See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315787513

Classification of Lunge Biomechanics with Multiple and Individual Inertial

Measurement Units

Article  in  Sports Biomechanics · April 2017

DOI: 10.1080/14763141.2017.1314544

CITATIONS READS

6 1,457

5 authors, including:

Martin O'Reilly Darragh Whelan

University College Dublin University College Dublin
32 PUBLICATIONS   104 CITATIONS    17 PUBLICATIONS   85 CITATIONS   

SEE PROFILE SEE PROFILE

Eamonn Delahunt Tomas E Ward

University College Dublin Dublin City University
149 PUBLICATIONS   3,584 CITATIONS    253 PUBLICATIONS   2,880 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The Marigot Osteoarthritis Nutritional Intervention (MOANi) trial View project

Utilization of inertial measurement units to analyse lower limb movement in athletes with chronic ankle instability during sports related tasks View project

All content following this page was uploaded by Martin O'Reilly on 06 February 2018.

The user has requested enhancement of the downloaded file.

 Classification of Lunge Biomechanics with Multiple and
Individual Inertial Measurement Units

Martin A. O’Reilly1, 2, Darragh F. Whelan1, 2, Tomas E. Ward3, Eamonn Delahunt2,

Brian Caulfield 1, 2

1
Insight Centre for Data Analytics, University College Dublin, Ireland

2
School of Public Health, Physiotherapy and Sports Science, University College
Dublin, Ireland

3
Insight Centre for Data Analytics, Maynooth University, Ireland

Corresponding author:

Name: Martin O’Reilly

Address: Insight UCD,

Science Centre EAST

Belfield

Dublin 4

EIRE

Email: martin.oreilly@insight-centre.org
 1. Abstract

Lunges are a common, compound lower-limb resistance exercise. If

completed with aberrant technique, the increased stress on the joints used may
increase risk of injury. This study sought to first investigate the ability of inertial
measurement units (IMUs), when used in isolation and combination, to (a)
classify acceptable and aberrant lunge technique (b) classify exact deviations in
lunge technique. We then sought to investigate the most important features and
establish the minimum number of top-ranked features and decision trees that are
needed to maintain maximal system classification efficacy. Eighty volunteers
performed the lunge with acceptable form and 11 deviations. Five IMUs
positioned on the lumbar spine, thighs and shanks recorded these movements.
Time and frequency domain features were extracted from the IMU data and used
to train and test a variety of classifiers. A single IMU system achieved 83%
accuracy, 62% sensitivity and 90% specificity in binary classification and a five
IMU system achieved 90% accuracy, 80% sensitivity and 92% specificity. A five
IMU set-up can also detect specific deviations with 70% accuracy. System
efficiency was improved and classification quality was maintained when using
only 20% of the top-ranked features for training and testing classifiers.

Keywords: Wearable sensors, biomedical technology; lower extremity; inertial

measurement units

2. Introduction

The lunge is a compound full-body exercise commonly used in strength and

conditioning (S&C), injury risk screening and rehabilitation. The movement is weight
bearing and promotes muscle activation patterns similar to those of gait (Riemann,
Lapinski, Smith, & Davies, 2012), while also being critical in certain sporting activities
(Cronin, Mcnair, & Marshall, 2003). The lunge is also an evaluative movement to
assess an athletes' risk of incurring lower limb musculoskeletal injury as it is
representative of lower extremity function during activity (Powden, Hoch, & Hoch,
2015). Furthermore, it is an excellent rehabilitative exercise because it can be completed
at home and requires no equipment. As such it is used for rehabilitating conditions such
as patellofemoral pain (Escamilla, 2001) and following ligament injury (Alkjær,
Simonsen, Magnusson, Aagaard, & Dyhre-Poulsen, 2002).

The challenging nature of the lunge means it is difficult to maintain good technique
throughout the movement, which is important to reduce stress on knee, hip and ankle
joints (Farrokhi et al., 2008). Clinical staff such as physiotherapists and S&C coaches
can provide technique feedback to maintain acceptable lunge technique. This is often
completed using two distinct methods; sophisticated biomechanical assessment or
subjective examination. Both of these have a number of limitations. The use of 3-D
motion capture systems is expensive and the application of skin-mounted markers may
hinder normal movement (Ahmadi et al., 2014; Bonnechere et al., 2014). Furthermore,
data processing can be time intensive and specific expertise is often required to interpret
the processed data and to make recommendations on the observed results. Therefore,
these systems are not frequently used to assess lunge technique beyond the research
laboratory (Bonnet, Mazza, Fraisse, & Cappozzo, 2013). In a clinical setting, subjective
evaluation has the potential for bias and poor to moderate reliability (Whatman, Hing, &
Hume, 2012). This may be due to the experience of the rater, the method used to rate
performance of the exercise (ordinal vs. dichotomous scales) or the instructions given to
the raters (Chmielewski et al., 2007; Whatman et al., 2012). As the lunge has a large
number of potential deviations (Table 1), subjective analysis is difficult and often
flawed.

Wearable inertial measurement units (IMUs) offer the potential for low cost, objective
biomechanical analysis that can be completed in a clinical environment. Self-contained,
wireless IMU devices are easy to set-up and allow for the acquisition of human
movement data in unconstrained environments (McGrath, Greene, O’Donovan, &
Caulfield, 2012). These IMUs are small, inexpensive sensors that consist of
accelerometers, gyroscopes and magnetometers. They are able to acquire data
pertaining to the linear and angular motion of individual limb segments and the centre
of mass of the body. In this paper the term IMU system will be used to describe the
IMU sensors, the IMU signals, the associated signal processing applied to them and the
output of the exercise classification algorithms. These IMU systems are ideally suited to
quantify performance of these movements in a clinical setting, as they are not hampered
by location, occlusion and lighting issues unlike other biomechanical analysis tools
(Morris, Saponas, Guillory, & Kelner, 2014).

A growing body of scientific literature has investigated the ability of IMU systems to
assess technique in order to provide this holistic exercise analysis (Giggins, Sweeney, &
Caulfield, 2014; Melzi, Borsani, & Cesana, 2009; O'Reilly et al., 2015; Pernek, Kurillo,
Stiglic, & Bajcsy, 2015; Taylor, Almeida, Hodgins, & Kanade, 2012; Velloso, Bulling,
Gellersen, Ugulino, & Fuks, 2013; Whelan, O'Reilly, Ward, Delahunt, & Caulfield,
2015, 2016). However there is minimal work published evaluating the ability of an IMU
system to accurately quantify lunge biomechanics. Fitzgerald et al. (2007) used a
system that involved ten inertial sensors incorporated within a body suit to
automatically monitor an individual's lunge and tested the system using a case study.
Visual analysis of the IMU gyroscope signals identified lower limb movement
deviations in the injured athlete when compared to the non-injured athlete. Leardini et
al. (2014) and Tang et al. (2015) examined IMUs potential to track body segments
while completing the lunge. Both concluded that they had good accuracy compared to a
laboratory based optical measurement system such as ViconTM. Chen, Jafari, and
Kehtarnavaz (2015) found that combining IMUs with the Microsoft KinectTM, lunges
could be identified with 100% accuracy. Gowing et al. (2014) demonstrated IMUs
could identify a range of movements with greater accuracy than the Microsoft Kinect
(91% overall accuracy), however no specific results were given for the lunge. To our
knowledge no study has analysed the ability of an IMU based system to objectively
analyse lunge technique.

In summary, the lunge is a compound full-body exercise that is included in

resistance training, rehabilitation programs and musculoskeletal injury risk screening
protocols. To date, biomechanical analysis of the movement has come in the form of
expensive lab based evaluation or subjective clinical evaluation. Both of these have
inherent disadvantages and there is a need for biomechanical analysis of the lunge that
is easy to use, low-cost and provides objective data. IMUs may provide a platform to
achieve this. The research question this study seeks to address is: ‘How well can an
IMU-based system quantify lunge technique?’ We hypothesise that (a) IMUs, used in
combination or isolation, positioned on the lumbar spine, thighs and shanks, may be
capable of distinguishing between acceptable and aberrant lunge technique; (b) these
IMU systems may also be capable of detecting specific deviations from acceptable
lunge technique; (c) that system efficiency may be improved, whilst maintaining
accuracy, by finding a selection of most important features to train the classification
algorithms and through optimising other classifier settings.

3. Methods

This study was undertaken to determine the minimal IMU sensor set that can
discriminate between different levels of lunge performance and identify aberrant
exercise technique. Data were acquired from participants as they completed the lunge
with acceptable technique for ten repetitions. IMU data were then acquired while three
repetitions of the same exercise were completed with eleven commonly observed
deviations from acceptable technique.

Participants

Eighty healthy volunteers (57 males, 23 females, age: 24.7 ± 4.9 years, height: 1.75 ±
0.09 m, body mass: 76.0 ± 13.3 kg) were recruited for the study. No participant had a
current or recent musculoskeletal injury that would impair his or her lunge performance.
All participants had prior experience with the exercise and completed it regularly as part
of their own training regime for at least one year. Each participant signed a consent
form prior to completing the study. The University Human Research Ethics Committee
approved the study protocol.

Exercise Technique and Deviations

Participants completed the initial lunge with acceptable technique as described by the
National Strength and Conditioning Association (NSCA) guidelines (Baechle & Earle,
2004). This involved participants placing their left foot in front of the torso and right
foot behind the torso with toes pointing forward and the torso kept upright. This torso
position was maintained throughout the movement. The downward phase then started
from this position. The leading left knee was flexed in order to lower the trailing right
knee toward the floor. The lead knee was kept directly over the lead foot that remained
on the floor. The leading knee continued to flex until it was roughly 90 degrees
perpendicular with the lower leg in the sagittal plane and the trailing knee was 3-6 cm
above the floor. The upward phase immediately followed, whereby the lead knee
extended back to starting position whilst an upright torso posture was maintained.

Table 1 below shows deviations completed by participants in this study. In this study
there were no controls of the severity each participant completed the deviations with.

*Table 1 about here*

Experimental Protocol

When participants arrived to the laboratory the testing protocol was explained to them.
Following this they completed a ten minute warm-up on an exercise bike maintaining a
power output of 100 W at 75-85 revolutions per minute. Next, IMUs were secured on
the participant at the following five locations; the spinous process of the 5th lumbar
vertebra, the mid-point of both femurs (determined as half way between the greater
trochanter and lateral femoral condyle), and on both shanks 2 cm above the lateral
malleolus (Figure 1). The orientation and location of the IMUs were consistent for all
study participants.

*Figure 1 about here*

A pilot study was used to determine an appropriate sampling frequency and the ranges
for the accelerometer and gyroscope on board the IMUs (Shimmer 3, Shimmer, Dublin,
Ireland). In the pilot study squat data was collected at 512 Hz. A Fourier transform was
then used to detect the characteristic frequencies of the signal which were all found to
be less than 20 Hz. Therefore, a sampling frequency of 51.2 Hz was deemed appropriate
for this study based upon the Nyquist criterion and the Shannon sampling theorem
(Jerri, 1977). The Shimmer IMU was configured to stream tri-axial accelerometer (+/-2
g), gyroscope (+/-500 o/s) and magnetometer (+/-1 Ga) data with the sensor ranges
chosen also based upon data from the pilot study. The IMU was calibrated for these
specific sensor ranges using the Shimmer 9DoF Calibration application.

Participants were then instructed on how to complete the lunge with acceptable
technique and biomechanical alignment as outlined in the NSCA guidelines and
explained in the ‘Exercise Technique and Deviations’ section. They completed ten
repetitions with this acceptable technique. Once the lunge had been completed with
acceptable technique the participant was instructed to complete the exercise with the
deviations specified in Table 1. They completed three repetitions of each deviation as
required. Verbal instructions and a demonstration were provided to all participants and
they completed trial repetitions to ensure they were comfortable completing the
deviations. All lunges were completed using body weight only. A chartered
physiotherapist and an S&C trained individual were present throughout all data
collection to ensure the lunge had been completed as instructed. If the movement
completed was not in accordance with the description in Table 1, the participants were
asked to repeat the exercise.

Data Analysis

Nine signals were obtained from each inertial measurement unit; accelerometer x, y, z
gyroscope x, y, z and magnetometer x, y, z. The direction of each of the above axes is
relative to the IMUs. A precise description of these axes and signals can be found in
Shimmer's online documentation for users
(http://www.shimmersensing.com/support/wireless-sensor-networks-documentation).
These signals were low-pass filtered at fc=20 Hz using a Butterworth filter of order n =
8 to remove high frequency noise and ensure all data analysed related to each
participant's movement. In particular, this filter helped remove high frequency, high
amplitude spikes in the magnetometer data that occur due to electromagnetic
interference. Six additional signals were then computed. The 3-D orientation of the
inertial measurement unit was computed from the accelerometer, gyroscope and
magnetometer signals using the gradient descent algorithm as developed by Madgwick ,
Harrison, & Vaidyanathan (2011) which resulted in the quaternion W, X, Y and Z
signals. The acceleration magnitude was also computed from the accelerometer x, y and
z signals. Finally the gyroscope magnitude was computed from the gyroscope x, y and z
signals.

Each repetition from each exercise was extracted from the IMU data and resampled to a
length of 250 samples in order to facilitate for where participants had completed lunges
at different tempos. Descriptive features were then computed from the aforementioned
15 signals. Time-domain and frequency-domain descriptive features were computed in
order to describe the pattern of each of the 15 signals when the exercise was completed.
The features were chosen based on a visual analysis of the data and similar work
completed previously in the field (Giggins et al., 2014; O'Reilly et al., 2015; Whelan et
al., 2015). These features were namely; signal peak, valley, range, mean, standard
deviation, skewness, kurtosis, signal energy, level crossing rate, variance, 25th
percentile, 75th percentile, median and the variance of both the approximate and
detailed wavelet coefficients using the Daubechies 5 mother wavelet to level 6. This
resulted in 16 features for each of the 16 available signals producing a total of 240
features per IMU.

Feature importance was calculated to reduce model complexity and develop an

understanding of which of the aforementioned features are most valuable for correct
classification outcomes. This in turn could lead to a more efficient end user application,
increasing the rate of feedback to the user and improving the computational efficiency
off the system. We used the method described by Liaw and Wiener (2002) to establish
variable importance. Initially benchmark accuracy was established through using all of
the computed features to train and test a random forests classifier. The process was to
then permute the values of each feature and measure how much the permutation
decreases the accuracy of the model. For unimportant features, the permutation should
have little to no effect on model accuracy, while permuting important features should
significantly decrease it. Following the permutation of each feature, features were
ranked based on their importance with the most important feature being that which
caused the largest reduction in accuracy when permuted and so on.
The random-forests method was employed to perform classification (Breiman, 2001).
During analysis several types of classifiers were tested including k-nearest neighbours,
support vector machines and naïve Bayes classifiers, however none were shown to
provide improved results on this dataset and some of these classifiers increased
computational time required. The random-forests method was therefore chosen due to
its superior classification results and efficiency versus these other methods
classification. Random forests classifiers are also easily implemented and reproduced.
The number of trees used in an end-user system will depend on application
requirements, where fewer trees will increase system efficiency but reduce classification
efficacy. In order to investigate this relationship, for the IMU based lunge technique
classification system, the out of bag error of a random forest was computed when using
1 to 500 trees. Out of bag error, is a method of measuring the prediction error of random
forests and other machine learning models. It uses bootstrap aggregating to sub-sample
data sampled used for training. OOB is the mean prediction error on each training
sample xᵢ, using only the trees that did not have xᵢ in their bootstrap sample (James,
Witten, Hastie, & Tibshirani, 2013). This process was completed for both binary and
multi-class lunge technique classification systems. A lower out of bag error score
suggests a lower probability of misclassification but an increased number of trees will
decrease computational efficiency.

Classifiers were initially developed using 400 trees and all available features. They were
evaluated for the following ten combinations of variables; the 1200 (5x240) variables
computed from every IMU, the 720 (3x240) variables from both shanks and the Lumbar
IMU, the 720 variables from both thighs and the Lumbar IMU, the 480 (2x240)
variables from both shanks, the 480 variables from both thighs and the 240 variables
from each of the five individual IMUs. This was to compare classification scores using
five IMUs, three IMUs, two IMUs and one IMU at different anatomical locations and to
establish model accuracy without considering efficiency due to feature selection.

Classifiers were then developed using subsets of features based on their ranked
importance which was computed using the aforementioned random-forests variable
permutation method (Liaw & Wiener, 2002). For each IMU combination, the features
ranked in the top 2% were first used as training and test data. The accuracy of the
system was determined using this reduced set of variables. Following this, the accuracy
of the system was determined when using the top 4% of features and so on in
increments of 2% until all variables were used again. The accuracy was plotted versus
the percentage of variables used in order to build an understanding of the relationship of
model efficiency and accuracy.

Initially binary classification was evaluated to establish how effectively each individual
IMU and each combination of IMUs could distinguish between acceptable and aberrant
lunge technique. All repetitions of acceptable performance of the lunge were labelled '0'
and all repetitions of the lunge performed with one of the deviations as outlined in Table
1 were labelled '1'. Multi-label classification was then evaluated on the IMU data set to
investigate how effectively each individual IMU and each combination of IMUs could
be used to discriminate between correct performance of the lunge exercise and each of
the eleven deviations from correct technique as described in Table 1. All repetitions of
normal performance of the lunge remained labelled as '0' and each of the different
deviations were labelled '1-11'.

Identifying the type of deviation occurring could allow for an automated biofeedback
system to give more specific feedback to a user than a binary classifier is capable of.
Moreover, they may require less IMUs to perform effective classification than a
classifier that identifies the exact deviation in lunge technique that a user is making. In
order to investigate this idea, the twelve classes were grouped as described in Table 2,
to establish the ability of each IMU set-up to identify the type of deviation that was
occurring. The four groups were chosen by a Chartered Physiotherapist and a person
trained in strength and conditioning following discussions with clinical professionals.
They pertain to the users' lower-limb alignment throughout each repetition of the
exercise, the user's starting position and their lumbar tracking.
 *Table 2 about here*

All of the above classifiers were developed using all computed features and subsets of
the top features. System quality for both sets of classifiers was compared using the
accuracy, sensitivity and specificity metrics.

The quality of the lunge exercise classification was established using leave-one-subject-
out-cross-validation (LOSOCV) and the random-forests classifier with 400 trees. Each
participant's data corresponds to one fold of the cross validation. At each fold, one
participant's data is held out while the random forests classifier is trained and then this
held out data is used to assess the classifier's ability to correctly categorise new data it is
presented with. The use of LOSOCV ensures that there is no biasing of the classifiers,
whereby the test subjects' data is completely unseen by the classifier prior to testing.
Previous research has shown that not employing this method of testing can greatly skew
results (Taylor, Almeida, Kanade, & Hodgins, 2010). In our system each individual
repetition was classified.

The scores used to measure the quality of classification were total accuracy, average
sensitivity and average specificity. Accuracy is the number of correctly classified
repetitions of all the exercises divided by the total number of repetitions completed; this
is calculated as the sum of the true positives (TP) and true negatives (TN) divided by
the sum of the true positives, false positives (FP), true negatives and false negatives
(FN):

𝑇𝑃+𝑇𝑁
1) 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃+𝐹𝑃+𝑇𝑁+𝐹𝑁

In binary classification the '0' label was considered positive and '1' label as negative.
Therefore a single sensitivity and specificity score is suitable to describe the quality of
the various IMU combinations. In multi-class classification, for every IMU combination
evaluated the sensitivity and specificity were calculated for each of the twelve
deviations and then the mean and standard deviation across the twelve values was taken,
using the formulas below:

𝑇𝑃
2) 𝑆e𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑇𝑃+𝐹𝑁

𝑇𝑁
3) 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑁+𝐹𝑃

Sensitivity measures the effectiveness of a classifier at identifying a desired label, while

specificity measures the classifiers ability to detect negative labels. A 'positive' label is
the desired label. For example in binary classification a lunge completed with correct
form is considered the positive class.

In reviewing the accuracy, sensitivity and specificity scores produced by each classifier,
80% or higher was considered a 'good' quality result, 60-79% was considered a
'moderate' result and anything less than 59% was deemed a poor result. The authors
chose these values after reviewing existing literature on identifying deviations from
acceptable exercise performance using data derived from IMUs. In reviewing such
literature, an existing accepted standard for a good, moderate or poor classifier could
not be found. Therefore, the above system was agreed on by the authors to facilitate
interpretation of our range of results.

4. Results

Results are first presented from the classifiers trained with all available features. The
most important features for model development are then presented. The subsequent
effect of feature selection on classification is then presented.

Classification with Full Feature Set

Total accuracy scores for each IMU position are shown in Table 3. A single IMU set-up
is able to detect whether a lunge is completed correctly with 72-82% accuracy but can
only detect the exact mistake a user is making with 35-40% accuracy. A multi-IMU set-
up using 5 IMUs worn on the lumbar, both thighs and both shanks is capable of
distinguishing between good and bad performance of the lunge with 90% accuracy.
Furthermore, this IMU set-up can detect the exact mistake a user is making from the list
shown in Table 1 with 70% accuracy.

*Table 3 about here*

Sensitivity and specificity scores are shown for all IMU combinations in Table 3.
Moderate sensitivity and specificity scores are achieved for a single IMU system in
Binary classification. The IMU on the right thigh is capable of 78% sensitivity and 83%
specificity. The left shank was the poorest position for binary classification using a
single IMU with a sensitivity of 40% and specificity of 82%. The five IMU set-up is
most effective for both binary and multi-class classification. A reduced IMU set using
three IMUs positioned on the lumbar and both shanks also produces good classification
accuracy.

Figures 2 and 3 are confusion matrices showing the exact percentage of repetitions
correctly and incorrectly classified. The rows demonstrate the class the rep actually
belongs to and the columns show which class the classifier outputted. In both figures the
top left shows the percentage of TPs, top right: FNs, bottom right: TNs and bottom left:
FNs. The true positive rate for the single IMU on the right thigh is 66% and for the five
IMU set is 71%. The true negative rate for the single IMU on the right thigh is 88% and
for the five IMU set is 95%.

*Figure 2 about here*

*Figure 3 about here*

Figure 4 is a confusion matrix for identifying the type of deviation occurring using the
IMUs on the lumbar and both shanks. Mistakes pertaining to lumbar tracking are best
detected with an 89% rate. Mistakes relating to lower limb alignment and starting
position are the most mutually confused categories.
 *Figure 4 about here*

Figure 5 shows a confusion matrix for multi-class classification (exact deviation) when
using the 5 IMU set-up. 'Normal' (N) performance of the squat is detected with a 68%
TP rate. 6% of normal lunges were confused by the classifier to be 'step too short' (STS)
and 7% of normal lunges were mistaken to be 'step too long' (STL). The 'push back'
(PB) deviation is the worst detected deviation with a TP rate of just 46%, almost 10% of
PB reps were mistaken to be 'normal' (N) by the system. The best detected deviation
was 'stutter step' (SS) with a TP rate of 81%.

*Figure 5 about here*

Feature Importance

Table 4 highlights the 15 most important features for each type of classifier evaluated.

*Table 4 about here*

Classification with Top-Ranked Features

Figure 6 shows the quality of multi-class classification of exact deviation when using 1-
100% of features ranked in order of importance. It is evident using just 20% of the top
ranked features produces similar classification quality to using 100% of the features.
Similarly, Figure 7 demonstrates the quality of binary classification when using 1-100%
of the features computed from the IMU located on the right thigh. A similar trend in
accuracy, sensitivity and specificity scores was evident across all the classification
problems discussed in this paper when using just 20% of the most important features.

*Figure 6 about here*

*Figure 7 about here*

Number of Trees Used Versus Out of Bag Error

Figure 8 demonstrates the out of bag error versus number of trees used in the random
forests classifiers. It is evident that the cost (computational efficiency)- benefit
(improved classification scores) of using increased tree numbers starts to diminish after
utilising 100 to 150 trees for classification.

*Figure 8 about here*

5. Discussion and Implications

In relation to hypothesis (a), the results presented in this paper show that a system
consisting of 5 lower limb IMUs can distinguish between acceptable and aberrant lunge
technique with 90% accuracy, 80% sensitivity and 92% specificity. A system based on
data derived from 3 IMUs achieves 87% accuracy, 73% sensitivity and 91% specificity.
The similar classification scores with a reduced IMU set-up arise because the features
from both thighs and both shanks are likely to correlate. This provides minimal
additional information to aid classification quality. The system using data solely from
the right thigh IMU achieved 82% accuracy, 78% sensitivity and 83% specificity. This
IMU position vastly outperforms the others when used in isolation, particularly when
considering the sensitivity achieved. This is likely because this position on the trailing
leg for the lunges studied captures the differences between acceptable and aberrant
lunges most optimally as at the other IMU positions, the inter-subject variability of how
lunges are performed exceeds the intra-subject variability between acceptable and
aberrant lunges to a greater extent.

In examining hypotheses (b) of this study, it has been shown that multi-label
classification of specific deviations is possible with 70% accuracy, 70% sensitivity and
97% specificity using data from the full 5 IMU set up. A 3 IMU set up may provide
increased system practicality and efficiency compared to a 5 IMU set-up, however its
accuracy and sensitivity drop by 8-10%. A system consisting of two IMUs did not
produce favourable results for multi-label classification. These results are likely to be
related to amount of classes the system is attempting to identify (a total of 12 as shown
in Table 1). Furthermore, these deviations manifest more severely at specific anatomical
points, making it difficult for a reduced IMU set-up to identify. When using reduced
IMU set ups for detecting these specific mistakes, it is likely the features will be less
discriminative between each class when only coming from a single IMU. It should also
be noted that the number of instances of each class in multi-label classification is
considerably lower than that used for the binary classifiers where repetitions from 11
classes of deviations were pooled together to make one larger ‘aberrant’ class,
negatively impacting multi-label classification scores. Interestingly, when pooling types
of deviations together to create larger classes (Table 2), an increase of 4-16% accuracy
was observed. This may be due to the reduced likelihood of classifiers ‘confusing’ these
deviations or because each class now contains more instances than were available for
specific multi-label classification. However, two and single IMU set-ups still showed
poor sensitivity and accuracy with this form of classification. Subsequently it would be
difficult to implement effectively in a real world environment.

This work also identifies two methods of increasing system efficiency corresponding
with hypothesis (c) of this study. Figure 6 and Figure 7 show the ability of a reduced
feature selection to obtain similar accuracy, sensitivity and specificity scores compared
to using all features. In these cases, using just 20-30% of top-ranked features to train
and evaluate the random forests classifiers produces equivalent accuracy, sensitivity and
specificity results. Figure 8 also demonstrates that using a reduced number of trees in
each random forests classifier can maintain the misclassification probability as
determined by out of bag error. It is difficult to draw conclusions as to how many trees
and what proportion of top ranked features should be used by people implementing
systems such as those described in this paper. These decisions should be made
considering the specific application of the systems, the computational processing
devices available (e.g. on device or cloud based analysis) and a domain specific
judgement on the accuracy-efficiency trade off. It is however shown that a broad range
of choices will result in near equivalent classification scores.
The top-ranked features shown in Table 4 demonstrate which features are most
important for each type of classifier. There is a large degree of variety in the important
IMU positions, signals and feature types (time domain, frequency domain and time-
frequency domain). As such, it would be difficult to predict which signals and features
are of most importance in advance of creating IMU based exercise analysis systems for
other exercises. The authors would recommend following a similar process of
computing a large number of diverse features from a variety of signals and then
identifying the most important features for each specific application. Interestingly a
number of features from the magnetometer and quaternion signals are highly ranked.
The authors note that these signals were not analysed in similar research (Giggins et al.,
2014; Pernek, Hummel, & Kokol, 2013; Taylor et al., 2012; Velloso et al., 2013) and
that perhaps they could be beneficial to include in future IMU based exercise analysis
systems.

It is difficult to directly compare results with previous work in the area due to
differences in exercises investigated, IMU positions and overall size of datasets
analysed. The majority of research to date has investigated the ability of IMU systems
to monitor technique in simple exercises such as straight leg raises (Taylor et al., 2012),
dumbbell curls (Velloso et al., 2013) or heel slides (Giggins et al., 2014). This work
builds on previous research by analysing lunge form. The lunge is a compound, non-
symmetrical exercise making it more difficult to track than single joint exercises. While
some authors have tracked the lunge using IMUs (Fitzgerald et al., 2007; Leardini et al.,
2014; Tang et al., 2015), none have quantitatively assessed technique using the IMUs.
The complex nature of the lunge means a number of mistakes can occur at a variety of
joint positions, making it more difficult to ascertain exact deviations using an IMU
sensor set. This explains the moderate to low sensitivity and accuracy scores for multi-
label classification seen in Table 3.

The lower sensitivity and accuracy scores in Table 3 for multi label classification are
especially prevalent in reduced IMU set-ups (less than three IMUs). A minimal IMU
set-up is advantageous as it is less cumbersome, easier to operate and reduces the risk of
sensor placement error. Furthermore, it would reduce cost for end users. A number of
authors have evaluated exercise performance with a single IMU, reporting higher
overall accuracy scores than this study (Giggins et al., 2014; O'Reilly et al., 2015;
Pernek et al., 2013; Whelan et al., 2015). The scores presented here for a single IMU
performing multi-class recognition may be lower for a number of reasons. The lunge is
a multi-joint lower limb exercise unlike those presented in Giggins et al. (2014) and
Pernek et al. (2013). The use of multiple joints makes it more difficult to detect
deviations with a minimal IMU set-up due to the increased number of possible
deviations and additional complexity of exercise biomechanics. Previous work (O'Reilly
et al., 2015; Whelan et al., 2015) evaluated multi-joint exercises (squat and single leg
squat) using a single IMU set-up and found moderate to good overall accuracy scores.
However, the number of potential deviations that the authors attempted to classify with
the IMUs was reduced compared to the eleven deviations shown in this paper (Table 1).
This makes it more difficult for the IMUs to classify each deviation, explaining some of
the lower scores shown in Table 3.

While this paper does show encouraging results for using an IMU system to track lunge
biomechanics, there are a number of contextual factors that must be considered. Firstly,
the dataset collected and analysed in this study may be limited in its transferability to
real-world applications as all exercise deviations were deliberately induced. When
deviations occur naturally, the exact way in which they present may differ from the
induced deviations investigated in this study. Moreover, there were no controls as to the
severity with which each participant performed a deviation and therefore it is possible
that naturally occurring lunge deviations in a ‘real life’ application may be more acute,
or occur in a more idiosyncratic fashion than those used in the presented classification
systems. No gold-standard 3-dimensional motion capture system was used to confirm
that each deviation occurred and the 11 deviations studied may be a non-exhaustive list
of lunge deviations. Another potential limitation of the current system in terms of its
transferability to real world application is that IMU positioning was strictly consistent
and local frame IMU data was used. As such the IMU systems presented may only
produce these results when set up and analysed in identical fashion to this study.
Inexperienced users may place the IMUs with incorrect orientation or at different
positions, affecting system accuracy. One such method to combat this could be to
convert the local frame IMU data to a global frame that utilises vertical acceleration.
However, as many of the signals and features used in this study are derived from local
frame IMU data the authors anticipate this could have a negative consequence on
classification quality. Most notably, this may happen where information is lost in
converting a signal from local frame to global frame. In the current system, the
accelerometer x signal measures both inertial and gravitational acceleration with respect
to the x-axis of the IMU, converting signals such as this to acceleration in the global
frame may cause loss of information, for instance that which came from the effect of
gravitational acceleration with respect to the IMU’s x-axis. This limitation requires
further investigation and the authors recommend that IMU positioning and its
importance should be made clear in instructions to all users of the current system.

Future work will involve 3 key areas of focus: (a) addressing the aforementioned
limitations of the system, including developing and evaluating lunge classification
systems with naturally occurring technique deviations (b) improving system accuracy
through collecting a larger dataset and investigating additional analysis methodologies
and (c) evaluating the usability, functionality and perceived impact of IMU based
exercise analysis systems such as that described in this paper. A larger dataset should
improve system accuracy and may allow for the application of deep learning
classification techniques to the data that could further heighten system accuracy. In
order to enable the collection of a larger data set, a tablet based tool is currently being
developed which will allow for exercise professionals to simultaneously collect IMU
and video data, which is then automatically segmented in to exercise repetitions and can
be labelled by the exercise professional. The labelled IMU data can be appended to
existing classifiers such as those described in this paper. This should take data
collection out of the hands of researchers and in to the hands of exercise professionals,
allowing for the creation of larger datasets. A qualitative and quantitative evaluation of
the system described in this paper will also be undertaken in order to establish its
usability, functionality and perceived benefit. This should help inform steps to improve
the development of IMU based exercise analysis systems and subsequently what future
work may also involve.
The lunge is an important movement in S&C, musculoskeletal injury risk screening and
rehabilitation. The ability to objectively quantify lunge biomechanics using low cost
IMU technology would have practical implications in all of these settings. In an S&C
setting the ability to remotely monitor lunge biomechanics may allow for guidance in
how to complete the lunge exercise. This could help in achieving exercise goals and
reducing the risk of injury by allowing coaches to monitor lunge biomechanics and
progression. In musculoskeletal and sports medicine, the lunge is a common screening
test as it is weight bearing and functional (Cook, Burton, & Hoogenboom, 2006;
Powden et al., 2015). An IMU system that can identify aberrant biomechanics would
allow for quicker and more objective injury risk identification and stratification. In a
rehabilitation setting, the lunge is often completed following injury and surgery (Hall et
al., 2015). Monitoring technique would be important to prevent further injury and
possibly allow exercises to be completed at home without the need for constant
supervision, reducing overall healthcare costs. Furthermore, the IMUs would allow the
transfer of exercise data to a cloud-based server. This means therapists can assess
patient compliance and technique during the prescribed exercise without the need for
constant monitoring.

6. Conclusion

The results of this study show that an IMU system is able to classify lunge technique as
acceptable or aberrant with good accuracy using a five, three, two or single IMU set-up.
A five and three IMU set-up can detect specific deviations in a person's lunge
biomechanics with moderate accuracy. This is diminished with a reduced IMU set, even
with a broader grouping of deviations. It is also shown that a more streamlined feature
selection technique has similar outcomes to using all features to train the classifier,
allowing for the possibility of an enhanced user experience. Taken together, these
results suggest that a wearable IMU system has the potential to monitor lunge
biomechanics, which would have important implications in injury risk screening,
strengthening and rehabilitation.
 7. Acknowledgements
This project is partly funded by the Irish Research Council as part of a Postgraduate
Enterprise Partnership Scheme with Shimmer (EPSPG/2013/574) and partly funded by
Science Foundation Ireland (SFI/12/RC/2289). The authors would also like to thank
UCD Sport for providing equipment which was used in this study. The authors Martin
O’Reilly and Darragh Whelan contributed equally to this study.

8. References

Ahmadi, A., Mitchell, E., Destelle, F., Gowing, M., O'Connor, N. E., Richter, C., &
Moran, K. (2014, June). Automatic activity classification and movement
assessment during a sports training session using wearable inertial sensors. In
Proceedings of the 2014 11th International Conference on Wearable and
Implantable Body Sensor Networks (BSN) (pp. 98-103). IEEE.

Alkjær, T., Simonsen, E. B., Magnusson, S. P., Aagaard, H., & Dyhre-Poulsen, P.
(2002). Differences in the movement pattern of a forward lunge in two types of
anterior cruciate ligament deficient patients: copers and non-copers. Clinical
Biomechanics, 17, 586-593.

Baechle, T. R., & Earle, R. W. (2004). Resistance Training Exercise Techniques. In

NSCA's Essentials of Personal Training: Champaign, IL.: Human Kinetics.

Bonnechere, B., Jansen, B., Salvia, P., Bouzahouene, H., Omelina, L., Moiseev, F.,
Van Sint Jan, S. (2014). Validity and reliability of the Kinect within functional
assessment activities: Comparison with standard stereophotogrammetry. Gait &
Posture, 39, 593-598.

Bonnet, V., Mazza, C., Fraisse, P., & Cappozzo, A. (2013). Real-time estimate of body
kinematics during a planar squat task using a single inertial measurement unit.
IEEE Transactions on Biomedical Engineering, 60, 1920-1926.

Breiman, L. (2001). Random forests. Machine learning, 45, 5-32.

Chen, C., Jafari, R., & Kehtarnavaz, N. (2016). A real-time human action recognition
system using depth and inertial sensor fusion. IEEE Sensors Journal, 16, 773-
781.

Chmielewski, T. L., Hodges, M. J., Horodyski, M., Bishop, M. D., Conrad, B. P., &
Tillman, S. M. (2007). Investigation of clinician agreement in evaluating
movement quality during unilateral lower extremity functional tasks: a
comparison of 2 rating methods. Journal of Orthopaedic & Sports Physical
 Therapy, 37, 122-129.

Cook, G., Burton, L., & Hoogenboom, B. (2006). Pre-participation screening: the use of
fundamental movements as an assessment of function - part 1. North American
Journal of Sports Physical Therapy, 1, 62-72.

Cronin, J., Mcnair, P., & Marshall, R. (2003). Lunge performance and its determinants.
Journal of Sports Sciences, 21, 49-57.

Escamilla, R. F. (2001). Knee biomechanics of the dynamic squat exercise. Medicine &
Science in Sports & Exercise, 33, 127-141.

Farrokhi, S., Pollard, C. D., Souza, R. B., Chen, Y.-J., Reischl, S., & Powers, C. M.
(2008). Trunk position influences the kinematics, kinetics, and muscle activity
of the lead lower extremity during the forward lunge exercise. Journal of
Orthopaedic & Sports Physical Therapy, 38, 403-409.

Fitzgerald, D., Foody, J., Kelly, D., Ward, T., Markham, C., McDonald, J., & Caulfield,
B. (2007, August). Development of a wearable motion capture suit and virtual
reality biofeedback system for the instruction and analysis of sports
rehabilitation exercises. In Proceedings of the 29th Annual International
Conference of the IEEE on Engineering in Medicine and Biology Society
(EMBC) (pp. 4870-4874). IEEE.

Giggins, O. M., Sweeney, K. T., & Caulfield, B. (2014). Rehabilitation exercise

assessment using inertial sensors: a cross-sectional analytical study. Journal of
Neuroengineering and Rehabilitation, 11, 158-168.

Gowing, M., Ahmadi, A., Destelle, F., Monaghan, D. S., O’Connor, N. E., & Moran, K.
(2014, January). Kinect vs. low-cost inertial sensing for gesture recognition. In
Proceedings of the 2014 International Conference on Multimedia Modeling (pp.
484-495). Springer International Publishing.

Hall, M., Nielsen, J. H., Holsgaard-Larsen, A., Nielsen, D. B., Creaby, M. W., &
Thorlund, J. B. (2015). Forward lunge knee biomechanics before and after
partial meniscectomy. The Knee, 22, 506-509.

James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). Classification. In An
introduction to statistical learning (Vol. 6). New York: Springer.

Jerri, A. J. (1977). The Shannon sampling theorem—Its various extensions and

applications: A tutorial review. Proceedings of the IEEE, 65, 1565-1596.

Leardini, A., Lullini, G., Giannini, S., Berti, L., Ortolani, M., & Caravaggi, P. (2014).
Validation of the angular measurements of a new inertial-measurement-unit
based rehabilitation system: comparison with state-of-the-art gait analysis.
Journal of Neuroengineering and Rehabilitation, 11, 1-7.

Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R

News, 2, 18-22.

Madgwick, S. O., Harrison, A. J., & Vaidyanathan, R. (2011, June). Estimation of IMU
 and MARG orientation using a gradient descent algorithm. In proceedings of
the 2011 IEEE International Conference on Rehabilitation Robotics (ICORR)
(pp. 1-7). IEEE.

McGrath, D., Greene, B. R., O’Donovan, K. J., & Caulfield, B. (2012). Gyroscope-
based assessment of temporal gait parameters during treadmill walking and
running. Sports Engineering, 15, 207-213.

Melzi, S., Borsani, L. P., & Cesana, M. (2009, June). The virtual trainer: Supervising
movements through a wearable wireless sensor network. In Proceedings of the
6th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad
Hoc Communications and Networks Workshops, 2009 (SECON Workshops)
(pp. 1-3). IEEE.

Morris, D., Saponas, T. S., Guillory, A., & Kelner, I. (2014, April). RecoFit: using a
wearable sensor to find, recognize, and count repetitive exercises. In
Proceedings of the 32nd annual ACM conference on Human factors in
computing systems (pp. 3225-3234). ACM.

O'Reilly, M., Whelan, D., Chanialidis, C., Friel, N., Delahunt, E., Ward, T., &
Caulfield, B. (2015, June). Evaluating squat performance with a single inertial
measurement unit. In Proceedings of the 2015 IEEE 12th International
Conference on Wearable and Implantable Body Sensor Networks (BSN) (pp. 1-
6). IEEE.

Pernek, I., Hummel, K. A., & Kokol, P. (2013). Exercise repetition detection for
resistance training based on smartphones. Personal and Ubiquitous Computing,
17, 771-782.

Pernek, I., Kurillo, G., Stiglic, G., & Bajcsy, R. (2015). Recognizing the intensity of
strength training exercises with wearable sensors. Journal of Biomedical
Informatics, 58, 145-155.

Powden, C. J., Hoch, J. M., & Hoch, M. C. (2015). Reliability and minimal detectable
change of the weight-bearing lunge test: A systematic review. Manual Therapy,
20, 524-532.

Riemann, B. L., Lapinski, S., Smith, L., & Davies, G. (2012). Biomechanical analysis
of the anterior lunge during 4 external-load conditions. Journal of Athletic
Training, 47, 372-378.

Tang, Z., Sekine, M., Tamura, T., Tanaka, N., Yoshida, M., & Chen, W. (2015).
Measurement and Estimation of 3D Orientation using Magnetic and Inertial
Sensors. Advanced Biomedical Engineering, 4, 135-143.

Taylor, P. E., Almeida, G. J., Hodgins, J. K., & Kanade, T. (2012, August). Multi-label
classification for the analysis of human motion quality. In Proceedings of the
2012 Annual International Conference of the IEEE on Engineering in Medicine
and Biology Society (EMBC) (pp. 2214-2218). IEEE.

Taylor, P. E., Almeida, G. J., Kanade, T., & Hodgins, J. K. (2010, August). Classifying
human motion quality for knee osteoarthritis using accelerometers. In
 Proceedings of the 2010 Annual International Conference of the IEEE on
Engineering in Medicine and Biology Society (EMBC) (pp. 339-343). IEEE.

Velloso, E., Bulling, A., Gellersen, H., Ugulino, W., & Fuks, H. (2013, March).
Qualitative activity recognition of weight lifting exercises. In Proceedings of the
4th Augmented Human International Conference (pp. 116-123). ACM.

Whatman, C., Hing, W., & Hume, P. (2012). Physiotherapist agreement when visually
rating movement quality during lower extremity functional screening tests.
Physical Therapy in Sport, 13, 87-96.

Whelan, D., O'Reilly, M., Ward, T., Delahunt, E., & Caulfield, B. (2015, October).
Evaluating performance of the single leg squat exercise with a single inertial
measurement unit. In Proceedings of the 3rd 2015 Workshop on ICTs for
improving Patients Rehabilitation Research Techniques (REHAB) (pp. 144-
147). ACM.

Whelan, D., O'Reilly, M., Ward, T., Delahunt, E., & Caulfield, B. (2016). Evaluating
Performance of the Lunge Exercise with Multiple and Individual Inertial
Measurement Units. In Proceedings of the 10th EAI International Conference on
Pervasive Computing Technologies for Healthcare (pp. 101-108). ACM.
Figure 1. Image showing the lunge exercise and the five sensor positions: (1) The
spinous process of the 5th lumbar vertebra, (2&3) the mid-point of both femurs on the
lateral surface (determined as half way between the greater trochanter and lateral
femoral condyle), (4&5) and on both shanks 2cm above the lateral malleolus
Figure 2. Heatmap confusion matrix showing binary classification results with right
thigh IMU. Rows show the actual class of a repetition and columns show the classifier's
prediction.
Figure 3. Heatmap confusion matrix showing binary classification results with all five
IMUs. Rows show the actual class of a repetition and columns show the classifier's
prediction.
Figure 4. Heatmap confusion matrix showing multi-label classification results
(deviation type) with lumbar & shank IMUs. Rows show the actual class of a repetition
and columns show the classifier's prediction.
Figure 5. Heatmap confusion matrix showing multi-class classification results with all
five IMUs. Rows show the actual class of a repetition and columns show the classifier's
prediction.
Figure 6. Graph showing accuracy, sensitivity and specificity scores in multi-class
classification (exact deviation) with all 5 IMUs when using 1-100% of features ranked
in order of importance.
Figure 7. Graph showing accuracy, sensitivity and specificity scores in binary
classification (acceptable or aberrant technique) with the right thigh IMU when using 1-
100% of features ranked in order of importance.
Figure 8. The relationship between out of bag error and number trees used for the
random forests classifiers when using all training data for binary classification using the
right thigh IMU and all training data for multi-class classification using all 5 IMUs.
Table 1. List and description of Lunge exercise technique deviations used in this study

Deviation Explanation
N Normal lunge
KVL Left knee coming towards mid-line during
downwards phase
KVR Left knee moving away from mid-line during
downward phase
KTF Left knee ahead of toes during downward phase
HSR Excessive lean to left hand side during entire
lunge exercise
HSL Excessive lean to right hand side during entire
lunge exercise
BO Excessive flexion of hip and torso during entire
lunge exercise
BFO Right foot externally rotated
SS Loss of balance during upward phase resulting
stuttered steps
PB Pushing backwards during upwards phase
STS Starting stance too short
STL Starting stance too long
Table 2. Grouped deviations used for classifying the type of deviation a system user may be making

Group Name Numeric Label Deviations in Group

Normal 0 N
Lower Limb Alignment 1 KVL
BFO
KVR
KTF
PB
SS
Starting Position 2 STS
STL
Lumbar Tracking 3 BO
HSL
HSR
Table 3. Total accuracy, sensitivity and specificity scores from LOSOCV for binary and both multi-
Label classifiers

Binary (%) Multi-Label (%) Multi-Label (%)

(Correct or Incorrect) (Type of Deviation) (Specific Deviation)
Acc Sens Spec Acc Sens Spec Acc Sens Spec
All 5 IMUs 90 80 92 67 71 ± 13 90 ± 3 70 70 ± 10 97 ± 1
Lumbar + Shanks 87 73 91 66 69 ± 14 88 ± 5 60 62 ± 13 96 ± 2
Lumbar + Thighs 87 73 91 65 70 ± 13 88 ± 4 61 61 ± 13 96 ± 2
Shanks 84 67 89 56 50 ± 15 85 ± 4 50 49 ± 23 95 ± 2
Thighs 85 69 90 53 52 ± 16 84 ± 1 48 44 ± 2 95 ± 3
Left Shank 72 40 82 46 39 ± 13 81 ± 5 35 38 ± 25 94 ± 3
Left Thigh 77 51 83 44 42 ± 4 82 ± 4 37 39 ± 3 94 ± 3
Lumbar 77 50 87 56 59 ± 19 85 ± 6 40 43 ± 24 95 ± 3
Right Thigh 82 78 83 52 47 ± 18 84 ± 2 39 33 ± 25 94 ± 3
Right Shank 81 58 88 44 42 ± 15 81 ± 5 33 29 ± 24 94 ± 3
 Table 4. Top 15 features for binary and multi-label classification

Rank Binary Multi-Label Multi-Label

(Correct or Incorrect) (Type of Deviation) (Specific Deviation)
1 ’Right Shank - ’Right Thigh - ’Left Shank -
Gyroscope X - Magnetometer X - Gyroscope X-
Fractal Dimension’ 75th Percentile’ Energy’
2 ’Right Shank - ’Left Shank - ’Left Shank -
Roll - Accelerometer Z- Gyroscope Z-
Approx Wavelet Coefficients’ RMS’ Level Crossing Rate’
3 ’Right Thigh - ’Lumbar - ’Right Shank -
Magnetometer X - Acceleration magnitude- Accelerometer Y-
Variance’ Detailed Wavelet Coefficients’ Standard Deviation’
4 ’Right Thigh - ’Right Thigh - ’Left Shank -
Quaternion W - Quaternion Y- Magnetometer X-
Approx Wavelet Coefficients’ Level Crossing Rate’ Kurtosis’
5 ’Left Shank - ’Left Thigh - ’Right Thigh -
Magnetometer Z - Quaternion X- Pitch-
Detailed Wavelet Coefficients’ Max’ Standard Deviation’
6 ’Left Thigh - ’Right Thigh - ’Right Thigh -
Magnetometer Y - Quaternion Y- Gyroscope Magnitude-
Range’ Skewness’ Max’
7 ’Left Thigh - ’Right Shank - ’Right Thigh -
Acceleration magnitude - Accelerometer Z- Magnetometer X-
RMS’ Skewness’ Range’
8 ’Lumbar - ’Left Thigh - ’Right Thigh -
Magnetometer X - Roll- Acceleration magnitude-
Min’ Variance’ Max’
9 ’Lumbar - ’Left Thigh - ’Left Shank -
Quaternion W - Gyroscope Magnitude- Gyroscope Magnitude-
Mean’ 75th Percentile’ Standard Deviation’
10 ’Lumbar - ’Lumbar - ’Left Shank -
Quaternion W - Gyroscope Y- Quaternion X-
RMS’ Variance’ Variance’
11 ’Right Thigh - ’Right Thigh - ’Right Shank -
Accelerometer Y - Roll- Gyroscope Y-
Variance’ Approx Wavelet Coefficients’ Max’
12 ’Right Shank - ’Right Shank - ’Right Shank -
Accelerometer X - Accelerometer Y- Acceleration magnitude-
Level Crossing Rate’ Kurtosis’ Detailed Wavelet Coefficients’
13 ’Right Shank - ’Right Shank - ’Left Shank -
Magnetometer X - Yaw- Gyroscope X-
Energy’ Energy’ Fractal Dimension’
14 ’Left Shank - ’Left Shank - ’Left Shank -
Accelerometer X - Pitch- Accelerometer Z-
Range’ Energy’ Max’
15 ’Left Shank - ’Lumbar - ’Left Thigh -
Magnetometer Z - Quaternion Y- Gyroscope Z-
Range’ Standard Deviation’ Max’

View publication stats