Content uploaded by Bo Zhou
Author content
All content in this area was uploaded by Bo Zhou on Dec 14, 2017
Content may be subject to copyright.
Never Skip Leg Day: A Novel Wearable Approach
to Monitoring Gym Leg Exercises
Bo Zhou, Mathias Sundholm, Jingyuan Cheng, Heber Cruz, and Paul Lukowicz
German Research Center for Artificial Intelligence (DFKI)
and University of Kaiserslautern
Trippstadter Straße 122 , 67663, Kaiserslautern, Germany
{bo.zhou, mathias.sundholm, jingyuan.cheng, heber.cruz zurian, paul.lukowicz}@dfki.de
Abstract—We present a wearable textile sensor system for
monitoring muscle activity, leveraging surface pressure changes
between the skin and an elastic sport support band. The sensor is
based on an 8×16 element fabric resistive pressure sensing matrix
of 1cm spatial resolution, which can be read out with 50fps
refresh rate. We evaluate the system by monitoring leg muscles
during leg workouts in a gym out of the lab. The sensor covers the
lower part of quadriceps of the user. The shape and movement of
the two major muscles (vastus lateralis and medialis) are visible
from the data during various exercises. The system registers the
activity of the user for every second, including which machine
he/she is using, walking, relaxing and adjusting the machines; it
also counts the repetitions from each set and evaluate the force
consistency which is related to the workout quality.
6 people participated in the experiment of overall 24 leg
workout sessions. Each session includes cross-trainer warm-up
and cool-down, 3 different leg machines, 4 sets on each machine.
Plus relaxing, adjusting machines, and walking, we perform
activity recognition and quality evaluation through 2-dimensional
mapping and the time sequence of the average force. We have
reached 81.7% average recognition accuracy on a 2ssliding
window basis, 93.3% on an event basis, and 85.6% spotting
F1-score. We further demonstrate how to evaluate the workout
quality through counting, force pattern variation and consistency.
Index Terms—pressure sensor matrix; wearable garment;
sport science; muscle activity monitoring; gym fitness tracking.
I. INTRODUCTION
A. Background and Motivation
Muscle activity study is an important topic in life and sport
science. Understanding how muscles move during various
activities and exercises can help us compare the effect of
different exercises, choose more efficient and secure methods
for exercising the muscles and inducing muscle hypertrophy,
which asserts its role in fields such as body building, profes-
sional sport, injury rehabilitation, elderly caring, etc.
Today muscle monitoring is largely restricted to constrained
lab experiments. Long term monitoring under real life condi-
tions is difficult due to lack of unobtrusive mobile systems. As
reviewed in details in Section I-C, most existing approaches
such as EMG, MRI, USI, etc require either bulky hardware
or special attachment of electrodes to the skin, making wide
spread deployment unfeasible. While simpler approaches such
as force sensitive resistors attached to elastic bands [1] or
inertial measurement units that measure skin surface motion
have been studied, they provide only very limited amount of
information about the muscle.
B. Paper Contribution
This work demonstrates an alternative approach based on
textile ressitive pressure mapping sensors, which has been
used for activity recognition with sportmats [2] and tablecloths
[3] during our previous work. We propose a novel wearable
system that can be used in daily environments out of the
laboratories for monitoring muscle activities, and evaluate its
performance in a gym leg workout scenario. The sensor system
is integrated into the form of an elastic sport band, which
is unobtrusive and does not limit the movement freedom of
the user while doing various exercises. At the same time
with an 8×16 elements grid of 1cm pitch, a sampling rate
of 50fps (frames per second) and high dynamic range, the
system provides detailed information about muscle actions.
Thus, the core contributions of this paper are as follows:
•Novel wearable sensing approach: The wearable sen-
sor we have designed is an air-permeable fabric matrix
structure that measures force distribution on high spacial
and temporal resolutions. It is integrated into an off-
the-shelf sport band, suitable for most exercises and
everyday use. Compared with current IMU or EMG based
approaches, our sensor acquires information about the
muscle activities without direct skin-electrode contact.
•Leg exercise action recognition and quality evaluation:
–single action detection: Based on the muscle movement
signature of different exercises, we are able to distin-
guish among the activities of four machine exercises,
adjusting machines, workout pause and walking, on
a fine time granularity. In short, the system knows
exactly each second the user is doing which exercise,
how long the user takes breaks between each sets.
–repetition counting: By inspecting the repeating pat-
terns within each workout set, the system automati-
cally counts the repetitions without training templates.
Therefore, applications such as an automatic personal
workout diary can be easily achieved.
–Workout Quality Evaluation: Based on the common
gym experience that during an ideal and efficient
workout, the person should perform each action in a
fluent, stable pattern with the targeted muscles. We
abstract information describing the force consistency
among each repetition from every set as quality meters;
then we compare this individual information with the
user’s self report and the observer’s evaluation about
the effort the user puts on each set.
–The high framerate muscle force pressure mapping
information can be further used for the study of life
and sport science.
C. Related Work
To the best of our knowledge, state of the art methods for
detecting muscle activities are:
1) Electromyography (EMG) detects the biopotential from
skeletal muscles, which is directly related to the muscle
activity. It is widely used for studying muscles and
considered as the golden standard[4][5]. While traditional
EMG instruments are bulky, the state of the art hardware
has minimal footprint with off-the-shelf components,
which is suitable for wearable applications [6][7]. Main-
stream EMG electrodes need to have direct contact with
the muscle or the skin, and requires stable and minimal
impedance [8] to have reliable signal, which is not prac-
tical concerning conditions such as extreme movement
causing electrode shifting and excessive perspiration that
are common during normal exercises. Recently, contact-
less electrodes that are suitable for EMG signals are being
developed such as the works in [9][10].
2) Magnetic resonance imaging (MRI) forms images of the
body and internal organs. It is typically used before and
after exercise to observe the muscle structural differences.
[11] The MRI equipment is stationary and radioactive,
limiting its use only in lab conditions.
3) Ultrasound imaging (USI) also forms internal organ imag-
ing. While traditional USI equipment is not portable
and can only be used before and after exercise, such
as the study carried out in [12]; recently wearable USI
approaches are being proposed and modeled [13]. In [14],
Sikdar et al. using a single scan element on the wrist,
were able to robustly identify the movement of individual
fingers. Ultrasound therapy is also a versatile treatment
for musculoskeletal conditions, and recently wearable
applications are emerging as in [15].
4) Inertial Measurement Units (IMU) are nowadays a
standard feature of smartphones, smartwatches, fitness
trackers[16]. IMUs track the linear and angular accelera-
tion, thus can be interpreted into the movement of body
parts where the sensor is fixed, which helps repetition
counting and posture correction. [17] This, however, does
not translate directly into specific muscle activities: e.g.
using a smartwatch, biceps curl with a heavy barbell
results in similar signals as moving arms without weight
in the same speed and pattern; while the muscles undergo
different activities.
5) Force Sensitive Resistors (FSR) are carbon polymer thin
films with printed conductive electrodes, which respond
to pressure force. Previous works include muscle force
evaluation during exercise [18] and gesture recognition
by measuring forearm muscles [19], with one or several
Fig. 1. Wearable leg muscle surface pressure mapping system hardware
(1)prototype components; (2) fixing position and example pressure mapping
data, brighter color indicates higher pressure
FSRs fixed on the corresponding muscles. However, using
discrete sensors limits the level of detail that can be
sensed and creates placement issues (as the sensors need
to always be placed at the same location). Several sensor
systems for robotic sensing skins also offers similar
measurement, such as opto-electronic sensors with sili-
con casting [20]; however, their size and flexibility are
currently not suitable for sport wearable devices.
6) Textile capacitive sensors based on a capacitance change
between two conductive layers have also been used for
muscle monitoring [21]. Proper analog design can achieve
sensing of intense muscle movement as well as minor
tissues changes such as pulse [22].
7) Other medical measurements are used in lab studies such
as vascular and metabolic responses [23], which include
measuring blood flow and taking blood samples before
and after targeted exercises from the specific muscle; or
spectroscopy which attaches optical probes onto skin to
measure tissue oxygen saturation [24].
II. HA RDWARE SYSTE M
Our force sensitive resistance matrix sensor is essentially a
carbon-polymer fabric sheet, covered with a group of parallel
stripe electrodes of 1cm pitch on each side; the groups on both
sides are perpendicular to each other, forming a matrix struc-
ture with the intersections from the top-view. The electrodes
are also thin metal fibers, woven into a non-conductive fabric
sheet. The entire sensor is therefore made out of fabrics, and
is flexible and air permeable. We stitch an 8×16 sensor patch
onto an off-the-shelf elastic sportband, which is commonly
used for supporting joints and muscles. As no electrode-skin
contact is required, we cover the sensors with cohesive textile
band to shield it from sweat during intensive workout. The
sensor is wrapped around the leg to cover the lower part of
vastus lateralis and vastus medialis from the quadriceps muscle
group as in Fig. 1(2).
The electrodes are connected to our custom data acquisition
hardware based on our previous work [25]. The hardware
scans the entire matrix at a speed of 50fps by powering
one row electrode with 3.3Veach time, pulling the rest row
electrodes to ground; then measures the local resistance of
the carbon-polymer sheet through a voltage divider at every
column electrode with a 24-bit, 16-channel high performance
ADC. An FPGA controls the logic and offers the scan voltage.
To ensure reliable data-rate, we implement a FIFO-USB
protocol to send the data to a smartphone by a USB-OTG
cable, which also powers the hardware. The smartphone runs
an Android app to save and visualize the data. The hardware
and smartphone can be put inside the user’s pocket during the
workout. Fig. 1(1) shows the sensor system.
III. EXP ER IM EN T SET UP
A. Experiment Design
To evaluate the system’s capability of recording leg muscle
movement during gym workout, and distinguishing various
workout actions that are focused on different muscles, we
design our experiment as a complete gym leg workout session.
To introduce controlled movement, and for safety concerns,
we avoid exercises with free weights such as barbell squats;
instead, we choose gym machines that are designed to limit the
freedom of movement and hence, in most cases, the movement
of the weight is initiated by the targeted muscles. Each session
includes the following procedures:
1) Warm up with the Cross Trainer ’7-minute warm up’
program (Fig. 2 A);
2) For each machine, perform 3 sets of each 12 repetitions,
followed by one set, in which the participant performs
until failure (if the participant feels confident that failure
will not happen, till only 15 reps). The machines include:
a) Leg Press Machine (Fig. 2 B)
b) Seated Leg Curl Machine (Fig. 2 C)
c) Leg Extension Machine (Fig. 2 D)
3) Cool down with the Cross Trainer ’7-minute warm up’
program, same as step 1.
The participants are free to take pauses, drink, walk around
or any normal activities inside the gym. The system will
record all of those non-workout movements as ”NULL” class.
Every participant recorded four sessions, numbered Day1−4.
Between every workout day, the participants take several
days off to rest the leg muscles. For Day1and Day3, the
weight of each machine is constant; while for Day2and
Day4, the weight is increased after each set. The actual
constant weight, starting weight and increment are participant
dependent because the subjects have different muscle strength
and their performance actually improves from Day1to Day4.
In Step 2, the order of the three machines are shuffled within
the four days, so that the participants begin and end with a
different machine each day to eliminate the possible bias that
he/she might be already tired upon arriving at the last machine.
The exact wrapping tension and position of the sensor is
not precisely defined; instead, the only standards are (1) the
shapes of the muscles are visible on the app while the users
stand and tighten their legs; (2) the users can fully curl their
legs easily (stand and squat) with the band. The user can adjust
the band if it slides during training.
TABLE I
CLA SS DEFI NI TIO N
Class Definition
Workout Activities
Class 1 Cross Trainer
Class 2 Leg Press
Class 3 Seated Leg Curl
Class 4 Leg Extension
Non-workout Activities
Class 5 Mounting/Dismounting/Adjusting machines
Class 6 Pause (on or off machines)
Class 7 walking
B. Participants Demography
Overall 6participants recorded 24 complete sessions, con-
taining 288 sets of the three leg machines and 48 sets of
the cross trainer. To ensure safety while performing those
heavy weight exercises, the participants are regular members
at the gym, and they are aware of how to perform all of the
exercises safely. One participant is a professional gym trainer,
also a university student in sport science. Three participants
are university students who do regular sports or gym exercises,
but are not professional athletes. Two participants are full-time
researchers in the institute and do occasional sports. One of the
three students is female, and the other participants are male.
Their ages range: 21 −27, heights range: 171 −187cm, and
weights range: 70kg −85kg.
C. Annotation Method
The ground truth is annotated manually, detailed to the
second of the start and end of each set or other movements.
The class definitions are listed in Table I. The single repetition
is not annotated, because (1) counting can be easily imple-
mented by the movement of the weights (e.g. some smart
training machines are commercially available); (2) exercises
with weights are slower, less consistent compared to free-
hand exercises (e.g. push-ups), therefore, usually the exact
separation between adjacent reps is not clear simply by looking
at the person’s movement. Instead, the repetition number and
weight of each set are recorded in the form of a normal
workout diary by the experiment supervisor.
After each set, the participant is asked to rate his/her effort
of the past set from difficulty scores of easy (under weight),
normal (the person feels fair effort from the muscles moving
the weight), hard (the person needs to continue focus and self-
motivate moving the weight, movements may be inconsistent),
limit (the person’s movements are close to failure in moving
the weight). Then the experiment supervisor, with a third
person who looks at the recorded video, gives a rating of the
effort based on the participant’s rating, movement consistency
and facial expressions. It is worth mentioning that, the effort is
a subjective score; and the effort does not necessarily increase
as the set proceeds or weight increases. During the experiment,
sometimes if the participant takes a longer pause, he/she might
regain more strength, and the next set can be easier than the
previous one.
Fig. 2. Example of an experiment session’s data was explained in Section IV-A, and example pressure mapping signals of classes C1(A), C2(B), C3(C),
C4(D)
IV. ANA LYSI S AN D RES ULTS
A. Signal Processing and Feature Extraction
The signal from the sensor is essentially a time sequence
(stream) of 2D pressure distributions (frames). We use signal
conditioning schemes similar to the ones introduced in [2] and
[3]. The 2D frame is first spatially up-sampled from 8×16
to 16 ×32. For the training and initial recognition we pre-
segment the data according to the annotation (spotting as in
Section IV-C is done on non-segmented data). We then use
a sliding window (window size 8s, step size 2s) approach to
step through the stream of data, each window is denoted as i.
We extract features from two aspects:
•Temporal: For each frame we compute three central
moments (sum wiand centroid xi, yi) and the pixel
value of the pixel (σi) which has the maximum standard
deviation during the current window.
•Spacial: We then calculate the inter frames, which are dif-
ference per pixel between every pair of adjacent frames.
They represent the change of the pressure mapping. We
then sum those inter frames within the window, resulting
in a single augmented frame ADi, which represents the
overall 2D pressure change in the window.
Several practical reasons could cause the sensor to have
different offset values, such as the difference in wrapping
tensions and positions, sliding and adjusting of the sensor
band. To make the system more robust against such variations,
we avoid involving absolute values in extracting features. For
each temporal sequence s∈ {wi, xi, yi, σi}, we first normalize
sso the average value is 0, and the standard deviation(SDs)
is 1. We filter the signal to remove DC and high frequency
noise with zero-phase digital filtering. Then we compute the
following features, denoted as Ft(s):
1) magnitude range (max(s)−min(s));
2) average absolute of the 1st-order derivatives (avg(|s0|));
3) standard deviation of the 1st-order derivatives (SD(|s0|));
4) range of the 1st-order derivatives (max(s0)−min(s0));
5) central frequency of the FFT spectrum (without 0Hz);
6) we divide the spectrum into 5 even portions of frequency
band, use the mean magnitude of each as a feature;
7) we sort the data into a histogram of five bins, and use
the count number of each bin as five features;
8) using half of SDsas minimum peak height, we find the
local maxima and minima. We use the amount of maxima,
minima, and maxima-to-minima ratio as three features.
Again, to make the feature extracted from the augmented
frame translation invariant, instead of directly using ADi,
we look for the maximum pixel within a field that is 4
pixels retracted from the four borders. This pixel represents
the most movement during the window. We select the 9×9
Fig. 3. Leave-1-day-out average result
region, centered at this pixel as the region of interest ROIi.
Then we compute the first three cental moments and Hu’s 7
moments[26] of ROIi, denote those 10 features as Fs(ROIi).
Overall 82 features are extracted from every window:
F(i) = {Ft(wi), Ft(xi), Ft(yi), Ft(σi); Fs(ROIi)}.
B. Window Based and Event Based Classification
First we consider classification results on the basis of
individual windows. We use the confidence-based AdaBoost
algorithm ConfAdaBoost.M1 [27] with decision trees as the
base classifier. We randomly pick an even number of feature
data from every class for balanced training. Since the amount
of data from every class is very unbalanced, 9 classifiers are
trained with such random pick process; the final result is the
majority of the 9 classifiers’ outputs. We perform 10-fold
cross validation of the complete dataset from each person.
This yields over 95% average accuracy; however, this is over
optimistic: the data from the same repetition set have larger
similarity compared with other sets. Therefore, for everyone’s
4 days, we perform leave-1-day-out cross validation.
Fig. 3 (1) shows the average result of all participants
as confusion matrix; the detailed precision, recall, F1and
ACC of every day from every person are listed in Fig. 5.
Among the results, the gym trainer - ID5, has the highest
scores, with minimum daily scores deviation. We speculate
this to be a result of the participant doing the same machine
exercise in a very consistent manner across different sessions,
as his workout experience and body build allow him to easily
Fig. 4. Event-based Leave-1-day-out average result
Fig. 5. Distribution of leave-1-day-out individual day results
control the weights. We then performed leave-one-participant-
out validation to examine the system’s robustness against
encountering unregistered users; the confusion matrix is shown
in Fig. 3 (2).
Next, we elevate classification to event based level. For
every segmented set of C1−C4, we use the majority of all
windows’ output as its final class. We combine the remaining
non-workout classes into C5−7. As shown in Fig. 4, there
exists accuracy improvement in every class, and the resulting
overall F1 scores increase to 0.943 and 0.852, as this process
essentially smooths the window based result.
In this application, it is important to first separate the ma-
chine workout activities (C1−C4) from non-workout activities
(C5−C7), and then further distinguish which machine the
user is using. From the confusion matrix, it is obvious the most
misclassifications happen within the workout classes, and non-
workout classes; while very little confusion happens between
the two categories. Among the four machines, the cross trainer
is the most distinct from others; while it is misclassified with
walking in some occasions, it is reasonable since the workout
is a stepping action. In the person independent case, the con-
fusions among C2, C3, C 4are increased but not significantly.
The accuracy degradation from person dependent to inde-
pendent case is also a result of the participants’ physiology
variety in terms of muscle size, density, etc. As a baseline,
we speculate the system to be more robust against switching
users with more training data that cover more body types, and
that in practical use case such a personal device can be easily
trained with individual users for optimum accuracy.
C. Activity Detection (Spotting)
As a system that accompanies the user out of the lab
environment, it should be able to automatically spot the user’s
activities in a continuous data stream without annotation. The
challenge is to be able to separate the relevant activities from
aNULL class containing anything that the user can possibly
do in between.
For activity spotting we directly use the same sliding win-
dow istep and size as in Section IV-A to scan through the data
without annotation, and extract the same set of features F(i).
We assume a person dependent use case for optimal accuracy.
For each day’s data, we first train the ConfAdaBoost.M1
classifier with the annotated data from other three days of
the user (leave-1-day-out); during the training we combine
C1−C4,C5−C7into two classes C1−4,C5−7to distinguish
between workout activities and non-workout activities. We test
the classifier with F(i)from every window of this day, and
name the output as the binary spotting result, in which 010
indicates the workout activities. Singular events are removed,
then adjacent 010s are grouped into ’instances’, each with the
start and end points, then instances with less than 10 seconds
gap are merged.
Then, we train two different sets of classifiers: one trained
with data from {C1, C2, C 3, C4}, and tested on the windows
marked as workout instances; another with {C5, C6, C 7}and
tested on the remaining windows. The results within each
workout instance are majority-voted. We compare this final
spotting result with the ground truth on a 2 seconds (window
step) temporal precision. The resulting F1-score is on average
0.779; however, only 31 out of 336 workout sets are not
correctly spotted and classified (error <10%). The major
misses are among C5, C6, C 7, which also do not have a clear
separation during the annotation, for example: when the user
is approaching a machine, he/she might as well starts adjusting
the machine during the last few steps; when a user is relaxing,
he/she might adjust the machine to release his body, or stand
and slightly step around; some adjusting machine movements
might last much shorter than the window size (pulling the
weight pin). After all, the muscles are not actively asserting
force on a workout level. Therefore, we combine C5−C7in
the spotting result, reducing the classes from 7 to 5, and the
resulting average F1-score is 0.856.
Fig. 6 shows the spotting result of the experiment session
in Fig. 2; Fig. 7 shows the distribution of the F1-scores of
every session from every participants. The spotting result of
each person accords with the classification result in Fig. 5.
D. Counting and Workout Quality
To automatically assess not only the type of training, but
also the amount is relevant. We thus next attempt to count the
number of repetitions within each spotted instance. For every
instance, we use the sum w(t)of each frame and run a sliding
window to locate the local maxima with the minimum peak
height criteria of median(w) + SD(w)/2, and local minima
smaller than median(w)−SD(w)/2(wis the samples of the
current window instead of the whole instance). Then define the
Fig. 6. Spotting example for the session in Fig. 2, before and after combining
C5, C6, C 7
Fig. 7. Distribution of spotting F1-scores of all participants
smallest minima between two adjacent maxima as the start
of a partition, and end of the previous partition. While this
already defines a naive counting algorithm, it is sensitive to
minor variations in the timing and execution of the individual
repetition. Therefore, we use dynamic time warping (DTW)
to inspect whether the same pattern have been repeated, or
it is an over counting of an locally abnormal peak. In our
previous work [2], we have used DTW for counting body
weight exercises (e.g. push-ups) on a pressure sensitive mat,
one template was chosen as the standard to best represent
Fig. 8. Examples of partition and counting result. Red vertical lines mark the partition separations, orange and dash purple horizontal lines mark the standard
deviation and median of each window.
Fig. 9. Histogram of counting errors
each exercise. Because the repetitions here are not annotated
as explained in Section III-C, rather than choosing a standard
template, we derive a template for each set, by a recursive
process:
1) the partitions are separated into pairs of two;
2) the time series of each pair are warped to have the same
length in time domain to have the optimum distance using
dynamic programming [28];
3) a new time series is calculated as the average of the two
warped series;
4) by generating one new time series from each pair, the
collection of new time series are grouped into pairs of
two and go to step 2 until only one time series is left.
The remaining single time series is then taken as the template
of the current set.
Then, we look for the partitions shorter than half of the
average partition length in the current set, which is an over
counting candidate. We derive three time series for each
candidate: its own partition, the previous partition plus itself,
itself plus the following partition. Then we match each of the
three series with the set’s template, and calculate the mean
value of the difference between the two warped series as the
DTW warping error. The series with the smallest warping error
is the one that best matches the template. Hence, if either the
previous plus itself, or itself plus the next has the smallest
error among the three, the corresponding two partitions are
merged. Compared to our previous work in [2], DTW is
used for correcting the counting results instead of the major
counting criteria. To further save computational time, for cases
that over counting candidates do not exist in a set, or for the
cross trainer with hundreds of repetitions, the DTW correction
process is not performed.
While counting of movement repetitions can be easily done
through wearable IMUs or machines with position sensors on
the weights/cable, the actual meaningful information is how
much effort the user is asserting with the targeted muscles,
and the consistency of the force patterns during the workout.
For each partition’s time series w, we calculate the following
properties to represent the consistency variation:
1) p1length of the partition in seconds
2) p2range of absolute value (max(w)−min(w))
3) p3average of absolute value (avg(w))
4) p4DTW distance normalized by warping path as ex-
plained in [2]
5) p5DTW warping error as previously explained.
The template of the DTW is calculated as the same recursive
process, but after removing over counting. Fig. 6 shows
an example of the partitioning process from each machine
workout class, and the according p1−5. In the end, p1−5
outliers (within a dataset, points whose distance to the median
is greater than the standard deviation) at the beginning and
end of the set are removed and the number of the remaining
partitions is the final counting result. We compare the counting
results with the ground truth, and show the histogram of errors
in Fig. 9, from which, zero error takes the majority. In Fig.
6, outliers are kept to show how different the pvalues are for
signals out of the set. p1−5describe the speed, force intensity
and pattern variations, thus can be used as a new measure for
evaluating the workout consistency in future studies.
E. Observation of Warm Up Process
Muscle warm up is an important process for improving
sport performance and safety [29] [30] [31]. In this subsection
we demonstrate the capability of the system for evaluating
the warm up phase. While the skin temperature of the used
muscles indicates the warm up status; muscle warm up is,
by nature, the process of increased blood flow, oxygen,
metabolism and adrenaline in the muscle, which also results
in a muscle volume and strength improvement. With a fixed
elastic band around the muscle, this improvement causes an
increment in the average pressure which can be measured by
our sensor.
In Fig. 2, it is visible that the average w(t)during the
warm up is increasing, while during the cool down phase,
the increment is not as obvious; this difference exists in every
experiment session. While the magnitude range for each step is
smaller during warm up than during cool down in the particu-
lar session of Fig. 2, this does not apply to all workout sessions
(in some sessions the difference is not distinguishable or even
reversed). To quantify this and review all experiment sessions
statistically, we fit each cross-trainer session’s x=w(t)with
the following two functions: (1)y=k×x+bto approximate
the linear increment; (2)y=m×(1 −e−x/a)to approximate
the start to saturation trend. Fig. 10 shows the resulting k,
b,aand mfor every session(values are normalized for each
coefficient to compare only the difference), from which, it can
be concluded that during the warm up, the average pressure
increases more, with a lower starting value; and the average
pressure saturates faster after the main workout routines, with
a smaller start-to-saturation range.
F. Individual Set Difficulty
As introduced in Section III-C, each workout set has been
rated an effort score. We investigate the relevance between
the subjectively perceived effort and the consistency of the
Fig. 10. Distribution of muscle warm up parameters. x-axis: 1=warm up
sessions, 2=cool down sessions. Higher kmeans more increment; higher b
means more starting tension; higher ameans slower to saturation; higher m
means bigger start-to-saturation range
Fig. 11. Distribution of d9and d10. (before and after grouping easy-normal
and hard-limit; left: 1-easy, 2-normal, 3-hard, 4-limit; right: 1-easy & normal,
2-hard & limit)
force measured by the sensor. To avoid building upon errors
from counting, we do not use p1−5in Section IV-D; instead,
for every set, we use the same peak detection scheme on the
average pressure data with a sliding window, and also calculate
the standard deviation of the signal within the window SDi.
Then we derive the following pairs of features that describes
the signal consistency:
1) avg (d1) and SD (d2) of local maxima heights;
2) avg (d3) and SD (d4) of local minima heights;
3) avg (d5) and SD (d6) of local maxima distances;
4) avg (d7) and SD (d8) of local minima distances;
5) avg (d9) and SD (d10 ) of SDi;
The last pair of features are shown in Fig. 11, from which the
trend of more significant deviation related to higher difficulty
score is observable. To determine how well those features are
separated among the four rating classes (easy: 1; normal: 2,
hard: 3, limit: 4), we perform 10-fold cross validation with the
ConfAdaBoost.M1 algorithm. The resulting average accuracy
is 42.50%, which is not high, yet still above random (with
four balanced classes, the accuracy of random selection is
25%). We group easy and normal as class 1; hard and limit as
class 2, resulting in a cross validation accuracy of 72.99%. As
the ground truth is subjective, our sensor offers a new aspect
for evaluating the quality of the work out by measuring the
consistency within each repetition set.
V. CONCLUSION AND FUTURE WORK
In conclusion, this work has demonstrated a novel wearable
system based on fabric force mapping sensor matrix, which
can measure the muscle movement during various sport activ-
ities. The system is evaluated in a leg workout scenario, which
can be used for not only sport activity recognition, but also
quality evaluation. The major advantages of the system are:
(1) the high amount of sensing points make the device more
robust to fixing variations; (2) its air permeability, flexibility
and possibility to be isolated with sweat absorbing textiles
make it very suitable for sport activities without introducing
uncomfortableness and restriction; (3) the muscle information
can be extracted goes beyond mere movement as shown by
the warm-up and difficulty level analysis.
After the exploratory study, we will put on more engineering
effort to minimize the electronic footprint and implement
battery power and wireless transmission, along with designing
different sizes for different muscle regions, to enable the
system to be evaluated in a wider range of exercises, as well
as other sport activities.
ACKNOWLEDGMENT
This work was partially supported by the collaborative
project SimpleSkin under contract with the European Commis-
sion (#323849) in the FP7 FET Open framework. The support
is gratefully acknowledged.
The authors would also like to thank the gym Unifit
Kaiserslautern and trainer Robert Gressnich, as well as other
experiment participants Orkhan Amirhslanov, Finja Coerdt,
Sebastian Immel and Patrick Mathy.
REFERENCES
[1] P. Lukowicz, F. Hanser, C. Szubski, and W. Schobersberger, “Detecting
and interpreting muscle activity with wearable force sensors,” Lecture
Notes in Computer Science, vol. 3968, pp. 101–116, 2006.
[2] M. Sundholm, J. Cheng, B. Zhou, and et al, “Smart-mat: Recognizing
and counting gym exercises with low-cost resistive pressure sensing
matrix,” in UbiComp ’14. ACM, 2014.
[3] B. Zhou, J. Cheng, M. Sundholm, and et al, “Smart table surface: A
novel approach to pervasive dining monitoring,” in PerCom 2015, March
2015, pp. 155–162.
[4] A. P. Marsh and P. E. Martin, “The relationship between cadence and
lower extremity emg in cyclists and noncyclists,” Medicine and science
in sports and exercise, vol. 27, no. 2, pp. 217–225, 1995.
[5] S. T. McCaw and D. R. Melrose, “Stance width and bar load effects on
leg muscle activity during the parallel squat,” Medicine and science in
sports and exercise, 1999.
[6] W. Huang, M. Kreil, J. Cheng, and et al, “Development of compact
versatile acquisition platform for wireless sensing applications,” in
Applied Mechanics and Materials, vol. 645, 2014, pp. 274–280.
[7] Athos, “www.liveathos.com.”
[8] S. Day, “Important factors in surface emg measurement,” Bortec
Biomedical Ltd publishers, pp. 1–17, 2002.
[9] A. Ianov, H. Kawamoto, and Y. Sankai, “Development of a capacitive
coupling electrode for bioelectrical signal measurements and assistive
device use,” in Complex Medical Engineering (CME). IEEE, July 2012,
pp. 593–598.
[10] J. Taelman, T. Adriaensen, C. van der Horst, T. Linz, and A. Spaepen,
“Textile integrated contactless emg sensing for stress analysis,” in
Engineering in Medicine and Biology Society (EMBS). IEEE, Aug
2007, pp. 3966–3969.
[11] T. Abe, C. F. Kearns, and Y. Sato, “Muscle size and strength are
increased following walk training with restricted venous blood flow from
the leg muscle, kaatsu-walk training,” Journal of Applied Physiology,
vol. 100, no. 5, pp. 1460–1466, 2006.
[12] Y. Takai, M. Ohta, R. Akagi, and et al, “Validity of ultrasound muscle
thickness measurements for predicting leg skeletal muscle mass in
healthy japanese middle-aged and older individuals,” in Journal of
Physiological Anthropology, vol. 32, 2013.
[13] M. Tsakallakis and N. Bourbakis, “A wearable ultrasound multi-
transducer array system for abdominal organs monitoring,” in Bioin-
formatics and Bioengineering (BIBE), 2013 IEEE 13th International
Conference on, Nov 2013, pp. 1–5.
[14] S. Sikdar, H. Rangwala, E. Eastlake, I. Hunt, A. Nelson, J. Devanathan,
A. Shin, and J. Pancrazio, “Novel method for predicting dexterous
individual finger movements by imaging muscle activity using a wear-
able ultrasonic system,” Neural Systems and Rehabilitation Engineering,
IEEE Transactions on, vol. 22, no. 1, pp. 69–76, Jan 2014.
[15] G. Lewis, M. Langer, C. Henderson, and R. Ortiz, “Design and evalua-
tion of a wearable self-applied therapeutic ultrasound device for chronic
myofascial pain,” vol. 39, no. 8, 2013, pp. 1429 – 1439.
[16] Gymwatch, “www.gymwatch.com/sensor.”
[17] E. Velloso, A. Bulling, H. Gellersen, W. Ugulino, and H. Fuks, “Qual-
itative activity recognition of weight lifting exercises,” in Proceedings
of the 4th Augmented Human International Conference, ser. AH ’13.
New York, NY, USA: ACM, 2013, pp. 116–123.
[18] M. Kreil, G. Ogris, and P. Lukowicz, “Muscle activity evaluation using
force sensitive resistors,” in Medical Devices and Biosensors, 2008.
ISSS-MDBS 2008. 5th International Summer School and Symposium on,
June 2008, pp. 107–110.
[19] A. Dementyev and J. A. Paradiso, “Wristflex: Low-power gesture input
with wrist-worn pressure sensors,” in Proceedings of the 27th Annual
ACM Symposium on User Interface Software and Technology, ser. UIST
’14. New York, NY, USA: ACM, 2014, pp. 161–166.
[20] M. Donati, N. Vitiello, S. Maria De Rossi, T. Lenzi, S. Crea, A. Per-
sichetti, F. Giovacchini, and B. Koopman, “A flexible sensor technology
for the distributed measurement of interaction pressure,” in Sensors,
vol. 12, 2013.
[21] J. Meyer, P. Lukowicz, and G. Tr¨
oster, “Textile pressure sensor for
muscle activity and motion detection,” in Wearable Computers, 2006
10th IEEE International Symposium on. IEEE, 2006, pp. 69–72.
[22] J. Cheng, O. Amft, G. Bahle, and P. Lukowicz, “Designing sensitive
wearable capacitive sensors for activity recognition,” Sensors Journal,
IEEE, vol. 13, no. 10, pp. 3935–3947, Oct 2013.
[23] L. Lawrenson, J. G. Poole, J. Kim, and et al, “Vascular and metabolic
response to isolated small muscle mass exercise: effect of age,” American
Journal of Physiology - Heart and Circulatory Physiology, vol. 285,
no. 3, pp. H1023–H1031, 2003.
[24] A. Nygren, K. Rennerfelt, and Q. Zhang, “Detection of changes in
muscle oxygen saturation in the human leg: a comparison of two
near-infrared spectroscopy devices,” Journal of Clinical Monitoring and
Computing, vol. 28, no. 1, pp. 57–62, 2014.
[25] B. Zhou, J. Cheng, M. Sundholm, and P. Lukowicz, “From smart
clothing to smart table cloth: Design and implementation of a large
scale, textile pressure matrix sensor,” in ARCS2014. Springer LNCS,
2014.
[26] M.-K. Hu, “Visual pattern recognition by moment invariants,” Informa-
tion Theory, IRE Transactions on, vol. 8, no. 2, pp. 179–187, February
1962.
[27] A. Reiss and D. Stricker, “Aerobic activity monitoring: towards a
long-term approach,” International Journal of Universal Access in the
Information Society (UAIS), 2013.
[28] H. Sakoe and S. Chiba, “Readings in speech recognition,” A. Waibel and
K.-F. Lee, Eds. San Francisco, CA, USA: Morgan Kaufmann Publishers
Inc., 1990, ch. Dynamic Programming Algorithm Optimization for
Spoken Word Recognition, pp. 159–165.
[29] A. Fradkin, T. Zazryn, and J. Smoliga, “Effects of warming-up on
physical performance: a systematic review with meta-analysis,” in J
Strength Cond Res., 2010.
[30] D. High, E. Howley, and F. Don, “The effects of static stretching and
warm-up on prevention of delayed-onset muscle soreness,” in Research
Quarterly for Exercise and Sport, 1989, pp. 357–361.
[31] M. Mohr, P. Krustrup, L. Nybo, J. Nielsen, and B. J., “Muscle tempera-
ture and sprint performance during soccermatches – beneficial effect of
re-warm-up at half-time,” in Scand J Med Sci Sports, 2004, pp. 156–162.
