLUV
*RLQJ8SVWDLUV
5XQQLQJ
6WDQGLQJ
:DONLQJ
6OHHSLQJ
'ULYLQJ
2QH RI WKH SRVVLEOH FDXVHV RI WKH YDULDELOLW\ RI WKH UHVXOWV EHWZHHQ WKH SUHYLRXV ZRUN DQG WKH $QGURLG OLEUDU\
GHYHORSHGPD\EHWKHUHVWULFWLRQVLQKDUGZDUHDQGVRIWZDUHYHULILHGLQWKHPRELOHGHYLFHVWKDWZLOOEHDQDO\]HGLQWKH
QH[WVHFWLRQ
/LPLWDWLRQV
'XH WR WKH VRIWZDUH DQG KDUGZDUH OLPLWDWLRQV RI WKH PRELOH GHYLFHV WKHLU XVH IRU WKH UHFRJQLWLRQ RI $'/ DQG
HQYLURQPHQWVKDVVHYHUDOFKDOOHQJHVZKLFKDUHHYLGHQFHGLQVHYHUDOVWXGLHV>@7KHVHDUH
7KHXVHRIVHYHUDOW\SHVRIFRQQHFWLRQ HJ:L)LDQG%OXHWRRWK GHFUHDVHVWKHXVDEOHEDWWHU\WLPH
'XULQJWKHSURFHVVLQJRIWKHVHQVRUVlGDWDWKHSHUIRUPDQFHRIWKHPRELOHGHYLFHPD\GHFUHDVH
'XHWRWKHH[HFXWLRQRIVHYHUDOWDVNVDWWKHVDPHWLPHDQGWKHUHGXFHGUHVRXUFHVRIWKHVHGHYLFHVWKHIUHTXHQF\
RIWKHGDWDDFTXLVLWLRQLVQRWFRQVWDQWRYHUWKHWLPH
7KHSODFHPHQWRIWKHPRELOHGHYLFHLVRQHRIWKHPRVWLPSRUWDQWSUREOHPVEHFDXVHLWLVYHU\LPSRUWDQWIRUWKH
DFTXLVLWLRQRIYDOLGGDWDZKHUHWKHUHDUHQRVWXGLHVSURYLGLQJWKHDEVROXWHRSWLPDOSRVLWLRQWKDWFDQEHJRRGIRU
DQ\W\SHRIUHFRJQLWLRQRQHZDQWVWRSHUIRUPEHFDXVHLWGHSHQGVRQWKHSXUSRVHRIWKHVWXG\
,IWKHSURFHVVLQJRIWKHVHQVRUVlGDWDLVSHUIRUPHGDWDUHPRWHVHUYHUDSHUPDQHQWGDWDFRQQHFWLRQVKRXOGEH
DYDLODEOHWKLVZLOODOVRFRQWULEXWHWRIXUWKHUGHFUHDVHWKHXVDEOHEDWWHU\WLPH
7KH QXPEHU RI VHQVRUV DYDLODEOH LQ WKH FRQVXPHU PRELOH GHYLFHV LV GLIIHUHQW E\ KDUGZDUH YHUVLRQ DQG
PDQXIDFWXUHUVEXWWKHGHYHORSHGPHWKRGVVKRXOGEHDIXQFWLRQRIWKHQXPEHURIVHQVRUVDYDLODEOHRQWKHVH
GHYLFHVPRUHRYHUGLIIHUHQWPDQXIDFWXUHUVPD\UHSRUWGLIIHUHQWHUURUPDUJLQVDQGHYHQGLIIHUHQWRXWSXWYDOXH
UDQJHV
5HJDUGLQJ WKH XVH RI PXOWLSOH GHYLFHV WKH VHFXULW\ SULYDF\ DQG UHVLOLHQFH RI WKH GDWD H[FKDQJHG RYHU WKH
GLIIHUHQWW\SHVRIQHWZRUNFRQQHFWLRQVVKRXOGEHWDNHQLQWRDFFRXQWIRUWKHFRUUHFWXVHRIWKHPRELOHV\VWHPV
)RUWKHPLQLPL]DWLRQRIWKHHIIHFWVRIWKHDIRUHPHQWLRQHGOLPLWDWLRQVVHYHUDOVWUDWHJLHVDUHSURSRVHGLQRXUVWXG\
WKHVHDUH
7KHVHUHVHDUFKVWXGLHVPD\RQO\XVHWKHVHQVRUVDYDLODEOHLQFRQVXPHUPRELOHGHYLFHVUHGXFLQJWKHFRQVWUDLQWV
ZLWKWKHFRQQHFWLRQWRH[WHUQDOVHQVRUVDQGWKHEDWWHU\OLIHWLPH
7KHXVHRIOLJKWZHLJKWPHWKRGVPD\UHGXFHWKHSUREOHPVZLWKWKHUHGXFHGSRZHUSURFHVVLQJRIWKHVHGHYLFHV
$QRWKHUSRVVLEOHVROXWLRQLVWRHQDEOHDQGGLVDEOHWKHDFTXLVLWLRQRIWKHVHQVRUVlGDWDRYHUWKHWLPHEHFDXVHLWLV
QRWQHFHVVDU\WRKDYHDFRQWLQXRXVUHDOWLPHVHQVRUGDWDDFTXLVLWLRQLQRUGHUWRUHFRJQL]HWKH$'/DQGWKHLU
HQYLURQPHQWV
7KH FUHDWLRQ RI PHWKRGV DGDSWHG WR WKH GLIIHUHQW VHWV RI VHQVRUV DYDLODEOH LQ WKH FRQVXPHU PRELOH GHYLFHV
141
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
LQFUHDVHVWKHDYDLODELOLW\RIWKHPHWKRGVIRUDODUJHJURXSRISHRSOH
7KHXVHRIORFDOGDWDSURFHVVLQJUHGXFHVWKHWLPHWRJLYHIHHGEDFNWRWKHXVHUEXWLWDOVRUHGXFHVWKHSUREOHPV
ZLWKWKHVHFXULW\DQGSULYDF\RIWKHSHUVRQDOGDWDGHVSLWHLQFUHDVLQJWKHFRPSXWDWLRQEXUGHQRIWKHGHYLFH
&RPSDULVRQZLWK3ULRU:RUN
%DVHGRQWKHSUHYLRXVUHVHDUFKVWXGLHVDYDLODEOHLQWKHOLWHUDWXUH7DEOHSUHVHQWVWKHFRPSDULVRQEHWZHHQWKH
PLQLPXPPD[LPXPDQGDYHUDJHRIWKHDFFXUDF\DFKLHYHGDQGWKHUHVXOWVUHSRUWHGE\WKH$QGURLGOLEUDU\GHYHORSHG
7KHUHVXOWVWKDWDUHFRPSDUDEOHDUHKLJKOLJKWHGLQ7DEOHYHULI\LQJWKDWWKHKLJKHVWLPSURYHPHQWZDVYHULILHGLQWKH
UXQQLQJDFWLYLW\ZKHQZHFRPSDUHGRXUUHVXOWVZKLWWKHDYHUDJHDFFXUDF\DFKLHYHGLQSULRUZRUN)RUWKDWDFWLYLW\WKH
PLQLPXPDQGWKHPD[LPXPDFFXUDF\IRXQGHGDUHDQGUHVSHFWLYHO\UHWXUQHGDSRVLWLYH DQG
QHJDWLYHLPSURYHPHQW
!""#
$"#
%"#
&"#
'"#
("#
)"#
*"#
F0E/8-0G603>
M:8<96>:A3>
73?0B/01
D8/E/01
;-/01<B-C0>3?/8>
@?A./01
;-/01<:=>3?/8>
9:00/01
,--./012./34560
738663
+"#
HDI
HE68?16<H44:8?4K
J?L/G:0<H44:8?4K
J/0/G:G<H44:8?4K
)LJ &RPSDULVRQEHWZHHQWKHPLQLPXPPD[LPXPDQGDYHUDJHRIWKHDFFXUDFLHVREWDLQHGLQWKHOLWHUDWXUHUHYLHZZLWKWKH
FRPPRQ$'/DQGHQYLURQPHQWVUHFRJQL]HGE\WKH$QGURLGOLEUDU\GHYHORSHG
7DEOH&RPSDULVRQEHWZHHQWKHPLQLPXPPD[LPXPDQGDYHUDJHRIWKHDFFXUDFLHVREWDLQHGLQWKHOLWHUDWXUHUHYLHZDQG
WKHDFFXUDF\RIRXUUHVXOWVLQWKHUHFRJQLWLRQRIWKHVHOHFWHG$'/DQGHQYLURQPHQWV
b
(QYLURQPHQWV
$'/(QYLURQPHQWV
%HGURRP
&RRNLQJNLWFKHQ
$YHUDJH
$FFXUDF\
$YJ
1$
2XU5HVXOWV
25
,PSURYHPHQWV
250LQ 250D[ 25$YJ
*\P
1$
&ODVVURRP
1$
+DOO
1$
:DWFKLQJ79OLYLQJURRP
1$
%DU
1$
/LEUDU\
1$
6WUHHW
142
0LQLPXP
0D[LPXP
$FFXUDF\ 0LQ $FFXUDF\ 0D[
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
(Table 33) contd.....
$'/(QYLURQPHQWV
b
$'/
0LQLPXP
0D[LPXP
$FFXUDF\ 0LQ $FFXUDF\ 0D[
$YHUDJH
$FFXUDF\
$YJ
2XU5HVXOWV
25
,PSURYHPHQWV
250LQ 250D[ 25$YJ
5XQQLQJ
*RLQJXSVWDLUV
:DONLQJ
*RLQJGRZQVWDLUV
'ULYLQJ
6WDQGLQJ
6OHHSLQJ
1$
2EV1$PHDQVWKDWWKH$'/DQGHQYLURQPHQWVZHUHQRWUHFRJQL]HGLQUHVHDUFKVWXGLHVDYDLODEOHLQWKHOLWHUDWXUH
)ROORZLQJ WKH FRPSDUDEOH UHVXOWV SUHVHQWHG LQ )LJ ZH YHULILHG WKDW RXU UHVXOWV DUH DOZD\V KLJKHU WKDQ WKH
DYHUDJHRIWKHUHVXOWVREWDLQHGLQWKHSUHYLRXVUHVHDUFKVWXGLHVLQWKHOLWHUDWXUH
+RZHYHUWKHVLJQLILFDQFHRIWKHLPSURYHPHQWV VHH25$YJLQ7DEOH ZDVHYDOXDWHGZLWKWKH6WXGHQWlVWWHVW
IRUDVDPSOH$VSUHVHQWHGLQ7DEOHWKHQRUPDOLW\DVVXPSWLRQZDVDFKLHYHGDQGLWZDVDVVHVVHGE\WKH6KDSLUR
:LONWHVW 76: SBYDOXH 63666WDWLVWLFV6RIWZDUH Y,%06366&KLFDJR,/ >@ZDVXVHGWR
SHUIRUP WKH VWDWLVWLFDO WHVWV 7KH LPSURYHPHQWV IRU ZKLFK WKH SYDOXH RI WKH WHVW LV ORZHU WKDQ DUH FRQVLGHUHG
VWDWLVWLFDOO\VLJQLILFDQWDWOHYHORIFRQILGHQFH
7DEOH1RUPDOLW\WHVWV
b
.ROPRJRURY6PLUQRY6WDWLVWLFV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH
3UREDELOLW\
6KDSLUR:LON
6WDWLVWLFV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH
3UREDELOLW\
25
$YJ
7KHUHVXOWVRIWKH6WXGHQWlVWWHVWIRUWKHVDPSOHXVHGDUHSUHVHQWHGLQ7DEOHUHSRUWLQJDYDOXHRIWKHDYHUDJHRI
LPSURYHPHQWV HTXDOV WR $FFRUGLQJ WR WKH 6WXGHQWlV WWHVW WKLV YDOXH LV VWDWLVWLFDOO\ VLJQLILFDQW W
SBYDOXH
7DEOH9DOXHVRIWKH6WXGHQW VWWHVWIRUDVDPSOHPHDQIRUFRPSDULQJRXUUHVXOWVDQGDYHUDJHDFFXUDF\
7HVW9DOXH
b
b
WVFRUH
'HJUHHVRI)UHHGRP
6LJQLILFDQFH3UREDELOLW\
%LODWHUDO
25$YJ
$YHUDJHRIWKH'LIIHUHQFH
b
&RQILGHQFH,QWHUYDORI'LIIHUHQFH
/RZHVW
+LJKHVW
,QDGGLWLRQWKHVDPHSURFHGXUHZDVGRQHIRUDQDO\]HWKHLPSURYHPHQWVREWDLQHGZKHQFRPSDUHGRXUUHVXOWVZKLW
WKHPLQLPXPDQGPD[LPXPUHVXOWIRXQGHGLQWKHOLWHUDWXUH VHH250LQDQG250D[LQ7DEOHUHVSHFWLYHO\
,Q7DEOHWKHQRUPDOLW\DVVXPSWLRQZDVDFKLHYHG 6KDSLUR:LONWHVWIRU250LQDQG250D[UHVSHFWLYHO\
76: SBYDOXH DQG 76: SBYDOXH
7DEOH1RUPDOLW\WHVWV
b
.ROPRJRURY6PLUQRY6WDWLVWLFV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH
3UREDELOLW\
6KDSLUR:LON
6WDWLVWLFV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH
3UREDELOLW\
250LQ
250D[
%DVHGLQ7DEOHWKHUHVXOWVRI6WXGHQWlVWHVWVKRZLQRQHKDQGWKDWRXUUHVXOWVDQGWKHPLQLPXPDFFXUDF\IRXQG
LQWKHOLWHUDWXUHDUHVWDWLVWLFDOO\GLIIHUHQW W SBYDOXH DQGWKH\DUHKLJKHU DYHUDJHRIWKHGLIIHUHQFH
EHWZHHQ25DQG0LQLVHTXDO ,QWKHRWKHUKDQGZKHQFRPSDULQJRXUUHVXOWVDQGWKHPD[LPXPDFKLHYHGLQ
OLWHUDWXUHZHFDQDFFHSWWKHQHJDWLYHDYHUDJHRIWKHGLIIHUHQFHEHWZHHQWKHP DYHUDJHRIWKHGLIIHUHQFHEHWZHHQ25
DQG0D[LVHTXDO LVHTXDO]HUR W SBYDOXH
143
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
7DEOH9DOXHVRIWKH6WXGHQW VWWHVWIRUDVDPSOHPHDQIRUFRPSDULQJRXUUHVXOWVZLWKPLQLPXPDQGRXUUHVXOWVZLWK
PD[LPXPDFFXUDF\
7HVW9DOXH
b
b
b
&RQILGHQFH,QWHUYDORI'LIIHUHQFH
WVFRUH
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH3UREDELOLW\
%LODWHUDO
$YHUDJHRIWKH'LIIHUHQFH
/RZHVW
+LJKHVW
250LQ
250D[
)LQDOO\EDVHGRQWKHDFFXUDFLHVSUHVHQWHGLQWKH7DEOHDQGWKHQXPEHURI$'/(QYLURQPHQWVUHFRJQL]HGLQWKH
OLWHUDWXUHSUHVHQWHGLQWKH7DEOHZHFDQFRQFOXGHWKDWRXUPHWKRGLVEHWWHUWKDQRWKHUDYDLODEOHLQWKHOLWHUDWXUH
UHYLHZEHFDXVHRXULPSURYHPHQWLVUHOLDEOHEHFDXVHZHUHFRJQL]HDPDMRUQXPEHURI$'/(QYLURQPHQWVWKDQWKH
RWKHUSUHYLRXVZRUNVDQGZHREWDLQHGVWDWLVWLFDOO\VLJQLILFDQWUHVXOWV
7DEOH 1XPEHU RI $'/(QYLURQPHQWV UHFRJQL]HG E\ WKH VWXGLHV ZKHUH WKH PLQLPXP PD[LPXP DQG DYHUDJH RI WKH
DFFXUDFLHVREWDLQHGLQWKHOLWHUDWXUHUHYLHZZHUHYHULILHG
$'/(QYLURQPHQWV
1XPEHURI$'/
(QYLURQPHQWVLQWKH
6WXG\ZLWK0D[LPXP
$FFXUDF\
1XPEHURI$'/
(QYLURQPHQWVLQWKH
6WXG\ZLWK0LQLPXP
$FFXUDF\
$YHUDJHRIWKHQXPEHURI
$'/(QYLURQPHQWV
5HFRJQL]HGLQWKH
/LWHUDWXUH
&RRNLQJNLWFKHQ
6WUHHW
5XQQLQJ
*RLQJXSVWDLUV
:DONLQJ
*RLQJGRZQVWDLUV
'ULYLQJ
6WDQGLQJ
b
(QYLURQPHQWV
$'/
2XU1XPEHURI
$'/
(QYLURQPHQWV
,QDGGLWLRQZHSHUIRUPHGWKHFRPSDULVRQEHWZHHQRXWQXPEHURI$'/(QYLURQPHQWVUHFRJQL]HGZLWKWKHPHDQ
QXPEHU RI $'/(QYLURQPHQWV UHFRJQL]HG LQ WKH OLWHUDWXUH HYDOXDWLQJ WKH VLJQLILFDQFH RI RXU QXPEHU RI
$'/(QYLURQPHQWV UHFRJQL]HG ZLWK WKH 6WXGHQWlV WWHVW IRU D VDPSOH $V SUHVHQWHG LQ 7DEOH WKH QRUPDOLW\
DVVXPSWLRQ ZDV DFKLHYHG DQG LW ZDV DVVHVVHG E\ WKH 6KDSLUR:LON WHVW 76: SBYDOXH 6366
6WDWLVWLFV6RIWZDUH Y,%06366&KLFDJR,/ >@ZDVXVHGWRSHUIRUPWKHVWDWLVWLFDOWHVWV7KHLPSURYHPHQWVIRU
ZKLFKWKHSYDOXHRIWKHWHVWLVORZHUWKDQDUHFRQVLGHUHGVWDWLVWLFDOO\VLJQLILFDQWDWOHYHORIFRQILGHQFH
7DEOH1RUPDOLW\WHVWV
b
0HDQQXPEHURI$'/
(QYLURQPHQWV
.ROPRJRURY6PLUQRY
6WDWLVWLFV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH
3UREDELOLW\
6KDSLUR:LON
6WDWLVWLFV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH
3UREDELOLW\
7KHUHVXOWVRIWKH6WXGHQWlVWWHVWIRUWKHVDPSOHXVHGDUHSUHVHQWHGLQ7DEOHUHSRUWLQJDYDOXHRIWKHQXPEHURI
$'/(QYLURQPHQWV UHFRJQL]HG HTXDOV WR $FFRUGLQJ WR WKH 6WXGHQWlV WWHVW WKLV YDOXH LV VWDWLVWLFDOO\ VLJQLILFDQW
W SBYDOXH
7DEOH9DOXHVRIWKH6WXGHQW VWWHVWIRUDVDPSOHPHDQIRUFRPSDULQJRXUQXPEHURI$'/(QYLURQPHQWVZLWKWKHDYHUDJH
QXPEHURI$'/(QYLURQPHQWVUHFRJQL]HGLQWKHOLWHUDWXUH
7HVW9DOXH
b
b
WVFRUH
0HDQQXPEHURI$'/
(QYLURQPHQWV
'HJUHHVRI
)UHHGRP
6LJQLILFDQFH3UREDELOLW\
%LODWHUDO
$YHUDJHRIWKH
'LIIHUHQFH
b
&RQILGHQFH,QWHUYDORI'LIIHUHQFH
/RZHVW
+LJKHVW
7KH GHYHORSPHQW RI WKH OLEUDU\ KDV VHYHUDO FKDOOHQJHV LQFOXGLQJ WKH OLPLWHG EDWWHU\ OLIHWLPH DQG WKH UHGXFHG
144
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
SURFHVVLQJSRZHURIWKHPRELOHGHYLFHV>@+RZHYHUWKHDFTXLVLWLRQRIVHFRQGVRIUDZGDWDHYHU\PLQXWHVXVLQJ
WKHROGHVWGHYLFHVWHVWHGKDVVKRZHGDPLQLPXPRIKRXUVRIEDWWHU\OLIHZLWKDQRUPDOXVDJHDQGDVFXUUHQWO\WKH
PRELOHGHYLFHVUHTXLUHGDGDLO\UHFKDUJHLWLVSRVVLEOHWRFRQVLGHUWKDWWKHDFTXLVLWLRQRIWKHVHQVRUVlGDWDZLWKWKLV
PHWKRGFDQEHLPSOHPHQWHG7KHSHUIRUPDQFHRIWKHPRELOHGHYLFHLVRQO\VWULFWO\DIIHFWHGGXULQJDVPDOOWLPHHYHU\
PLQXWHVDQGWKHHIIHFWVLQWKHSHUIRUPDQFHDUHQRWVLJQLILFDQW,WZLOOEHSUREDEO\VXIILFLHQWIRUWKHFKDUDFWHUL]DWLRQRI
OLIHVW\OHV EXW LW PD\ PLVV VRPH LPSRUWDQW HYHQWV WKDW PD\ QRW EH LGHQWLILHG LQFOXGLQJ IDOOLQJ DFWLYLWLHV 7KH
GHYHORSPHQWRIDPHWKRGWKDWLPSOHPHQWVDPRUHVLJQLILFDQWVDPSOLQJVWUDWHJ\ZLWKRXWGHFUHDVLQJWKHSHUIRUPDQFHDQG
WKH DYDLODELOLW\ RI UHVRXUFHV DW WKH PRELOH GHYLFH ZKLOH LGHQWLI\LQJ D ZLGHU QXPEHU RI $'/V UHTXLUHV DGGLWLRQDO
UHVHDUFK
7KHGDWDSURFHVVLQJLVH[HFXWHGDIWHUWKHGDWDDFTXLVLWLRQIRUWKHSUHSDUDWLRQRIWKHGDWDIRUIXUWKHUFODVVLILFDWLRQLQ
WKHODVWVWDJH7KHGDWDSURFHVVLQJZLOOFDOFXODWHWKHIHDWXUHVIURPWKHUDZGDWDDQGDIWHUWKDWLIWKHPLQLPXPUHTXLUHG
GDWD LV DYDLODEOH LH WKH DFFHOHURPHWHU GDWD LV FRUUHFWO\ DFTXLUHG WKH FDOFXODWHG IHDWXUHV ZLOO EH VHQW WR WKH
FODVVLILFDWLRQPRGXOHZKHUHWKHYHULILFDWLRQRIWKH$'/H[HFXWHGLVSHUIRUPHGDQGLQFDVHRIDVLPLODUSDWWHUQLVQRW
DOUHDG\WUDLQHGLQ$11FUHDWHGIRUWKHGHYHORSPHQWRIWKHPHWKRGLQ>@WKHGDWDZLOOEHGLVFDUGHG1HYHUWKHOHVV
WKHGDWDLVRQO\VWRUHGXQWLOWKHFODVVLILFDWLRQDIWHUZKLFKLWZLOOEHDXWRPDWLFDOO\GLVFDUGHGDYRLGLQJSUREOHPVUHODWHG
WRWKHVWRUDJHFDSDFLW\RIWKHVHGHYLFHV
&21&/86,21
7KH DXWRPDWLF UHFRJQLWLRQ RI $'/ DQG WKHLU HQYLURQPHQWV PD\ EH SHUIRUPHG ZLWK WKH VHQVRUV DYDLODEOH LQ
FRQVXPHUPRELOHGHYLFHVLQFOXGLQJWKHDFFHOHURPHWHUWKHJ\URVFRSHWKHPDJQHWRPHWHUWKHPLFURSKRQHDQGWKH*36
UHFHLYHU7KLVSDSHUFRQVLVWVRIWKHFUHDWLRQRIDQ$QGURLGOLEUDU\EDVHGRQDIUDPHZRUNIRUWKHUHFRJQLWLRQRI$'/DQG
WKHLUHQYLURQPHQWVFUHDWHGE\SURSHUO\PHUJLQJPHWKRGVDQGWRROVDYDLODEOHDW>@7KHOLEUDU\GHYHORSHG
VKRXOGEHDIXQFWLRQRIWKHQXPEHURIVHQVRUVDYDLODEOHLQWKHPRELOHGHYLFHVDQGDEOHWRSURYLGHDUDSLGIHHGEDFNWR
WKHXVHUWKDQNVWRWKHORFDOSURFHVVLQJRIWKHVHQVRUVlGDWD
7KH $QGURLG OLEUDU\ LPSOHPHQWHG LQFOXGHV VRPH VWDJHV RI WKH IUDPHZRUN WKDW KDYH EHHQ DOUHDG\ SURSRVHG
LQFOXGLQJ GDWD SURFHVVLQJ GDWD IXVLRQ DQG FODVVLILFDWLRQ PHWKRGV 7KH GDWD DFTXLVLWLRQ VWDJH KDV WR EH SHUIRUPHG
DFFRUGLQJWRWKHPHWKRGVVXSSRUWHGE\WKH$QGURLG2SHUDWLQJ6\VWHPLQHDFKGHYLFH5HODWHGWRWKHGDWDSURFHVVLQJ
SURSHU GDWD FOHDQLQJ PHWKRGV VKRXOG EH DSSOLHG D ORZ SDVV ILOWHU VKRXOG EH DSSOLHG IRU WKH PRWLRQ DQG PDJQHWLF
VHQVRUVlGDWD))7VKRXOGEHDSSOLHGIRUWKHDFRXVWLFVHQVRUVlGDWDDQGFRUUHFWIHDWXUHVVKRXOGEHH[WUDFWHGIRUIXUWKHU
IXVLRQDQGFODVVLILFDWLRQRIWKHVHQVRUVGDWD$IWHUWKHGDWDSURFHVVLQJVWHSWKHIHDWXUHVVKRXOGEHIXVHGQRUPDOL]HG
DQGWKH'11PHWKRGVKRXOGEHDSSOLHG7KHVWUXFWXUHRIWKHIUDPHZRUNIRUWKHFODVVLILFDWLRQLVVHSDUDWHGLQWRWKUHH
VWDJHVWKHUHFRJQLWLRQRIFRPPRQ$'/WKHUHFRJQLWLRQRIHQYLURQPHQWVDQGWKHUHFRJQLWLRQRI$'/ZLWKRXWERG\
PRWLRQ7KLVVWUXFWXUHLVSRVVLEOHEHFDXVHHDFKVWDJHXVHVGLIIHUHQWVHQVRUVlGDWD)LUVWO\WKHUHFRJQLWLRQRIFRPPRQ
$'/ PD\ XVH WKH DFFHOHURPHWHU PDJQHWRPHWHU DQG J\URVFRSH VHQVRUVl GDWD ZLWK WKH '11 PHWKRG UHSRUWLQJ DQ
DYHUDJHDFFXUDF\EHWZHHQDQGIRUWKHUHFRJQLWLRQRIJRLQJXSVWDLUVDQGGRZQVWDLUVZDONLQJUXQQLQJ
DQG VWDQGLQJ 6HFRQGO\ WKH UHFRJQLWLRQ RI WKH HQYLURQPHQWV PD\ XVH WKH PLFURSKRQH GDWD ZLWK WKH '11 PHWKRG
UHSRUWLQJ DQ DFFXUDF\ RI IRU WKH UHFRJQLWLRQ RI GLIIHUHQW VRXQG HQYLURQPHQWV LQFOXGLQJ EHGURRP
FRRNLQJNLWFKHQJ\PFODVVURRPKDOOZDWFKLQJ79OLYLQJURRPEDUOLEUDU\DQGVWUHHW)LQDOO\WKHUHFRJQLWLRQRI
$'/ZLWKRXWERG\PRWLRQPD\EHSHUIRUPHGZLWKWKHHQYLURQPHQWUHFRJQL]HGDQGWKHDFFHOHURPHWHUPDJQHWRPHWHU
J\URVFRSHDQG*36UHFHLYHUVHQVRUVlGDWDUHSRUWLQJDQDFFXUDF\RIIRUWKHUHFRJQLWLRQRIZDWFKLQJ79OLYLQJ
URRPVOHHSLQJGULYLQJDQGRWKHUXQGLIIHUHQWLDWHGVWDQGLQJDFWLYLWLHV,QVXPPDU\WKHRYHUDOODFFXUDF\RIWKH$QGURLG
OLEUDU\ LV EHWZHHQ DQG GHSHQGLQJ RQ WKH QXPEHU RI VHQVRUV XVHG DQG WKH QXPEHU RI $'/ DQG
HQYLURQPHQWVUHFRJQL]HG
7KHUHDUHVHYHUDOOLPLWDWLRQVUHODWHGWRWKHXVHRIPRELOHGHYLFHVWKDWFDQEHPLQLPL]HGZLWKVHYHUDOWHFKQLTXHV
LQFOXGLQJWKHXVHRIOLJKWZHLJKWPHWKRGVDQGORFDOSURFHVVLQJ7KHUHDUHQRRWKHUGDWDVHWVDYDLODEOHWKDWFDQEHXVHGWR
VLPXODWH WKH UHVXOWV REWDLQHG ZLWK WKH IUDPHZRUN ZH GHYHORSHG EHFDXVH WKH UHVXOWV DYDLODEOH LQ WKH OLWHUDWXUH DUH
SURYLGHGVHSDUDWHO\IRU$'/DQGHQYLURQPHQWV6RZHFRPSDUHGWKHDFFXUDFLHVREWDLQHGYHULI\LQJWKDWRXUUHVXOWV
H[KLELWDQRYHUDOOLPSURYHPHQWHTXDOVWRZKHQFRPSDUHGZLWKWKHDYHUDJHRIDFFXUDF\IRXQGLQWKHOLWHUDWXUH
:HDOVRFRPSDUHGRXUUHVXOWVZLWKWKHPLQLPXPDQGWKHPD[LPXPDFFXUDF\DYDLODEOHLQWKHOLWHUDWXUHWKDWSHUPLW
FRQFOXGHVWKDWRXUUHVXOWVGRQRWGLIIHUDWWKHPD[LPXPEXWWKH\DUHVWDWLFDOO\KLJKHUDWWKHPLQLPXPIRXQGLQWKH
OLWHUDWXUH
7RGDWHWKHUHDUHGLIIHUHQWW\SHVRIPRELOHGHYLFHVDYDLODEOHLQWKHPDUNHW7KHUHIRUHWKLVVWXG\ZDVIRFXVHGRQWKH
145
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
LPSOHPHQWDWLRQRIDQ$QGURLGOLEUDU\IRUWKHUHFRJQLWLRQRI$'/DQGWKHLUHQYLURQPHQWVZLWKWKHXVHRIDQ$QGURLG
VPDUWSKRQH7KHVHYHUDOW\SHVRIPRELOHGHYLFHVKDYHGLIIHUHQWDUFKLWHFWXUHVDQGVHQVRUVWKDWPD\UHWXUQGLIIHUHQWW\SHV
RIVHQVRUVlGDWD2XUPHWKRGLVRQO\DSSOLFDEOHZKHQWKHPRELOHGHYLFHLVLQWKHIURQWSRFNHWRIWKHXVHUlVSDQWVWKH
RWKHUSRVLWLRQVPD\UHWXUQLQFRUUHFWUHVXOWV)RUWKHXVHRIGLIIHUHQWGHYLFHV LHVPDUWZDWFK WKHPHWKRGVKRXOGEH
DGDSWHGEHFDXVHWKHVPDUWZDWFKLVQRUPDOO\RQWKHDUPRIWKHXVHUDQGWKHDFTXLUHGSDWWHUQVRIWKHVHQVRUVlGDWDDUH
GLIIHUHQWDFFRUGLQJWRWKHSRVLWLRQDQGWKHKDUGZDUHRIWKHVHGHYLFHV7KHFUHDWLRQRIDOLEUDU\WKDWVKRXOGEHDIXQFWLRQ
RIWKHQXPEHURIVHQVRUVWKHSRVLWLRQLQJRIGHYLFHDQGGLIIHUHQWDUFKLWHFWXUHVH[LVWLQJLQWKHPDUNHWLVDUHVHDUFKWKDW
VKRXOGEHGRQHLQWKHIXWXUH
,QFRQFOXVLRQWKH$QGURLGOLEUDU\GHYHORSHGSURYHVWKHUHOLDELOLW\RIWKHPRELOHGHYLFHVIRUWKHUHFRJQLWLRQRI$'/
DQGWKHLUHQYLURQPHQWVZLWKKLJKSUHFLVLRQ7KHOLEUDU\PD\EHXVHGLQIXWXUHVWXGLHVFRPELQLQJWKHUHFRJQLWLRQRIWKH
$'/DQGWKHLUHQYLURQPHQWVZLWKRWKHUSXUSRVHVIRUWKHGHYHORSPHQWRIDSHUVRQDOGLJLWDOOLIHFRDFK
(7+,&6$33529$/$1'&216(17723$57,&,3$7(
1RWDSSOLFDEOH
+80$1$1'$1,0$/5,*+76
1RDQLPDOVKXPDQVZHUHXVHGIRUVWXGLHVWKDWDUHWKHEDVLVRIWKLVUHVHDUFK
&216(17)2538%/,&$7,21
1RWDSSOLFDEOH
&21)/,&762),17(5(67
7KHDXWKRUVGHFODUHQRFRQIOLFWRILQWHUHVWILQDQFLDORURWKHUZLVH
$&.12:/('*(0(176
7KLVZRUNZDVVXSSRUWHGE\)&7SURMHFW8,'(($ (VWHWUDEDOKRIRLVXSRUWDGRSHORSURMHFWR)&7
8,'(($
7KH DXWKRUV ZRXOG DOVR OLNH WR DFNQRZOHGJH WKH FRQWULEXWLRQ RI WKH &267 $FWLRQ ,& k $$3(/( k
$UFKLWHFWXUHV $OJRULWKPV DQG 3URWRFROV IRU (QKDQFHG /LYLQJ (QYLURQPHQWV
5()(5(1&(6
>@
6DOD]DU/+$/DFHUGD71XQHV-9*UHVVHYRQ:DQJHQKHLP&$V\VWHPDWLFOLWHUDWXUHUHYLHZRQXVDELOLW\KHXULVWLFVIRUPRELOHSKRQHV,QW0RELOH+XP&RPSXW,QWHUDFW
>KWWSG[GRLRUJMPKFL@
>@
)RWL'.RNHWVX-6$FWLYLWLHVRIGDLO\OLYLQJ3HGUHWWLlVRFFXSDWLRQDOWKHUDS\3UDFWLFDOVNLOOVIRUSK\VLFDOG\VIXQFWLRQ
>@
7UDMNRYLN99ODKX*MRUJLHYVND(.RFHVNL6.XOHY,*HQHUDODVVLVWHGOLYLQJV\VWHPDUFKLWHFWXUHPRGHO,QWHUQDWLRQDO&RQIHUHQFHRQ0RELOH
1HWZRUNVDQG0DQDJHPHQW
>@
6LHJHO&+RFKJDWWHUHU$'RUQHU7(&RQWULEXWLRQVRIDPELHQWDVVLVWHGOLYLQJIRUKHDOWKDQGTXDOLW\RIOLIHLQWKHHOGHUO\DQGFDUHVHUYLFHVD
TXDOLWDWLYHDQDO\VLVIURPWKHH[SHUWVlSHUVSHFWLYHRIFDUHVHUYLFHSURIHVVLRQDOV%0&*HULDWU
>KWWSG[GRLRUJ@>30,'@
>@
%RWLD-$9LOOD$3DOPD-$PELHQW$VVLVWHG/LYLQJV\VWHPIRULQKRPHPRQLWRULQJRIKHDOWK\LQGHSHQGHQWHOGHUV([SHUW6\VW$SSO
>KWWSG[GRLRUJMHVZD@
>@
*DUFLD105RGULJXHV--3&(OLDV'&'LDV06$PELHQW$VVLVWHG/LYLQJ7D\ORU )UDQFLV
>@
'LPLWULHYVNL$=GUDYHYVNL(/DPHVNL37UDMNRYLN9$VXUYH\RI$PELHQW$VVLVWHG/LYLQJV\VWHPV&KDOOHQJHVDQGRSSRUWXQLWLHV,Q
,QWHOOLJHQW&RPSXWHU&RPPXQLFDWLRQDQG3URFHVVLQJ ,&&3 ,(((WK,QWHUQDWLRQDO&RQIHUHQFHRQSS
>KWWSG[GRLRUJ,&&3@
>@
6RXVD36DEXJXHLUR')HOL]DUGR9&RXWR53LUHV,*DUFLD1P+HDOWK6HQVRUVDQG$SSOLFDWLRQVIRU3HUVRQDO$LG,Q0RELOH+HDOWK
SS
>KWWSG[GRLRUJB@
>@
=GUDYHYVNL(/DPHVNL37UDMNRYLN9.XODNRY$&KRUEHY,*ROHYD5HWDO,PSURYLQJ$FWLYLW\5HFRJQLWLRQ$FFXUDF\LQ$PELHQW
$VVLVWHG/LYLQJ6\VWHPVE\$XWRPDWHG)HDWXUH(QJLQHHULQJ,((($FFHVV
>KWWSG[GRLRUJ$&&(66@
146
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
>@
&LSSLWHOOL(*DVSDUULQL6*DPEL(6SLQVDQWH6:cKVOdQ\-2UKDQ\,HWDO7LPHV\QFKURQL]DWLRQDQGGDWDIXVLRQIRU5*%GHSWKFDPHUDV
DQGLQHUWLDOVHQVRUVLQ$$/DSSOLFDWLRQV&RPPXQLFDWLRQ:RUNVKRS ,&&:
>KWWSG[GRLRUJ,&&:@
>@
'REUH&0DYURPRXVWDNLV&[*DUFLD1*ROHYD5,0DVWRUDNLV*$PELHQW$VVLVWHG/LYLQJDQG(QKDQFHG/LYLQJ(QYLURQPHQWV3ULQFLSOHV
7HFKQRORJLHVDQG&RQWURO%XWWHUZRUWK+HLQHPDQQ
>@
*ROHYD5,*DUFLD100DYURPRXVWDNLV&;'REUH&0DVWRUDNLV*6WDLQRY5HWDO$$/DQG(/(3ODWIRUP$UFKLWHFWXUH
>KWWSG[GRLRUJ%@
>@
$XJXVWR-&+XFK0.DPHDV$HWDO(GV+DQGERRNRI$PELHQW$VVLVWHG/LYLQJ7HFKQRORJ\IRU+HDOWKFDUH5HKDELOLWDWLRQDQG:HOO
EHLQJ$PVWHUGDP1HWKHUODQGV,263UHVV
>@
*DUFLD10$5RDGPDSWRWKH'HVLJQRID3HUVRQDO'LJLWDO/LIH&RDFK,Q
>KWWSG[GRLRUJB@
>@
.DUD0/DPRXFKL25DPGDQH&KHULI$$4XDOLW\0RGHOIRUWKH(YDOXDWLRQ$$/6\VWHPV
>KWWSG[GRLRUJMSURFV@
>@
6SLQVDQWH6*DPEL(5DIIDHOL/0RQWDQLQL/3DFLHOOR/%HYLODFTXD5HWDO7HFKQRORJ\EDVHGDVVLVWDQFHRISHRSOHZLWKGHPHQWLDVWDWH
RIWKHDUWRSHQFKDOOHQJHVDQGIXWXUHGHYHORSPHQWV+XPDQ0RQLWRULQJ6PDUW+HDOWKDQG$VVLVWHG/LYLQJ7HFKQLTXHVDQG7HFKQRORJLHV
>@
0 -DQXUD\ QG 6PDUWSKRQH 26 PDUNHW VKDUH IRUHFDVW _ 6WDWLVWLF $YDLODEOH KWWSVZZZVWDWLVWDFRP
VWDWLVWLFVPDUNHWVKDUHIRUHFDVWIRUVPDUWSKRQHRSHUDWLQJV\VWHPV
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD))URP'DWD$FTXLVLWLRQWR'DWD)XVLRQ$&RPSUHKHQVLYH5HYLHZDQGD5RDGPDSIRUWKH
,GHQWLILFDWLRQRI$FWLYLWLHVRI'DLO\/LYLQJ8VLQJ0RELOH'HYLFHV6HQVRUV %DVHO
>KWWSG[GRLRUJV@>30,'@
>@
3LUHV,0*DUFLD10)OhUH]5HYXHOWD)0XOWLVHQVRUGDWDIXVLRQWHFKQLTXHVIRUWKHLGHQWLILFDWLRQRIDFWLYLWLHVRIGDLO\OLYLQJXVLQJPRELOH
GHYLFHV3URFHHGLQJVRIWKH(&0/3.'''RFWRUDO&RQVRUWLXP(XURSHDQ&RQIHUHQFHRQ0DFKLQH/HDUQLQJDQG3ULQFLSOHVDQG3UDFWLFH
RI.QRZOHGJH'LVFRYHU\LQ'DWDEDVHV3RUWR3RUWXJDO
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD),GHQWLILFDWLRQRI$FWLYLWLHVRI'DLO\/LYLQJ8VLQJ6HQVRUV$YDLODEOHLQRIIWKHVKHOI
0RELOH'HYLFHV5HVHDUFKDQG+\SRWKHVLV
>KWWSG[GRLRUJB@
>@
,YDVFX7&LQFDU.'LQLV$1HJUX9$FWLYLWLHVRIGDLO\OLYLQJDQGIDOOVUHFRJQLWLRQDQGFODVVLILFDWLRQIURPWKHZHDUDEOHVHQVRUVGDWD,Q
>KWWSG[GRLRUJ(+%@
>@
.LOLQF2'DO]HOO$8OXWXUN,8\VDO,,QHUWLD%DVHG5HFRJQLWLRQRI'DLO\$FWLYLWLHVZLWK$11VDQG6SHFWURWHPSRUDO)HDWXUHV
>KWWSG[GRLRUJ,&0/$@
>@
3UDERZR200XWLMDUVD.6XSDQJNDW6+0LVVLQJGDWDKDQGOLQJXVLQJPDFKLQHOHDUQLQJIRUKXPDQDFWLYLW\UHFRJQLWLRQRQPRELOHGHYLFH
,QWHUQDWLRQDO&RQIHUHQFHRQ,&7)RU6PDUW6RFLHW\ ,&,66
>KWWSG[GRLRUJ,&766@
>@
%XMDUL $ /LFDU % 3DOD]]L &( 0RYHPHQW SDWWHUQ UHFRJQLWLRQ WKURXJK VPDUWSKRQHlV DFFHOHURPHWHU &RQVXPHU &RPPXQLFDWLRQV DQG
1HWZRUNLQJ&RQIHUHQFH &&1&
>KWWSG[GRLRUJ&&1&@
>@
9DOODEK30DOHNLDQ5<H1%RJDWLQRVND'&)DOOGHWHFWLRQXVLQJPDFKLQHOHDUQLQJDOJRULWKPV
>KWWSG[GRLRUJ62)7&20@
>@
3RPER1*DUFLD1%RXVVRQ.&ODVVLILFDWLRQWHFKQLTXHVRQFRPSXWHUL]HGV\VWHPVWRSUHGLFWDQGRUWRGHWHFW$SQHD$V\VWHPDWLFUHYLHZ
&RPSXW0HWKRGV3URJUDPV%LRPHG
>KWWSG[GRLRUJMFPSE@>30,'@
>@
'UXFNHU+&RUWHV&-DFNHO/'/H&XQ<9DSQLN9%RRVWLQJDQG2WKHU(QVHPEOH0HWKRGV1HXUDO&RPSXW
>KWWSG[GRLRUJQHFR@
>@
/RUHQ]L35DR55RPDQR*.LWD$,UUHUD)0RELOH'HYLFHVIRUWKH5HDO7LPH'HWHFWLRQRI6SHFLILF+XPDQ0RWLRQ'LVRUGHUV,(((6HQV
-
>KWWSG[GRLRUJ-VHQ@
>@
%D\DW$3RPSOXQ07UDQ'$$6WXG\RQ+XPDQ$FWLYLW\5HFRJQLWLRQ8VLQJ$FFHOHURPHWHU'DWDIURP6PDUWSKRQHVWK,QWHUQDWLRQDO
&RQIHUHQFH RQ )XWXUH 1HWZRUNV DQG &RPPXQLFDWLRQV )QFl WKH WK ,QWHUQDWLRQDO &RQIHUHQFH RQ 0RELOH 6\VWHPV DQG 3HUYDVLYH
&RPSXWLQJ 0RELVSFl $IILOLDWHG:RUNVKRSVYRO
>KWWSG[GRLRUJMSURFV@
>@
.KDOLID6+DVVDQ06HQHYLUDWQH$)HDWXUHVHOHFWLRQIRUIORRUFKDQJLQJDFWLYLW\UHFRJQLWLRQLQPXOWLIORRUSHGHVWULDQQDYLJDWLRQ
>KWWSG[GRLRUJ,&08@
>@
3L\DUH5/HH650RELOH6HQVLQJ3ODWIRUPIRU3HUVRQDO+HDOWK0DQDJHPHQWWK ,HHH,QWHUQDWLRQDO6\PSRVLXPRQ&RQVXPHU(OHFWURQLFV
,VFH
147
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
>KWWSG[GRLRUJ,6&(@
>@
.ZRQ<.DQJ.%DH&$QDO\VLVDQG(YDOXDWLRQRI6PDUWSKRQHEDVHG+XPDQ$FWLYLW\5HFRJQLWLRQ8VLQJD1HXUDO1HWZRUN$SSURDFK
,QWHUQDWLRQDO-RLQW&RQIHUHQFHRQ1HXUDO1HWZRUNV ,-&11
>KWWSG[GRLRUJ,-&11@
>@
:DQJ&;X<=KDQJ-<X:6:+00$0HWKRGIRU(YDOXDWLQJ&RQILGHQFHRI6PDUWSKRQH%DVHG$FWLYLW\5HFRJQLWLRQ,Q
>KWWSG[GRLRUJ7UXVW&RP@
>@
+VX<:&KHQ.+<DQJ---DZ)66PDUWSKRQHEDVHGIDOOGHWHFWLRQDOJRULWKPXVLQJIHDWXUHH[WUDFWLRQ
>KWWSG[GRLRUJ&,63%0(,@
>@
9DYRXODV*&KDW]DNL&0DOOLRWDNLV73HGLDGLWLV07VLNQDNLV07KH0REL$FW'DWDVHW5HFRJQLWLRQRI$FWLYLWLHVRI'DLO\/LYLQJXVLQJ
6PDUWSKRQHV3URFHHGLQJVRIWKH,QWHUQDWLRQDO&RQIHUHQFHRQ,QIRUPDWLRQDQG&RPPXQLFDWLRQ7HFKQRORJLHVIRU$JHLQJ:HOODQG(+HDOWK
,FWDZH
>KWWSG[GRLRUJ@
>@
%XMDUL$/LFDU%3DOD]]L&(0RYHPHQW3DWWHUQ5HFRJQLWLRQWKURXJK6PDUWSKRQHlV$FFHOHURPHWHU,(((&RQVXPHU&RPPXQLFDWLRQV
DQG1HWZRUNLQJ&RQIHUHQFH&FQFSS
>@
7RUUHV+XLW]LO&1XQR0DJDQGD05REXVWVPDUWSKRQHEDVHGKXPDQDFWLYLW\UHFRJQLWLRQXVLQJDWULD[LDODFFHOHURPHWHU
>KWWSG[GRLRUJ/$6&$6@
>@
%DMSDL$-LOOD97LZDUL919HQNDWHVDQ601DUD\DQDQ54XDQWLILDEOHILWQHVVWUDFNLQJXVLQJZHDUDEOHGHYLFHV&RQI3URF,((((QJ0HG
%LRO6RF
>KWWSG[GRLRUJ(0%&@>30,'@
>@
=DLQXGLQ 016 6XODLPDQ 01 0XVWDSKD 1 3HUXPDO 7 $FWLYLW\ 5HFRJQLWLRQ EDVHG RQ $FFHOHURPHWHU 6HQVRU XVLQJ &RPELQDWLRQDO
&ODVVLILHUV,(((&RQIHUHQFHRQ2SHQ6\VWHPV ,FRV
>KWWSG[GRLRUJ,&26@
>@
6HQ65DFKXUL..0XNKHUML$0LVUD$'LG\RXWDNHDEUHDNWRGD\"'HWHFWLQJSOD\LQJIRRVEDOOXVLQJ\RXUVPDUWZDWFK
>KWWSG[GRLRUJ3(5&20:@
>@
.ZDSLV]-5:HLVV*00RRUH6$$FWLYLW\UHFRJQLWLRQXVLQJFHOOSKRQHDFFHOHURPHWHUV6,*.''([SORU
>KWWSG[GRLRUJ@
>@
0LWFKHOO(0RQDJKDQ'2l&RQQRU1(&ODVVLILFDWLRQRIVSRUWLQJDFWLYLWLHVXVLQJVPDUWSKRQHDFFHOHURPHWHUV6HQVRUV %DVHO
>KWWSG[GRLRUJV@>30,'@
>@
/L3:DQJ<7LDQ<=KRX76/L-6$Q$XWRPDWLF8VHU$GDSWHG3K\VLFDO$FWLYLW\&ODVVLILFDWLRQ0HWKRG8VLQJ6PDUWSKRQHV,(((7UDQV
%LRPHG(QJ
>KWWSG[GRLRUJ7%0(@>30,'@
>@
=KDQJ/:X;/XR'5HDO7LPH$FWLYLW\5HFRJQLWLRQRQ6PDUWSKRQHV8VLQJ'HHS1HXUDO1HWZRUNV
>KWWSG[GRLRUJ8,&$7&6FDO&RP&%'&RP,R3@
>@
=KDR./'X-=/L&4=KDQJ&//LX+;X&+HDOWK\$'LDU\6\VWHP%DVHGRQ$FWLYLW\5HFRJQLWLRQ8VLQJ6PDUWSKRQH
>KWWSG[GRLRUJ0$66@
>@
'HUQEDFK6'DV%.ULVKQDQ1&7KRPDV%/&RRN'-6LPSOHDQG&RPSOH[$FWLYLW\5HFRJQLWLRQWKURXJK6PDUW3KRQHV
>KWWSG[GRLRUJ,(@
>@
6KHQ&&KHQ<)<DQJ*62Q0RWLRQ6HQVRU%HKDYLRU$QDO\VLVIRU+XPDQ$FWLYLW\5HFRJQLWLRQYLD6PDUWSKRQHV,HHH,QWHUQDWLRQDO
&RQIHUHQFHRQ,GHQWLW\6HFXULW\DQG%HKDYLRU$QDO\VLV ,VED 6HQGDL-DSDQSS
>KWWSG[GRLRUJ,6%$@
>@
9DOODEK30DOHNLDQ5<H1%RJDWLQRVND'&)DOO'HWHFWLRQ8VLQJ0DFKLQH/HDUQLQJ$OJRULWKPV
>KWWSG[GRLRUJ62)7&20@
>@
:X:'DVJXSWD65DPLUH]((3HWHUVRQ&1RUPDQ*-&ODVVLILFDWLRQDFFXUDFLHVRISK\VLFDODFWLYLWLHVXVLQJVPDUWSKRQHPRWLRQVHQVRUV0HG,QWHUQHW5HV H
>KWWSG[GRLRUJMPLU@>30,'@
>@
)DQJ / <LVKXL 6 :HL & 8S DQG GRZQ EXVHV DFWLYLW\ UHFRJQLWLRQ XVLQJ VPDUWSKRQH DFFHOHURPHWHU (OHFWURQLF DQG $XWRPDWLRQ &RQWURO
&RQIHUHQFH
>KWWSG[GRLRUJ,71(&@
>@
=GUDYHYVNL(/DPHVNL30LQJRY5.XODNRY$*MRUJMHYLNM'5REXVWKLVWRJUDPEDVHGIHDWXUHHQJLQHHULQJRIWLPHVHULHVGDWD
>KWWSG[GRLRUJ)@
>@
=GUDYHYVNL ( 5LVWHVND 6WRMNRVND % 6WDQGO 0 6FKXO] + $XWRPDWLF PDFKLQHOHDUQLQJ EDVHG LGHQWLILFDWLRQ RI MRJJLQJ SHULRGV IURP
DFFHOHURPHWHUPHDVXUHPHQWVRIDGROHVFHQWVXQGHUILHOGFRQGLWLRQV3/R62QH H
>KWWSG[GRLRUJMRXUQDOSRQH@>30,'@
>@
7DQJ&;3KRKD99$Q(PSLULFDO(YDOXDWLRQRI$FWLYLWLHVDQG&ODVVLILHUVIRU8VHU,GHQWLILFDWLRQRQ6PDUWSKRQHV
>KWWSG[GRLRUJ%7$6@
148
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
$QGURLG/LEUDU\IRU5HFRJQLWLRQRI$FWLYLWLHV
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
>@
%ORFK$(UGLQ50H\HU6.HOOHU7GH6SLQGOHU$%DWWHU\(IILFLHQW7UDQVSRUWDWLRQ0RGH'HWHFWLRQRQ0RELOH'HYLFHV
>KWWSG[GRLRUJ0'0@
>@
6KRDLE06FKROWHQ++DYLQJD3-07RZDUGV3K\VLFDO$FWLYLW\5HFRJQLWLRQ8VLQJ6PDUWSKRQH6HQVRUV
>KWWSG[GRLRUJ8,&$7&@
>@
0LHONH 0 %UiFN 5 6PDUWSKRQH DSSOLFDWLRQ IRU DXWRPDWLF FODVVLILFDWLRQ RI HQYLURQPHQWDO VRXQG 3URFHHGLQJV RI WKH WK ,QWHUQDWLRQDO
&RQIHUHQFH0L[HG'HVLJQRI,QWHJUDWHG&LUFXLWVDQG6\VWHPV0,;'(6
>@
*XR;7R\RGD</L++XDQJ-'LQJ6/LX<(QYLURQPHQWDOVRXQGUHFRJQLWLRQXVLQJWLPHIUHTXHQF\LQWHUVHFWLRQSDWWHUQV
>@
3LOORV$$OJKDPLGL.$O]DPHO13DYORY90DFKDQDYDMKDOD6$UHDOWLPHHQYLURQPHQWDOVRXQGUHFRJQLWLRQV\VWHPIRUWKH$QGURLG26
>@
%RXQWRXUDNLV99U\VLV/3DSDQLNRODRX*0DFKLQH/HDUQLQJ$OJRULWKPVIRU(QYLURQPHQWDO6RXQG5HFRJQLWLRQ7RZDUGV6RXQGVFDSH
6HPDQWLFV3URFHHGLQJVRIWKH$XGLR0RVWO\RQ,QWHUDFWLRQ:LWK6RXQG7KHVVDORQLNL*UHHFH
>KWWSG[GRLRUJ@
>@
0LHONH0%UXHFN5'HVLJQDQGHYDOXDWLRQRIDVPDUWSKRQHDSSOLFDWLRQIRUQRQVSHHFKVRXQGDZDUHQHVVIRUSHRSOHZLWKKHDULQJORVV&RQI
3URF,((((QJ0HG%LRO6RF
>KWWSG[GRLRUJ(0%&@
>@
6LJWLD66WDUN$0.UVWXORYLF63OXPEOH\0'$XWRPDWLF(QYLURQPHQWDO6RXQG5HFRJQLWLRQ3HUIRUPDQFH9HUVXV&RPSXWDWLRQDO&RVW
,((($&07UDQV$XGLR6SHHFK/DQJ3URFHVV
>KWWSG[GRLRUJ7$6/3@
>@
'XEH\+0HKO050DQNRGL\D.%LJ($5,QIHUULQJWKHDPELHQWDQGHPRWLRQDOFRUUHODWHVIURPVPDUWSKRQHEDVHGDFRXVWLFELJGDWD
VW
,((( ,QWHUQDWLRQDO&RQIHUHQFHRQ&RQQHFWHG+HDOWK$SSOLFDWLRQV6\VWHPVDQG(QJLQHHULQJ7HFKQRORJLHV &+$6(
>KWWSG[GRLRUJ&+$6(@
>@
/DQH1'*HRUJLHY34HQGUR/'HHS(DUUREXVWVPDUWSKRQHDXGLRVHQVLQJLQXQFRQVWUDLQHGDFRXVWLFHQYLURQPHQWVXVLQJGHHSOHDUQLQJ
3URFHHGLQJVRIWKH$&0,QWHUQDWLRQDO-RLQW&RQIHUHQFHRQ3HUYDVLYHDQG8ELTXLWRXV&RPSXWLQJ2VDND-DSDQ
>KWWSG[GRLRUJ@
>@
&KHIIHQD0)DOO'HWHFWLRQ8VLQJ6PDUWSKRQH$XGLR)HDWXUHV,(((-%LRPHG+HDOWK,QIRUP
>KWWSG[GRLRUJ-%+,@>30,'@
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD)8VHU(QYLURQPHQW'HWHFWLRQZLWK$FRXVWLF6HQVRUV(PEHGGHGRQ0RELOH'HYLFHVIRUWKH
5HFRJQLWLRQRI$FWLYLWLHVRI'DLO\/LYLQJ,Q5HYLHZ
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD)$0XOWLSOH6RXUFH)UDPHZRUNIRUWKH,GHQWLILFDWLRQRI$FWLYLWLHVRI'DLO\/LYLQJ%DVHG
RQ0RELOH'HYLFH'DWD,Q5HYLHZ
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD)6SLQVDQWH63DWWHUQ5HFRJQLWLRQ7HFKQLTXHVIRUWKH,GHQWLILFDWLRQRI$FWLYLWLHVRI'DLO\
/LYLQJXVLQJ0RELOH'HYLFH$FFHOHURPHWHU,Q5HYLHZ
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD)6SLQVDQWH6'DWD)XVLRQRQ0RWLRQDQG0DJQHWLF6HQVRUVHPEHGGHGRQ0RELOH'HYLFHV
IRUWKH,GHQWLILFDWLRQRI$FWLYLWLHVRI'DLO\/LYLQJ,QHQJU[LYRUJ[U],Q5HYLHZ
>@
3LUHV,0*DUFLD103RPER1)OhUH]5HYXHOWD)5RGUfJXH]1'9DOLGDWLRQ7HFKQLTXHVIRU6HQVRU'DWDLQ0RELOH+HDOWK$SSOLFDWLRQV6HQV
>KWWSG[GRLRUJ@
>@
$XJXVW 0XOWLVHQVRU GDWD IXVLRQ LQ PRELOH GHYLFHV IRU WKH LGHQWLILFDWLRQ RI DFWLYLWLHV RI GDLO\ OLYLQJ $YDLODEOH IURP
KWWSVDOODEGLXELSW
PHGLDZLNLLQGH[SKS$XJXVWBB
0XOWL
VHQVRUBGDWDBIXVLRQBLQBPRELOHBGHYLFHVBIRUBWKHBLGHQWLILFDWLRQBRIBDFWLYLWLHVBRIBGDLO\BOLYLQJ
>@
%RMLQRY+0LFKDOHYVN\<1DNLEO\*%RQHK'0RELOHGHYLFHLGHQWLILFDWLRQYLDVHQVRUILQJHUSULQWLQJDU;LYSUHSULQWDU;LY
>@
.DWHYDV.+DGGDGL+7RNDUFKXN/6HQVLQJNLW(YDOXDWLQJWKHVHQVRUSRZHUFRQVXPSWLRQLQL26GHYLFHVDU[LY
>KWWSG[GRLRUJ,(@
>@
6PDUWSKRQHV$YDLODEOHIURPKWWSVZZZETFRPSWVPDUWSKRQHV
>@
(QFRJ0DFKLQH/HDUQLQJ)UDPHZRUN$YDLODEOHIURPKWWSZZZKHDWRQUHVHDUFKFRPHQFRJ
>@
'HHSOHDUQLQJM2SHQVRXUFH'LVWULEXWHG'HHS/HDUQLQJIRUWKH-90$YDLODEOHIURPKWWSVGHHSOHDUQLQJMRUJ
>@
6DQWRV-5RGULJXHV--3&6LOYD%0&&DVDO-6DOHHP.'HQLVRY9$Q,R7EDVHGPRELOHJDWHZD\IRULQWHOOLJHQWSHUVRQDODVVLVWDQWVRQ
PRELOHKHDOWKHQYLURQPHQWV-1HWZ&RPSXW$SSO
>KWWSG[GRLRUJMMQFD@
>@
%HUW)*LDFRPHWWL0*XDODQR056LOLTXLQL56PDUWSKRQHVDQGKHDOWKSURPRWLRQDUHYLHZRIWKHHYLGHQFH-0HG6\VW
>KWWSG[GRLRUJV@>30,'@
>@
3RVODG68ELTXLWRXVFRPSXWLQJVPDUWGHYLFHVHQYLURQPHQWVDQGLQWHUDFWLRQV-RKQ:LOH\ 6RQV
>@
&KRXGKXU\7%RUULHOOR*&RQVROYR6+DHKQHO'+DUULVRQ%+HPLQJZD\%HWDO7KH0RELOH6HQVLQJ3ODWIRUP$Q(PEHGGHG$FWLYLW\
5HFRJQLWLRQ6\VWHP,(((3HUYDVLYH&RPSXW
149
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
7KH2SHQ%LRLQIRUPDWLFV-RXUQDO9ROXPH
3LUHVHWDO
>KWWSG[GRLRUJ0359@
>@
0RQWRUR0DQULTXH*+D\D&ROO36FKQHOOH:DOND',QWHUQHWRI7KLQJV)URP5),'6\VWHPVWR6PDUW$SSOLFDWLRQV
>@
$ULI0-(O(PDU\,0.RXWVRXULV''$UHYLHZRQWKHWHFKQRORJLHVDQGVHUYLFHVXVHGLQWKHVHOIPDQDJHPHQWRIKHDOWKDQGLQGHSHQGHQW
OLYLQJRIHOGHUO\7HFKQRO+HDOWK&DUH
>KWWSG[GRLRUJ7+&@>30,'@
>@
)LHOG$'LVFRYHULQJVWDWLVWLFVXVLQJ,%06366VWDWLVWLFV
j3LUHVHWDO
7KLVLVDQRSHQDFFHVVDUWLFOHGLVWULEXWHGXQGHUWKHWHUPVRIWKH&UHDWLYH&RPPRQV$WWULEXWLRQ,QWHUQDWLRQDO3XEOLF/LFHQVH &&%< D
FRS\RIZKLFKLVDYDLODEOHDW KWWSVFUHDWLYHFRPPRQVRUJOLFHQVHVE\OHJDOFRGH 7KLVOLFHQVHSHUPLWVXQUHVWULFWHGXVHGLVWULEXWLRQDQG
UHSURGXFWLRQLQDQ\PHGLXPSURYLGHGWKHRULJLQDODXWKRUDQGVRXUFHDUHFUHGLWHG
150
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
Chapter 5
Conclusions and Future Work
This chapter presents the main conclusions that result from the research work described in this
Thesis. Furthermore, it discusses research topics related with the work developed in the
doctoral programme, which may be addressed in the future.
1. Final Conclusions
The focus of this Thesis is related to the development of a novel framework for the recognition
of ADL and its environments that includes the selection of the best methods for different stages,
such as Data Acquisition, Data Fusion, Data Processing, Data Cleaning, Feature Extraction and
Classification, based on the sensors available in a smartphone. The most important stages
consist in the Data Fusion and Classification methods that mostly influence the results obtained
by the framework for the purpose of this Thesis.
Mobile devices have several available sensors, including the accelerometer, the gyroscope, the
magnetometer, the microphone and the Global Positioning System (GPS) receiver, which allow
the acquisition of different types of physical and physiological parameters. The mobile devices
may be considered as multi-sensor devices, which is the first part of the title of this Thesis.
There is a variety of mobile devices, but the scope of this Thesis focuses only in the use of
smartphones. The acquisition of the different parameters allows the recognition of Activities
of Daily Living (ADL), where the work of this Thesis is to implement data fusion and
classification methods for the recognition of ADL and its environments. There are several ADL
that can be recognized, but this Thesis focuses in the recognition of 7 ADL, i.e., walking,
running, going upstairs, going downstairs, standing/watching TV, sleeping and driving activities
and 9 environments, i.e., bedroom, bar, classroom, gym, kitchen, living room, hall, street and
library, with the sensors available in the mobile devices. These devices have several limitations,
such as limited power, processing and memory capabilities, which are minimized with the
151
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
implementation of different methods. Finally, the developed framework may be used for the
mapping of lifestyles with the ADL and environments recognized.
The main objective of the work reported in this Thesis is the creation of a new method and
framework for the recognition of ADL and environments with the fusion and classification of
the data acquired from a large set of sensors available in off-the-shelf mobile devices. The
developed method is a function of the number of available sensors embedded in the different
mobile devices. It should be implemented with lightweight algorithms in order to allow its
execution locally in the mobile devices. The main objective was divided into six smaller
objectives, which started with the literature review, verifying that there are no available
studies that use a large set of sensors available in the mobile devices and fuse the data acquired
with them, but it is a topic that is widely growing with different research works. The literature
review also reveals that Artificial Neural Networks (ANN) are widely used, and reported better
results than other methods. Finally, the literature review helped in the definition of the ADL
and environments to be recognized, and in the definition of the architecture of the novel
framework with Data Acquisition, Data Fusion, Data Processing, Data Cleaning, Feature
Extraction and Classification as stages. Among them, Data Fusion and Classification are the
main modules of the framework. Data Acquisition based using mobile devices may be performed
with standard methods. Data Processing, Data Cleaning and Feature Extraction methods depend
on the types of sensors used as presented in chapter 3, which is focused on the definition of
the architecture of the framework for the recognition of ADL and its environments. Next, data
acquisition was performed with a mobile application that allows the labelling of the ADL and
environments for further processing. In continuation, the acquired data was analyzed and the
different methods were implemented, including low pass filter, Fast Fourier Transform (FFT),
and so on. After the extraction of the features, three different implementations of ANN
methods had been tested, including Multilayer Perceptron (MLP) with Backpropagation,
Feedforward Neural Networks (FNN) with Backpropagation and Deep Neural Networks (DNN),
where these implementation were detailed in chapter 4. The different implementations
reported that the method that obtains the best results is the DNN method with normalized
data.
The first part of this Thesis, presented in chapter 2, consists in the literature review on the
different concepts related to the creation of a novel framework for the recognition of ADL and
its environments. Initially, the main proposed concepts are Data Acquisition, Data Processing,
Data Imputation and Data Fusion, which are explored and several frameworks were described
for each concept as well as their applicability. However, these solutions are difficult to be
adapted to the mobile devices, because the development of solutions for these devices should
take into account the different hardware restrictions that depend on the manufacturers.
Moreover, during the first part of the literature review, it was verified that Data Imputation
may be a part of Data Processing, therefore it was excluded from this research, as the
152
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
classification methods use the features extracted from the sensors’ signal, the inexistence of
data in some instants are not very relevant. The second part of the literature review consists
on the research on the validation methods, including Data Cleaning and Data Imputation
methods in the Data Processing module. Where, the Data Cleaning methods are mainly filtering
methods, including low pass filter, high pass filter and others; and the Data Imputation methods
reviewed are mainly the K-Nearest Neighbor (KNN) and their variants. Next, the classification
methods for the identification of the ADL and environments were reviewed. It was verified that
one of the most used methods with better results are ANN. As the acoustic data was also used
for the recognition of the environments, audio fingerprinting methods were also reviewed,
verifying that FFT should be implemented for the extraction of the frequencies. It was
concluded that for all types of sensors’ data, the most used features with better results are
statistical features.
The second part of this Thesis, presented in chapter 3, consists in the definition of the
architecture of the framework for the recognition of the ADL. In this chapter it is mentioned
that the execution starts with the Data Acquisition process. After the acquisition of the data,
it is sent to the Data Processing, including Data Cleaning, Data Imputation and Feature
Extraction. Data Cleaning will take into account the sensors used, implemented the low pass
filter method for the motion and magnetic sensors and the FFT for the acoustic sensors. Firstly,
for motion and magnetic sensors, the features extracted are the five greatest distances
between the maximum peaks combined with the average, standard deviation, variance and
median of the maximum peaks, and the standard deviation, average, maximum, minimum,
variance and median of the raw signal. Secondly, for the acoustic sensors, the features
extracted are the 26 Mel-Frequency Cepstral Coefficients (MFCC) combined with the standard
deviation, average, maximum, minimum, variance and median of the raw signal. Finally, for
the location sensors, the feature extracted is the distance travelled. After the feature
extraction and based on ANN, the classification method were implemented. The output should
be combined with the users’ agenda in future developments.
The third part of this Thesis, presented in chapter 4, consists in the explanation of the main
achievement of this Thesis, which is the implementation of the different methods in the
framework for the recognition of ADL and its environments. The implementation presented in
this Thesis was incremental, starting with the implementation with only one sensor, i.e.,
accelerometer, to the implementation with most common sensors in the smartphones, i.e.,
accelerometer, magnetometer, gyroscope, microphone and GPS receiver. The implementations
are performed with three different configurations of ANN, such as MLP method with
Backpropagation, FNN method with Backpropagation and DNN method with normalized and nonnormalized data. All implementations are based in groups of features extracted from the
different sensors’ data available in each study. The normalization for MLP and FNN methods
was performed with MIN/MAX normalizer and the normalization for DNN method was performed
153
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
with mean and standard deviation. The details about the configurations of the ANN methods
are available in chapter 4. The novel framework performs the recognition of ADL and
environments in three stages, these are the recognition of common ADL with motion and
magnetic sensors, the recognition of environments with acoustic sensors, and the recognition
of activities without motion with motion, magnetic and location sensors fused with the
environment recognized. Following the results obtained with the accelerometer data, the best
results were achieved with DNN method, reporting an accuracy of 85.89% with normalized data
and 80.35% with non-normalized data using cross-validation in the recognition of ADL. Following
the results obtained with the fusion of the accelerometer and the magnetometer data, the best
results were achieved also with the DNN method, reporting an accuracy of 86.49% with
normalized data and 70.43% with non-normalized data using cross-validation in the recognition
of ADL. Following the results obtained with the fusion of the accelerometer, the magnetometer
and the gyroscope data, the best results were achieved also with the DNN method, reporting
an accuracy of 89.51% with normalized data and 74.47% with non-normalized data using crossvalidation in the recognition of ADL. Following the results obtained with the acoustic data, the
best results were achieved with FNN method with Backpropagation, reporting an accuracy of
82.75% with normalized data and 86.50% with non-normalized data using cross-validation in the
recognition of environments. Following the results obtained with the fusion of previously used
sensors with the location sensors and/or the environment recognized, the best results were
achieved with DNN method, reporting an accuracy of 100% with the different combinations of
sensors in the recognition of activities without motion. Finally, the novel framework was
implemented and validated in a smartphone, reporting an overall accuracy between 58.02% and
89.15%.
The main objective of this Thesis was accomplished with the development of the novel
framework for the recognition of ADL and environments, recognizing a greater number of
ADL/environments than the previous works analysed and reporting an overall accuracy higher
than the mean of the results, and a similar accuracy than the highest results reported by the
previous works available in the literature.
2. Future Work
The recognition of ADL and its environments may be performed with other approaches, and the
results obtained by the framework may be improved with research and implementation of Data
Imputation methods to be incorporated into the Data Processing module. The Data Processing
154
Multi-sensor data fusion in mobile devices for the identification of Activities of Daily Living
module can be also improved with other approaches related to Data Validation and Data
Cleaning methods.
However, the most important module of the framework for the recognition of ADL and its
environments is the Classification module, in which other types of machine learning methods
may be implemented and tested with the datasets acquired during this Thesis, including
Support Vector Machines (SVM), Decision Trees and Ensemble Learning Methods (e.g., Adaboost)
in order to check the results are better with other machine learning methods.
Taking into account the limited battery, processing and memory capabilities of the
smartphones, the most important challenge in the development of the framework is to arrange
a method that acquires the data continuously without effects in the performance of the
smartphones in order to be able to recognize risk situations, e.g., falls of older people.
Currently, the method developed is prepared for smartphones. The implementation of the
framework in other devices (e.g., smartwatches) would require further experimentation,
acquiring new data and developing new classification methods adapted to these devices.
Consequently, the framework should be adapted to the specific characteristics of the mobile
device in use. One of the most important differences in the data acquisition between the
smartphone and the smartwatch is the position of them, because the smartphone can be placed
in the front pocket of the user’s trousers, but the smartwatch is commonly placed on the user’s
wrist.
As future work, the number of ADL and environments recognized by the framework may be
incremented, but this would also require further experiments. As final achievement, this
framework may be used in the development of a personal digital life coach.
155