This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE/ASME TRANSACTIONS ON MECHATRONICS 1

Foot2hip: A Deep Neural Network Model for

Predicting Lower Limb Kinematics From
Foot Measurements
Rishabh Bajpai , Graduate Student Member, IEEE, and Deepak Joshi, Member, IEEE

Abstract—Objective: This study aims to develop a neural sciences. Such an understanding may lead to the development of
network (foot2hip) for long-term recording of gait kinemat- prostheses, exoskeletons, and rehabilitation strategies for indi-
ics with improved user comfort. Methods: Foot2hip predicts viduals with gait disorders. Previously, researchers attempted to
ankle, knee, and hip joint angle profiles in the sagittal plane
using foot kinematics and kinetics during walking. Foot2hip understand gait by modeling its structural and functional charac-
consists of three convolutions, two max-pooling, two LSTM, teristics using recorded gait kinematics and kinetics. The widely
and three dense layers. An indigenously developed insole used and accurate motion capture system for recording gait
and an outsole were used to measure the kinetics and consists of several optical sensors and reflective markers. These
kinematics of the foot, respectively. Seven healthy partic- markers are placed on specific anatomical positions of the body
ipants were recruited to follow an experimental protocol
consisting of six walking conditions: slow, medium, fast and are tracked by optical sensors during the recording session.
walking speed, rearfoot, flatfoot, and forefoot landing pat- Later, the tracked positions are fed into an inverse kinematic
tern. Results: When tested for leave-one-out and nested framework to predict the movement pattern of the body parts.
cross-validation, foot2hip obtained 3.04◦ ± 0.20 RMSE and A similar and popular approach uses inertial measurement units
0.97 ± 0.01 correlation coefficient for knee joint, 1.7◦ ± 0.09
(IMUs) for a three-dimensional (3-D) motion capture system.
RMSE and 0.95 ± 0.01 correlation coefficient for hip joint,
and 1.32◦ ± 0.08 RMSE and 0.91 ± 0.02 correlation coeffi- Such a system is relatively easier to use and provides more
cient for ankle joint (averaged across all folds). Conclusion: flexibility to perform experiments outside the laboratory setting.
The prediction performance of foot2hip is encouraging and However, both IMU and optical sensor-based methods failed to
shows its applicability in accurately predicting lower limb address some fundamental aspects of data recording, namely,
kinematics with minimal wearables. Significance: The hard-
user’s comfort, ease of placement of markers or sensors, the pos-
ware used along with foot2hip is low cost ($268 + N ×
$35, N is the number of different foot sizes), comfortable, sibility of long-interval on-ground motion capture, affordability
and easy to use. Therefore, the system is suitable for most of the system, elimination of the requirement of an experienced
clinical and personal applications. clinical research engineer for system setup, high computational
Index Terms—Foot kinematics, gait analysis, human lo- cost, and user aversion to their personal use due to its low
comotion, mechatronics, predictive models. aesthetics.
To deal with the aforementioned problems, researchers pro-
posed some solutions. The first and most obvious solution was
I. INTRODUCTION to reduce the number of IMUs. In this direction, few studies
NDERSTANDING the movement patterns of different used a single IMU sensor for predicting either joint angles or
U parts of the human body during locomotion (gait) is of
utmost importance in the fields of sports, robotics, and medical
gait-related parameters. Sun et al. [1] proposed an artificial
neural network (NN) to predict the values of IMU sensors
virtually placed at the shin, thigh, and waist from one IMU
mounted at each shoe. Given that NNs were used in their work
better accuracy was expected. Also, for clinical and research
Manuscript received 16 October 2022; accepted 12 December 2022.
Recommended by Technical Editor D. Oetomo and Senior Editor H. applications, it is crucial to estimate joint (hip, knee, and ankle)
Qiao. This work was supported by the Prime Minister’s Research Fel- angles rather than segment orientations. A similar approach was
lows (PMRF) Scheme. (Corresponding author: Deepak Joshi.) used by Baghdadi et al. [2] to predict some lower-limb-related
This work involved human subjects or animals in its research. Ap-
proval of all ethical and experimental procedures and protocols was parameters such as hip acceleration, hip flexion, and trunk
granted by All-India Institute of Medical Sciences (AIIMS), New Delhi, posture from a single IMU using an extended Kalman filter
India ethics committee under Application No. IEC-222/04.05.2018. (KF) and an unscented KF. However, they used an expensive
The authors are with the Centre for Biomedical Engineering, Indian
Institute of Technology Delhi, New Delhi 110016, India, and also with IMU sensor and multiple computationally expensive algorithms
the All India Institute of Medical Science Delhi, New Delhi 110029, India for predicting the angles. In another work, instead of directly
(e-mail: rishabhbajpai24@gmail.com; joshid@iitd.ac.in). predicting gait parameters, Goulermas et al. [3] utilized eight
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TMECH.2022.3229404. regression algorithms including NN and linear regression to
Digital Object Identifier 10.1109/TMECH.2022.3229404 predict joint angle values from a single IMU mounted at the foot.

1083-4435 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE/ASME TRANSACTIONS ON MECHATRONICS

Since they had not used any dedicated algorithms for removing from a research group also used this approach. Sivakumar et al.
the drift of IMU values, their method is not suitable for long-term [12] predicted angle joint angle in the stance phase of the
data collection. Recently, Lee and Park [4] estimated ground gait cycle using GRF and feedforward NN, and in [13], they
reaction force (GRF), center of pressure (CoP), and joint torque predicted all three lower joint angles in sagittal plane during
from a single IMU attached to the sacrum using an artificial full gait cycle using GRF and wavelet NN. GRF contains useful
neural network model. Using a single IMU definitely improves but indirect information about the position of the lower limb
the ease of sensor placement and eventually ensures greater with respect to the body. Since the generally observed walking
comfort to the user. Although, its readings are unreliable, as pattern is symmetric and closely follows periodic movements,
even the most advanced signal processing algorithms are unable it can be predicted from the GRF data. However, for the clinical
to eliminate the effect of drift phenomenon in readings of a single population gait kinematics get significantly changed. Therefore,
IMU while estimating accurate position in a continuous motion more information may be required for estimating rarely observed
task [5]. Also, the computational cost of the prediction increases walking patterns.
with increased accuracy from an advanced signal processing
algorithm.
Therefore, for more reliable and practically applicable mo- A. Motivation and Contribution
tion capture systems, researchers tried multiple IMUs setups. The conventional methods of lower limb kinematic estimation
Recently, Sy et al. [6] exploited three IMUs to predict hip require wearing special suits or attachments to the body that
and knee angle profiles using constrained KF. Their method can cause uneasiness to the user, leading to deviation from
obtained a decent performance with a root-mean-squared error the natural gait. Since most analyses are performed in clinical
(RMSE) of 11.1◦ and 11.8◦ for predicting knee and hip joint gait laboratories, the recording is bounded with temporal and
angle profiles, respectively. The pseudomeasurements of the spatial constraints. Also, long-term recording in an IMU-based
pelvis’ location used in their method stop the pelvis from moving motion capture system requires periodical recalibration. The
away from the feet. It will, however, also make it difficult to accurate gait analysis is restricted to resourceful institutes since
accurately estimate poses during activities like sitting and lying the conventional motion capture system uses sophisticated and
down, where the pose is kept for a lot longer than a usual expensive apparatus. These systems are also extremely sensi-
gait cycle. Therefore, their method is not suitable for abnormal tive to marker/sensor misplacement; a slight misplacement of
gait applications. In two similar studies, Hu and Soh [7], and marker/sensor in children with cerebral palsy (CP) may result
Findlow [8] used two IMUs per foot to predict lower limb angles in increased measurement errors [14]. Moreover, these systems
with inverse kinematics (IK) and generalized NN, respectively. are time-taking to set up, require an experienced clinical gait
Their method not only involved expensive IMUs but also would laboratory engineer, and are less aesthetically appealing to be so-
be unsuitable for clinical uses where abnormal walking patterns cially acceptable for personal use [15]. Furthermore, a computer
and anatomical deviations are observed. Li et al. [9] attempted system that can handle computationally expensive operations is
to reconstruct walking dynamics using only two IMUs. Their required for real-time decoding of gait parameters using these
model obtained RMSE in a range of 5.3◦ and 6.8◦ for lower methods [15]. Broadly, three research gaps were identified based
limb joint angles. There are four major challenges in estimating on the literature: 1) expensive or less suitable sensors for long-
the positions of joints using IMUs. First, during integration, duration recording, 2) computationally expensive algorithms for
zero offset and added integration drift due to data noise need to accurate estimation of kinematics, and 3) no use of biomechanics
be corrected. Second, electromagnetic radiations and metallic of gait for reducing computational cost and improving estima-
objects can disturb the readings of IMU. Third, it computes tion accuracy. Utilizing the leverage of mechanical constraints of
displacement in the local coordinate system. These distance the lower body and anthropological measurements of the user,
values need to be converted to a global frame of reference in the lower limb can be approximated as a mechanism, where
order to compute joint angles. Also, calibration is required for the bones can be considered as links and body joints as lower
formulating the relation between sensors and anatomical body or higher pairs. The approximated mechanism can be used to
frames of reference. Fourth, anthropological details of the lower estimate the lower limb kinematics using foot kinematics during
limb are required for performing IK. Thus, researchers explored walking. However, such estimation would need computationally
the second approach of using sensor fusion techniques for joint expensive algorithms (for example, IK). Also, the calculation of
angle estimation. Yi et al. [10] used nine electromyography the relative position and orientation of the foot with respect to
(EMG) sensors and one IMU sensor placed at the knee joint to the trunk are not possible with the currently available physical
predict the knee angle profile. Their method achieved an RMSE sensors. Therefore, in the proposed work, simple and reliable
of 3.98◦ ; however, due to its high implementation cost, it lacks instrumentation was used to get long-duration foot kinematics
in demonstrating its applicability in practical scenarios. recordings, and an NN is developed to estimate lower limb
The third approach involves the use of sensors other than kinematics with robust and fast computation (41 ms). This
IMUs for real-time estimation of joint angles and gait-related study proposes an NN-based model named “foot2hip” to ad-
parameters. Mun et al. [11] estimated gait-related temporospatial dress the aforementioned limitations. For seamless replication
parameters by feeding features of the foot to a deep NN. Since of the hardware and algorithm, easily available sensors and
their algorithm relies on the foot structural features, it may microcontrollers were used in the study. The technical details
not work with the foot having deformities. Two more studies of the hardware are presented in [16] and [17] and the simplified

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BAJPAI AND JOSHI: FOOT2HIP: A DEEP NEURAL NETWORK MODEL FOR PREDICTING LOWER LIMB KINEMATICS 3

code of foot2hip implementation is publicly available online

at https://github.com/Rishabh-Bajpai/Foot2hip. Primary signif-
icant contributions and novelty are as follows.
The novelty of this study lies in twofold namely, conceptu-
alization of the solution by understanding various aspects of
the design, development and implementation, and development
of a novel NN model (foot2hip) for estimating lower limb
kinematics.
1) This is the first time these sensors have been used to
accomplish this, and the presented method requires no
IK computations or wearable sensors other than at the
foot.
2) An NN model, foot2hip, was developed after performing
rigorous experiments. Analyses of the activation function,
batch normalization parameter choices, and an ablation
study were performed to demonstrate the system’s sensi-
Fig. 1. Simplified model of the leg that can be used to predict hip,
tivity to NN architecture choices. knee, and ankle angles in the sagittal plane using IK. The negative
Secondary significant contributions are as follows. values in ROM are for flexion/dorsiflexion and positive values are for
1) Foot2hip is a generalized model that shows an excellent extension/plantar flexion.
prediction performance when tested on unseen subjects
without seeking any anthropometric measurements of the
subject.
2) Foot2hip is verified at various walking speeds and simu-
lated, commonly observed abnormal foot landing patterns
for demonstrating the applicability of foot2hip in clinical
settings.
3) Additional analyses were performed to assess the per-
formance of the system for the comfort of the user,
long-duration data recording, and real-time recording of
lower limb kinematics.
The rest of the work is as follows: The methodology used for
developing and validating foot2hip is explained in Section II.
Experimental results are presented in Section III. Discussion of Fig. 2. (a) Graphical representation of the used instruments. (b) Loca-
the results, comparison with the previous work, limitation, and tion of sensors in the outsole and data transmission channels. Sensors
future work is presented in Section IV. Finally, the conclusion 1–6 are six VCSEL sensors.
is drawn in Section V.
kinematics and kinetics of the foot, a more robust prediction
II. METHODOLOGY module can be developed. Therefore, an alternate approach
A. Solution Formation From the Prior Knowledge of Gait is required to predict joint angles from past and current foot
Biomechanics kinematics and kinetics values.

To understand the mechanics of gait, the lower limb can be

represented as a simple mechanism with certain assumptions. B. Instrumentation
Fig. 1 shows a leg as a mechanism consisting of four links and To measure the right lower limb joint angles in the sagittal
three rotational/revolute joints with motion in the sagittal plane. plane, three twin-axis-wired goniometers (Biometric, Ltd., U.K.,
Suppose the position and orientation of the foot with respect SG series) were used. However, only one axis was consid-
to the trunk are known. In that case, angles of the intermediate ered for this study. The goniometer data were collected at a
joints can be calculated using IK by incorporating knowledge sampling rate of 1000 Hz using a dedicated data acquisition
of constraints of the joint motion. However, calculation of the (DAQ) system (Biometrics, Ltd., U.K., DataLOG-MWX8). The
relative position and orientation of the foot with respect to CoP in the sagittal plane during walking was calculated from a
the trunk are not possible with the currently available physical wireless pressure insole developed by our research group [17]
sensors. Fortunately, if we leverage the dynamic properties of at a sampling rate of 64 Hz. The pressure insole measures
the mechanism, its current joint angles can be predicted using CoP in the sagittal plane by incorporating 16 force resistive
past and current values of foot measurements. Moreover, in the sensors with a range of 0–100 lbs (0–445 N). For measuring
swing phase of gait, the kinematics of the foot contains valuable distances of the foot from ground, an indigenously developed
information, and in the stance phase, the kinetics of the foot wireless outsole was used [16] at a sampling rate of 31 Hz. The
contains valuable information about the gait. By utilizing both outsole incorporated six Vertical Cavity Surface-Emitting Laser

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE/ASME TRANSACTIONS ON MECHATRONICS

Fig. 3. Graphical representation of protocols used during the experiment. First, a treadmill and protocol familiarizing session of 10.75 min was
conducted followed by a rest of 3 min. Then, gait data at a self-selected walking speed was recorded for 5 min. Then, protocol 1 was followed for
27.25 min. While the recording was on, a rest of 90 min was given to the subjects. Then, protocol 2 was followed for 19 min. Then, after a rest of 10
min, the gait data at a self-selected walking speed was recorded for 5 min.

(VCSEL) sensors at different anatomical locations of the foot, walking. Furthermore, the shift in CoP during footstrike results
as shown in Fig. 2. All three systems (i.e., goniometer, insole, in different foot landing strategies. This approach was used in the
and outsole) were time-synchronized by sending a simultaneous past for the same purpose of inducing foot-landing strategy [16].
triggering pulse of 50 ms to their respective DAQ systems. The Posterior CoP shift simulates the rearfoot landing pattern, medial
data were recorded from the right leg only. The cost of the outsole CoP shift simulates the flatfoot striking pattern, and the anterior
is approximately $116 plus $20 per foot size. The combined cost CoP shift simulates the forefoot landing pattern. Two trials of 90
of the insole and outsole is approximately $268 plus $35 per foot s were performed for each walking condition with a rest period
size. of 2 min between the trials.
The motivation for simulating different foot landing strate-
gies stemmed from commonly observed gait abnormalities in
C. Experimental Protocols children with CP [18]. Toe walking due to Achilles tendon
Data were collected from seven able-bodied male subjects contracture, flat foot walking due to multiple contractures, and
with a mean age of 26.1 ± 1.7 years, mean height of 177.6 ± excessive dorsiflexion in crouch gait due to spasticity may result
7.7 cm, and mean weight of 75.4 ± 12.1 kg. Volunteers were in a forefoot landing, flatfoot landing, and rearfoot landing,
examined before the recruitment and had no neurological, ambu- respectively [19]. Therefore, three abnormal gait patterns are
latory, or musculoskeletal dysfunction. This study was approved simulated to validate the applicability of foot2hip for patients
by the All-India Institute of Medical Sciences (AIIMS), New with gait abnormalities.
Delhi, India, ethics committee (Ref. no IEC-222/04.05.2018).
Mainly the data acquisition session was constituted of five parts:
1) treadmill and protocol familiarization session, 2) baseline D. Data Preprocessing
blocks, 3) protocol 1, 4) protocol 2, and 5) long rest periods The distances measured by the outsole were preprocessed
(see Fig. 3 for details). Treadmill and protocol familiarization using a fifth-order Butterworth low-pass filter of 10 Hz cut-off
sessions ensured that the subjects were able to simulate the al- frequency followed by KF with the variance in the measurement
tered foot landing patterns. Baseline blocks and long rest periods noise and process noise of 5 × 10−5 and 1 × 10−2 , respectively.
were included to understand the change of estimation error with The parameters of KF were selected for optimal tradeoff between
time. The details of the two experimental protocols are given as accuracy and response time [16]. After filtering, all three signals
follows. (joint angles, CoP, and distances) were downsampled to 30 Hz.
1) Protocol 1. Walking at Different Speeds: Each participant Then, all three time-locked signals were concatenated into a
wore an instrumented outsole, a pressure insole, and three unified data matrix, as shown in Fig. 4(a). The first six rows
goniometers at the hip, knee, and ankle joints. The data were contained foot to ground distances, seventh row contained CoP
recorded for three walking speeds viz., slow (1.8 km/h), medium values, and the last three rows contained angle values. The
(2.8 km/h), and fast (3.8 km/h). For each walking speed, three unified data matrix was split into two matrices that serve as
trials (each of 75 s duration) with a break period of 2 min in input and output data for training foot2hip. Since foot2hip was
between the trials were conducted. supposed to predict angle values at a time point “t” from current
2) Protocol 2. Altered Foot Landing Strategies: The partici- and past values of CoP and distances, separate input and output
pants used three different foot striking strategies while walking samples were extracted for each time point of the unified data
at their preferred walking (2.8 km/h, determined based on par- matrix (except for the first 149 time points). These samples were
ticipant feedback) speed on the treadmill: 1) rearfoot landing stacked to make input and output data matrices. Each sample
(with abnormal heel strike), 2) flatfoot landing (with flat foot of the input and output matrices is referred as a data point
strike), and 3) forefoot landing (with toe strike). The participants throughout the manuscript. For an output data point, the values
were shown visual CoP feedback on the monitor screen during of angles at a time point “t” were extracted from the unified data
the experiment to encourage them to modulate their CoP while matrix. The resulting dimension of an output data point was

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BAJPAI AND JOSHI: FOOT2HIP: A DEEP NEURAL NETWORK MODEL FOR PREDICTING LOWER LIMB KINEMATICS 5

E. Analysis 1. Finding a Suitable Model for the Best

Performance
Different types of NN had shown potential in using them for
angle profile generation during gait [20]. However, use of CNN,
LSTM, and dense NN has shown remarkable performance in
gait-related applications. The reason for the success of these NN
could be their ability to efficiently process the signal in spatial
and temporal dimensions. The selection of NNs architectures
and its parameters is a crucial step to archive desired results.
Specifically learning rate, activation function, and batch nor-
malization can significantly influence the training. To determine
these optimal parameters a grid search on the validation set was
performed. Other hyperparameters, such as number of layers,
filter size, layer parameters, stride, and pooling filter size, were
heuristically chosen.
1) Architecture of the Best Performing Model: Based on the
observations of the heuristic evaluation and grid search, a CNN-
LSTM-DNN-based model (foot2hip) was finally selected. The
architecture of foot2hip is shown in Table I. First, three convo-
lution layers with max pooling were applied to the signal for
removing irrelevant features from the signals. The dimensions
of the filters were set to one instead of two to avoid dilution
Fig. 4. Formation of input and output data matrices from distance (foot of signals in the channel domain. Two LSTM layers were used
to ground), CoP, and angle values. (a) First six rows (from the top) have
distances, seventh row has CoP values, eighth row has knee, ninth row for predicting decision parameters based on temporal dynamics
has ankle, and the tenth row has hip angle values. The blue shaded of the convoluted signals. At last, three dense layers were used
region corresponds to a data point (sample) of the input matrix, and to generate the required angles from the decision parameters.
the red shaded region corresponds to a data point of the output matrix.
(b) Input and output matrices for “N ” number of data points are shown, Architecture of foot2hip ensures the utilization of spatial, static,
where “N ” is the total number of data points. and dynamic features relevant for prediction of angles during
walking.

F. Analysis 2. Verifying Performance of the Model at

Untrained Walking Conditions
To verify the performance of foot2hip at untrained walking
conditions for a test (unseen) subject, a nested cross-validation
was performed. For a fold, the data were split into two parts
namely training subjects data and testing subject’s data. Then,
the training subjects’ data are further divided into two parts
according to the walking conditions. The first part of the dataset
contains data on all conditions (walking conditions) excluding
one condition (say testing condition). This part is used as the
training dataset for the given fold and the other part of the data
Fig. 5. Flowchart of the formation of input and output data matrices was discarded. Similarly, the testing subject’s data are further
from distance (foot to ground), CoP, and angle values.
divided into two parts according to the walking conditions. The
first part of the dataset contains data corresponding to the testing
1 × 3. For “N ” number of data points, the dimension of output condition. This part is used as testing data set for the given fold
matrix became N × 3. Similarly, for an input data point, values and the other part of the data was discarded. Then, foot2hip’s
of distances and CoP from “t − 149” to “t” time points were was trained on the training dataset and its performance was
extracted from the data matrix. The resulting dimension of an tested on the testing dataset. This process was repeated for each
input data point was 1×150×7 (5 s × 30 Hz = 150 time points, condition and each subject. For a better understanding of the
and 6 distance channels and one CoP channel = 7 channels). nested cross-validation, refer to Figs. 6 and 7.
For N number of data points, the dimension of the input matrix
became N ×150×7. For a better understanding of the formation
of input and output matrices, see Figs. 4 and 5. Later, the input G. Analysis 3. Ablation Study
and output matrices were standardized for zero mean and one Performance of three NN models consisting of CNN lay-
standard deviation, and standardization parameters were stored ers followed by dense NN layers (CNN-DNN), LSTM layers
for future reconstruction of the original data. followed by dense NN layers (LSTM-DNN), and CNN-LSTM

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE/ASME TRANSACTIONS ON MECHATRONICS

TABLE I
DETAILS OF THE FINAL ARCHITECTURE OF FOOT2HIP (CNN-LSTM-DNN MODEL) DECIDED AFTER PERFORMING GRID SEARCH AND USING HEURISTIC
APPROACH FOR CHOOSING HYPERPARAMETERS

compare them for better performance and biomechanical rele-

vance. Various combinations of number of layers, number of
filters, dimensions of the filters, and activation function were
tried for each model to figure out the best model. The architecture
of CNN-DNN, LSTM-DNN, and CNN-LSTM-DNN are given
in Tables II, III, and I, respectively.

H. Training of Model for Each Analyses

In total, 150 epochs with a batch size of 16 samples were
used during the training as they are found better than the other
combinations. A model checkpoint of early stopping with 10
sample points on validation loss was used to avoid overfitting.
For finding reconstruction error, the output and the predicted
data matrices are destandardized.

I. Evaluation Metrics
The performance of a predictive NN algorithm is often eval-
uated by the difference between the values of ground truth and
Fig. 6. Graphical representation of data splitting in nested cross- the predicted values. In this study, RMSE, mean absolute error
validation.
(MAE), and Pearson’s correlation coefficient (ρ) were used for
quantifying predictive performance of the model. Then, these
evaluation scales were averaged across the folds.

J. Environment Settings
For dataset formation and training of NN models, a system
having the following specifications was used; Intel(R) Core(TM)
i7 6700 CPU @ 3.40 GHz, 24 GB DDR4 RAM, 64-b Windows
10 Operating System, Matlab 2020a platform, python version
3.7.9, and tensorflow version 2.7.0.

III. RESULTS
Fig. 7. Flowchart of data splitting in nested cross-validation. A. Results. Analysis 1
Since the results of the Adam optimizer are generally better
layers followed by dense NN layers (CNN-LSTM-DNN) were than every other optimization algorithm, it was used for op-
tested on the dataset using leave-one-out cross-validation. For timization [21]. Training parameters such as learning rate of
each fold, data of one of the six subjects are treated as testing the Adam optimizer, activation function for CNN and LSTM
data, and data of the remaining subjects were treated as train- layers, and batch normalization layers were optimized for
ing data. Ablation studies were conducted on all three models foot2hip on nested
 cross-validation.
 The set used for learn-
(namely, CNN-DNN, LSTM-DNN, and CNN-LSTM-DNN) to ing rate was 0.01, 0.001, 0.0001 , activation function was

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BAJPAI AND JOSHI: FOOT2HIP: A DEEP NEURAL NETWORK MODEL FOR PREDICTING LOWER LIMB KINEMATICS 7

TABLE II
DETAILS OF THE FINAL ARCHITECTURE OF CNN-DNN MODEL DECIDED AFTER PERFORMING GRID SEARCH AND USING HEURISTIC APPROACH
FOR CHOOSING HYPERPARAMETERS

TABLE III
DETAILS OF THE ARCHITECTURE OF LSTM-DNN MODEL DECIDED AFTER PERFORMING GRID SEARCH AND USING HEURISTIC APPROACH
FOR CHOOSING HYPERPARAMETERS

TABLE IV
RESULTS OF THE GRID-SEARCH ANALYSIS

 
Sigmoid,
 ReLu, LeakyRelu,
 ELU , and batch normaliza-
tion was with, without during the grid search. The results of
the optimization process are shown in Table IV. We observed that
adding batch normalization did not much affect the performance
of the model. The model performed best on the validation set
with learning a rate of 0.01, activation function of ReLu, and
without batch normalization layers (shown in violet color in
Table IV). Based on the results, the Adam optimizer with a
learning rate of 0.001, β1 = 0.9, β2 = 0.999, and epsilon =
107 was selected to update the weights of the network during
the training. The β1 and β2 values represent the exponential
decay rates of the Adam optimizer algorithm’s first and second- Fig. 8. Foot2hip’s training and validation loss averaged across all folds.
moment estimations, respectively. The mean-square error was
used as the loss function.
model during training. The training and validation loss averaged
across all training folds of “Analysis 2” are shown in Fig. 8.
While performing nested cross-validation, the proposed NN
B. Results. Analysis 2 was tested on an unseen subject for unseen walking conditions.
During the training, the training data was further divided into Performances of foot2hip at various conditions are presented in
training and validation data for monitoring overfitting of the Table V. Each column in the table was averaged across subject

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE/ASME TRANSACTIONS ON MECHATRONICS

TABLE V
PERFORMANCE OF THE FOOT2HIP FOR NESTED CROSS-VALIDATION

TABLE VI
SAMPLE RESULT OF AN ABLATION STUDY

folds to compute the overall performance of the foot2hip for each TABLE VII
PERFORMANCE OF THREE NN MODELS FOR LEAVE-ONE-OUT
condition. One-way analysis of variance with the Bonferroni CROSS-VALIDATION
correction was used for testing the significance of the difference
in means of the distributions of RMSE, ρ, and MAE. Each
walking condition was compared with other walking conditions
in that protocol. For example, the slow walking condition was
compared with medium and fast walking conditions and rearfoot
walking condition was compared with flatfoot and forefoot
walking conditions. During the multiple-comparison results, a
p-value lesser than 0.05 was considered significant. The results
of this multiple comparisons suggest that foot2hip performed
better at slow and medium walking speed conditions than the
fast walking speed condition in predicting knee and hip angles.
Also, foot2hip performed better at rearfoot and flatfoot walking
conditions than the forefoot walking condition for all three joint
angles.

C. Results. Analysis 3
After considering many combinations of number of layers,
number of filters, filter dimensions, and activation functions,
one model from each category was selected. When all models
were tested on leave-one-out cross-validation, CNN-LSTM- Fig. 9. Predicted knee, hip, and ankle profiles for an unseen subject
DNN model (foot2hip) obtained the best performance (see Table and walking condition.
VII). Therefore, it was finally selected for further analysis. A
sample plot of the predicted joint angle profiles for a gait cycle by The table shows how the performance of foot2hip varies with
foot2hip is shown in Fig. 9. The profiles shown in Fig. 9 were pre- number of CNN, LSTM, and DNN layers in foot2hip model.
dicted for a test subject on an untrained condition. A video show-
ing simulation of the lower limb using predicted angle profiles
is uploaded at https://github.com/Rishabh-Bajpai/Foot2hip. An IV. DISCUSSION
iterative approach was used for selecting the combination of The results shown in Table VII indicate that the performance
model parameters for best performance. Since it is not feasible to of CNN-LSTM-DNN was better than CNN-DNN and LSTM-
show the performance of foot2hip for all combinations, a sample DNN models. Also, the results of Table VI indicate the crucial
result obtained from the ablation study is presented in Table VI. role of a sufficient number of CNN, LSTM, and DNN layers in

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BAJPAI AND JOSHI: FOOT2HIP: A DEEP NEURAL NETWORK MODEL FOR PREDICTING LOWER LIMB KINEMATICS 9

TABLE VIII
COMPARISON TABLE SHOWING JOINT ANGLE PREDICTION PERFORMANCE OF THE PREVIOUS WORKS AND THE PROPOSED WORK

accurately solving the given problem. From our previous knowl- Moreover, two studies used a foot kinematic measure, i.e.,
edge, we can say that CNN, LSTM, and DNN are specialized vertical GRF. Foot2hip not only obtained the best overall perfor-
in performing feature extraction or noise removal, dealing with mance but also considered many other aspects of data recording.
time-series-related data processing and regression, and mapping Therefore, it is suitable for both personal and clinical uses.
of latent information to meaningful data, respectively. These As problems of the IMU-based systems were highlighted in
specialized skills of the networks were required for solving the Section I, long-term recording using IMU may lead to unreliable
given problem of predicting joint angles from foot kinematics readings. To validate this aspect in the proposed method, the
and kinetics. Since only Kalman and low-pass filters were ap- prediction performance of the foot2hip using outsole was tested
plied to the distance data, more signal processing was needed for a continuous 3-h data recording session for seven subjects.
to find relevant features of the data for joint angle prediction. In the starting block (start baseline block) of duration 5 min,
Therefore, CNN layers were used to improve the signal-to-noise foot2hip obtained RMSE of 2.35◦ . And in the last block (end
ratio of the distance signal. On the other hand, as we know that baseline block) of the same duration, foot2hip obtained RMSE of
the estimation of joint angles using only present values of foot 2.29◦ . Throughout the experiment, the RMSE lies in the range of
kinematics and kinetics is not possible using the conventional IK (2.57◦ , 1.88◦ ). Thus, we can conclude that the change in predic-
approach without knowing the relative position and orientation tion performance with time is negligible [29]; therefore, foot2hip
of the trunk with respect to the foot. Therefore, an alternate is the better option than IMUs for long-term measurements.
approach of using past and current values of foot kinematics and
kinetics was required. LSTM layers provided robustness to the
B. Feasibility Check for Real-Time Applications
proposed alternate approach by efficiently predicting the angle
profiles. Similarly, DNN layers might be helped the LSTM layers In addition, to test the applicability of the proposed method in
in reducing their load of generating the expected angle profile. real time, the computation time durations of predicting joint an-
Results presented in Table V suggested that foot2hip per- gles using foot2hip along with wireless and wired outsole were
formed very well when tested on the unseen subject and walking calculated. First, the clock of Arduino and data recording system
condition. The maximum RMSE of 4.40◦ was observed for knee (a computer system) time-synchronized using a short UBS cable.
angle in the forefoot walking condition. A higher magnitude of Then, the data package sending time (recorded by the Arduino)
error in the forefoot walking condition was expected, knowing and data receiving time (recorded by the data recording system)
the fact that rearfoot walking and flat walking are more similar are noted down. For wireless mode, a mean delay of 213 ms was
to normal walking than forefoot walking. Similarly, since the observed, and for wired mode (long USB cable), a delay of 0 ms
range of motion (ROM) of lower limb joints increases under the was observed. An additional mean delay of 41 ms was observed
fast walking condition, higher prediction errors were expected for predicting angles from the received data using foot2hip
for the fast walking condition. However, even the maximum while testing on a system having the specifications mentioned in
observed errors are in an acceptable range [7], [20], [28], sug- Section II-J. Therefore, the accumulated time delay in real-time
gesting foot2hip can be used for predicting unseen subject’s angle prediction using the proposed system would be 254 ms
angle profiles even for abnormal walking patterns. Therefore, it for the wireless mode and 41 ms for the wired mode. Since the
has a great potential for recording gait kinematics in the clinical electromechanical delay in human skeletal muscles is found to
population. be in a range of 30–100 ms [30], the proposed system in wired
mode can be used for powering the exoskeleton in real time.
A. Comparison With the Previous Work
Table VIII shows the joint angle prediction performance of C. Filling the Research Gap
the previous works and the proposed work. Most of the studies The proposed method uses a self-develop outsole along with
used IMUs for prediction and obtained decent performance. artificial intelligence to estimate lower limb kinematics. The

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE/ASME TRANSACTIONS ON MECHATRONICS

TABLE IX
SUS SCORE CALCULATED FOR ALL PARTICIPANTS

main objective of the study is to address the aspects of data was found that the proposed method obtained excellent perfor-
collection neglected by the previous methods. Since the outsole mance for all tested walking conditions. However, there are a
is thin (15 mm), lightweight (125 gm), and less influencing the few limitations of this study. First, training the model on only
natural gait of the subjects, it provides better comfort to the users. a few walking conditions may lead to a false prediction when
The system usability scale (SUS) [31] was used to quantify the tested on an unseen patient with uncommon gait abnormality.
acceptability of the used hardware for lower limb kinematics Therefore, in the future, for training, a reinforcement learning-
estimation. All participants filled the questionnaire after com- based approach will be used to address the existence of endless
pletion of the experiment. The SUS consists of 10 questionnaire possible gait abnormalities. Moreover, foot2hip will be verified
items (see Table IX) on the Likert scale (1–5) for strongly on CP children in the future. The automated-gait assessment
disagree to strongly agree in increasing order. The SUS score score [32] and AbnormNet [33] developed by us will be used
was calculated by following the steps mentioned in the original with the outsole and foot2hip to provide affordable and accurate
article [31]. The averaged SUS score is given in Table IX. Since gait assessment to CP children. Second, in the experimental pro-
the average SUS of the participants was 84 ± 3.9 (acceptable tocol, the walking conditions were divided into separate trials.
SUS is more than 70), the used hardware was acceptable to However, in the future, to justify the applicability of foot2hip
use. The participants’ feedback after the experiments indicated for real-life applications, a protocol having continuously varying
that they did not feel any hindrance to the motion of limbs or walking conditions will be used. Third, foot2hip is an NN-based
discomfort from the outsole even after its long use. In addition, method that works on the self-learned set of rules. Since foot2hip
the preattached distance sensors to the outsole eliminate the need uses a black-box signal processing approach, it may make false
for an experienced clinical research engineer, reduce the time of predictions under unexpected conditions. Therefore, dedicated
sensor placement, and allow its use for personal applications. hardware will be developed for recording the relative position
Moreover, the wireless outsole enables users to wear any cloth of the hip with respect to the foot for utilizing the concept of IK.
and move around on any flat surface, thus improving the flex- Fourth, the outsole uses IR laser-based sensors for estimating the
ibility to perform experiments with lesser spatial and temporal distance. These sensors are not suitable for estimating distance
constraints compared to the previous motion capture systems. from a rough surface. However, these sensors can be used for
As mentioned before, the outsole’s performance is independent clinical or research purposes where the surface is almost flat (for
of its time of use. Therefore, it can be used for long-term example, treadmill and flat floor).
on-ground motion capture. Furthermore, the outsole provides a
low-priced alternative to wireless high-precision motion capture
($80 including DAQ system). The outsole consists of two parts, V. CONCLUSION
namely sensor modules and outsole templates. These sensor After understanding the shortcomings of the present methods
modules can be attached to templates of different sizes for of motion capture systems, an NN-based algorithm is developed
forming the outsole. For recording data from subjects of different for accurately predicting lower limb joint angle profiles during
foot sizes, multiple templates according to their foot sizes and walking from foot kinematics and kinetics. The proposed system
a single-sensor module pair are needed. Therefore, there is no successfully predicts ankle, knee, and hip angles from the dis-
need to buy multiple outsoles for subjects of different foot sizes. tances of various locations on the foot to the ground and sagittal
This ensures the affordability of the outsole for gait assessment CoP of the foot with a maximum RMSE of 4.62◦ . Foot2hip
centers, where patients with different foot sizes come frequently. addresses three major shortcomings of the present methods;
its setup time is very less compared to the present methods, it
ensures user comfort and it is suitable for long-term recordings.
D. Limitations and the Future Work
This study aims to develop an intelligent system to predict
the lower limb joint angle profiles during walking by addressing
the shortcomings of the previous methods. Moreover, the perfor- ACKNOWLEDGMENT
mance of the proposed method was tested on three commonly The authors would like to thank Mr. Rohan Khatavkar and Dr.
observed abnormal walking patterns for clinical applications. It Ashutosh Tiwari for providing feedback on the proposed system.

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BAJPAI AND JOSHI: FOOT2HIP: A DEEP NEURAL NETWORK MODEL FOR PREDICTING LOWER LIMB KINEMATICS 11

REFERENCES [25] X. Hu, C. Yao, and G. S. Soh, “Performance evaluation of lower limb
ambulatory measurement using reduced inertial measurement units and 3R
[1] Y. Sun, G.-Z. Yang, and B. Lo, “An artificial neural network framework for gait model,” in Proc. IEEE Int. Conf. Rehabil. Robot., 2015, pp. 549–554.
lower limb motion signal estimation with foot-mounted inertial sensors,” [26] F. J. Wouda et al., “Estimation of vertical ground reaction forces and
in Proc. IEEE 15th Int. Conf. Wearable Implantable Body Sensor Netw., sagittal knee kinematics during running using three inertial sensors,” Front.
2018, pp. 132–135. Physiol., vol. 9, 2018, Art. no. 218.
[2] A. Baghdadi, L. A. Cavuoto, and J. L. Crassidis, “Hip and trunk kinematics [27] J. Figueiredo, S. P. Carvalho, J. P. Vilas-Boas, L. M. Gonçalves, J. C.
estimation in gait through Kalman filter using IMU data at the ankle,” IEEE Moreno, and C. P. Santos, “Wearable inertial sensor system towards daily
Sensors J., vol. 18, no. 10, pp. 4253–4260, May 2018. human kinematic gait analysis: Benchmarking analysis to MVN biomech,”
[3] J. Goulermas, D. Howard, C. Nester, R. Jones, and L. Ren, “Regression Sensors, vol. 20, no. 8, 2020, Art. no. 2185.
techniques for the prediction of lower limb kinematics,” J. Biomech. Eng., [28] A. Sharma and E. Rombokas, “Improving IMU-based prediction of lower
vol. 127, no. 6, pp. 1020–1024, 2005. limb kinematics in natural environments using egocentric optical flow,”
[4] M. Lee and S. Park, “Estimation of three-dimensional lower limb kinetics IEEE Trans. Neural Syst. Rehabil. Eng., vol. 30, pp. 699–708, Mar. 2022.
data during walking using machine learning from a single IMU attached [29] J. L. McGinley, R. Baker, R. Wolfe, and M. E. Morris, “The reliability
to the sacrum,” Sensors, vol. 20, no. 21, 2020, Art. no. 6277. of three-dimensional kinematic gait measurements: A systematic review,”
[5] J. Zhao, “A review of wearable IMU (inertial-measurement-unit)-based Gait Posture, vol. 29, no. 3, pp. 360–369, 2009.
pose estimation and drift reduction technologies,” in J. Phys.: Conf. Ser., [30] P. R. Cavanagh and P. V. Komi, “Electromechanical delay in human skeletal
vol. 1087, no. 4, 2018, Art. no. 042003. muscle under concentric and eccentric contractions,” Eur. J. Appl. Physiol.
[6] L. Sy et al., “Estimating lower limb kinematics using a reduced wearable Occupat. Physiol., vol. 42, no. 3, pp. 159–163, 1979.
sensor count,” IEEE Trans. Biomed. Eng., vol. 68, no. 4, pp. 1293–1304, [31] J. Brooke et al., “SUS—A quick and dirty usability scale,” Usability Eval.
Apr. 2021. Ind., vol. 189, no. 194, pp. 4–7, 1996.
[7] X. Hu and G. S. Soh, “A study on estimation of planar gait kinematics using [32] R. Bajpai and D. Joshi, “A-GAS: A probabilistic approach for generating
minimal inertial measurement units and inverse kinematics,” in Proc. 36th automated gait assessment score for cerebral palsy children,” IEEE Trans.
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2014, pp. 6911–6914. Neural Syst. Rehabil. Eng., vol. 29, pp. 2530–2539, Nov. 2021.
[8] A. Findlow, J. Goulermas, C. Nester, D. Howard, and L. Kenney, “Pre- [33] R. Bajpai, A. Tiwari, D. Joshi, and R. Khatavkar, “AbnormNet: A neu-
dicting lower limb joint kinematics using wearable motion sensors,” Gait ral network based suggestive tool for identifying gait abnormalities in
Posture, vol. 28, no. 1, pp. 120–126, 2008. cerebral palsy children,” in Proc. Int. Conf. Advance. Technol., 2022,
[9] T. Li, L. Wang, J. Yi, Q. Li, and T. Liu, “Reconstructing walking dynamics pp. 1–5.
from two shank-mounted inertial measurement units,” IEEE/ASME Trans.
Mechatronics, vol. 26, no. 6, pp. 3040–3050, Dec. 2021.
[10] C. Yi et al., “Continuous prediction of lower-limb kinematics from multi-
modal biomedical signals,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 32, no. 5, pp. 2592–2602, May 2022.
[11] K.-R. Mun, G. Song, S. Chun, and J. Kim, “Gait estimation from anatomi-
cal foot parameters measured by a foot feature measurement system using Rishabh Bajpai (Graduate Student Member,
a deep neural network model,” Sci. Rep., vol. 8, no. 1, pp. 1–10, 2018. IEEE) received the B.Tech. degree in mechani-
[12] S. Sivakumar, A. Gopalai, D. Gouwanda, and K. H. Lim, “Estimation of cal engineering from the PDPM Indian Institute
joint angle from ground reaction force in human gait,” in Proc. IEEE-EMBS of Information Technology, Design and Manu-
Conf. Biomed. Eng. Sci., 2018, pp. 623–628. facturing Jabalpur, Jabalpur, Madhya Pradesh,
[13] S. Sivakumar, A. A. Gopalai, K. H. Lim, D. Gouwanda, and S. Chauhan, India, in 2020. He is currently pursuing the Ph.D.
“Joint angle estimation with wavelet neural networks,” Sci. Rep., vol. 11, degree with the Centre for Biomedical Engineer-
no. 1, pp. 1–15, 2021. ing, IIT Delhi, New Delhi, India, and the All In-
[14] M. Fonseca, X. Gasparutto, F. Leboeuf, R. Dumas, and S. Armand, “Impact dia Institute of Medical Sciences (AIIMS) Delhi,
of knee marker misplacement on gait kinematics of children with cerebral New Delhi, as a Prime Minister.s Research Fel-
palsy using the conventional gait model—A sensitivity study,” PLoS One, low (PMRF).
vol. 15, no. 4, 2020, Art. no. e0232064. He works in the area of brain.computer interface, rehabilitation en-
[15] M. N. H. Yunus, M. H. Jaafar, A. S. A. Mohamed, N. Z. Azraai, and M. S. gineering, neuroscience, electroencephalography, and development of
Hossain, “Implementation of kinetic and kinematic variables in ergonomic biomedical instrumentation for applications specific to assistive devices
risk assessment using motion capture simulation: A review,” Int. J. Environ. for disabled.
Res. Public Health, vol. 18, no. 16, 2021, Art. no. 8342.
[16] R. Bajpai, A. Tiwari, A. Jain, and D. Joshi, “A novel instrumented outsole
for real-time foot kinematics measurement: Validation across different
speeds and simulated foot landing in cerebral palsy gait,” IEEE Trans.
Instrum. Meas., vol. 71, May 2022, Art. no. 2003310.
[17] A. Tiwari and D. Joshi, “Template-based insoles for the center of pressure
estimation in different foot sizes,” IEEE Sens. Lett., vol. 4, no. 8, Aug. 2020,
Art. no. 6001804, doi: 10.1109/LSENS.2020.3010373. Deepak Joshi (Member, IEEE) received the
[18] R. Bajpai and D. Joshi, “Dynamic musculoskeletal simulation of a passive Ph.D. degree in biomedical engineering from the
exoskeleton for simulating contracture,” in Proc. 44th Annu. Int. Conf. Indian Institute of Technology (IIT) Delhi, New
IEEE Eng. Med. Biol. Soc., 2022, pp. 877–880. Delhi, India, in 2012.
[19] S. Gulati and V. Sondhi, “Cerebral palsy: An overview,” Indian J. Pedi- He is currently a faculty with the Centre for
atrics, vol. 85, no. 11, pp. 1006–1016, 2018. Biomedical Engineering, IIT Delhi, and the All
[20] R. Bajpai and D. Joshi, “MoveNet: A deep neural network for joint India Institute of Medical Sciences Delhi, New
profile prediction across variable walking speeds and slopes,” IEEE Trans. Delhi. He has been working in the area of
Instrum. Meas., vol. 70, Apr. 2021, Art. no. 2508511. neuroprosthetics and neurorehabilitation for last
[21] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” nearly 14 years. Before joining to IIT Delhi, he
2014, doi: 10.48550/arXiv.1412.6980. worked at various places including the National
[22] T. Cloete and C. Scheffer, “Benchmarking of a full-body inertial motion University of Singapore, Newcastle University U.K., and University of
capture system for clinical gait analysis,” in Proc. 30th Annu. Int. Conf. Oregon, USA. His current research interests combine experimental and
IEEE Eng. Med. Biol. Soc., 2008, pp. 4579–4582. computational techniques to understand the neural correlates during
[23] S. Tadano, R. Takeda, and H. Miyagawa, “Three dimensional gait anal- balancing and seamless transitions during walking. Besides that, he
ysis using wearable acceleration and gyro sensors based on quaternion is actively engaged in projects related to the development of wearable
calculations,” Sensors, vol. 13, no. 7, pp. 9321–9343, 2013. devices for applications specific to the diagnosis of neuromuscular dis-
[24] T. Seel, J. Raisch, and T. Schauer, “IMU-based joint angle measurement orders, assistive devices for the elderly and disabled, and biofeedback
for gait analysis,” Sensors, vol. 14, no. 4, pp. 6891–6909, 2014. for rehabilitation in stroke patients.

Authorized licensed use limited to: MAULANA AZAD NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2023 at 11:02:41 UTC from IEEE Xplore. Restrictions apply.