Validation of a Low-cost Inertial Exercise Tracker

Sarvenaz Salehi

and Didier Stricker

Daimler Protics, Germany

German Research Center for Artiﬁcial Intelligence (DFKI), Germany

Keywords:

Inertial Sensors, Body-IMU Calibration, Body Motion Tracking, Exercise Monitoring.

Abstract:

This work validates the application of a low-cost inertial tracking suit, for strength exercise monitoring. The

procedure includes an ofﬂine processing for body-IMU calibration and an online tracking and identiﬁcation of

lower body motion. We proposed an optimal movement pattern for the body-IMU calibration method from our

previous work. Here, in order to reproduce real extreme situations, the focus is on the movements with high

acceleration. For such movements, an optimal orientation tracking approach is introduced, which requires

no accelerometer measurements and it thus minimizes error due to outliers. The online tracking algorithm is

based on an extended Kalman ﬁlter(EKF), which estimates the position of upper and lower legs, along with

hip and knee joint angles. This method applies the estimated values in the calibration process i.e. joint axes

and positions, as well as biomechanical constraints of lower body. Therefore it requires no aiding sensors such

as magnetometer. The algorithm is evaluated using optical tracker for two types of exercises: squat and hip

abd/adduction which resulted average root mean square error(RMSE) of 9cm. Additionally, this work presents

a personalized exercise identiﬁcation approach, where an online template matching algorithm is applied and

optimised using zero velocity crossing(ZVC) for feature extraction. This results reducing the execution time

to 93% and improving the accuracy up to 33%.

1 INTRODUCTION

Strength training is one of the critical components of

the most ﬁtness and rehabilitation processes. Moni-

toring such exercises is beneﬁcial, in terms of perfor-

mance improvement, injury prevention and rehabili-

tation(Bleser et al., 2015).

Wearable systems, including multiple sensors

such as inertial measurement units(IMUs), provide an

efﬁcient solution for such applications. The kinemat-

ics analysis of body movements can be performed by

fusing and ﬁltering the well calibrated IMU measure-

ments and extracting higher level information, such

as joint angles and the segment positions (Yan et al.,

2017; Chardonnens et al., 2013). In the previous work

(Salehi et al., 2014), we presented the design and

development of a low-cost tracking suit which com-

posed of a wired network of IMUs and can be used for

a long time exercise monitoring. In the current work

the system is validated by a sequence of processes in-

cluding body-IMU calibration, body pose estimation

and personalized exercise identiﬁcation, in order to

provide a complete monitoring of user’s performance.

2 RELATED WORKS

2.1 Body-IMU Calibration

Body-IMU calibration is a key requirement for cap-

turing accurate body movements in applications based

on wearable systems (Zinnen et al., 2009). The

mounting positions of IMU with respect to joint is

critical information in joint angle estimation using ac-

celerometer measurements, especially during fast ro-

tations (Cheng and Oelmann, 2010) and when mod-

elling kinematic chains (Reiss et al., 2010). To obtain

such values from manual measurements or anthro-

pometric tables for different types of users is highly

error-prone and cumbersome. We previously pro-

posed a practical auto-calibration method for IMU

to body position estimation in (Salehi et al., 2015),

based upon a previously existing method by (Seel

et al., 2014). In contrast to (Seel et al., 2014), our

method considers three linked segments with IMUs

(pelvis, upper leg, lower leg) and, respectively, two

joints (hip, knee) in one estimation problem. This

makes it possible to beneﬁt from an additional con-

straint, which has shown theoretically and experimen-

Salehi, S. and Stricker, D.

Validation of a Low-cost Inertial Exercise Tracker.

DOI: 10.5220/0008965800970104

In Proceedings of the 9th International Conference on Sensor Networks (SENSORNETS 2020), pages 97-104

ISBN: 978-989-758-403-9; ISSN: 2184-4380

tally to provide more robust and accurate results under

suboptimal movement conditions. However, in exper-

iments with real data the dependency to global IMU

orientation could degrade the results due to magnetic

disturbances and outliers in accelerometer measure-

ments. In the current work the algorithm is evaluated

using data from two different types of movements.

In order to minimize the error due to existing out-

liers, especially in high acceleration, an optimal ori-

entation tracking algorithm is proposed, which esti-

mates relative rotations of the segments using gyro-

scope measurements. Therefore, the system is vali-

dated by this process, considering its applicability for

different types of users, who perform the movements

with different intensities and mostly with high accel-

eration.

2.2 Body Pose Estimation

In order to provide higher-level information such as

body segments positions and joint angles, as well as

3D visualization of the movements in the targeted ex-

ercise monitoring application, a variety of techniques

are proposed in the literature. Commonly, they as-

sume the body segments to be rigid bodies, which are

connected at joints.

A real-time motion tracking system was presented

in (Zhu and Zhou, 2004), using a linear Kalman ﬁl-

ter; the orientation of each body segment is calcu-

lated using accelerometer and magnetometer mea-

surements and fused with gyroscope measurements

in the Kalman ﬁlter, which results a smooth estima-

tion. This approach, however, doesn’t assume the

error caused by linear acceleration in accelerometer

measurement. Moreover, the magnetic disturbances

are not modelled. This could lead to faulty estimation

in the vicinity of ferromagnetic materials.

In order to consider the error of linear accelera-

tion, the idea of using physical and virtual sensor,

based on Newton-Euler equations, for estimating joint

angle is presented in (Dejnabadi et al., 2006). Ac-

cording to this method, acceleration of the joint can

be measured by placing a pair of virtual sensors on

the adjacent links at the joint center. In these ap-

proaches, joint angle is deﬁned by the difference of

direction in measured acceleration of sensors in the

joint frame. (Cheng and Oelmann, 2010) provides

a survey, which compares four different methods for

measuring the joint angle in two dimensions. They

proved that when the sensors are mounted far from

the joint center, especially in fast rotations a method

called CMRGD: common-mode-rejection with gyro

differentiation has less error and is easier to imple-

ment than the other methods. In this method, the

gyroscope measurements are numerically differenti-

ated in order to derive the angular acceleration, which

is required to calculate the acceleration vector of the

joint center.

Instead of using the magnetometer measurements,

as they can easily distort the orientation estima-

tion, different studies applied the biomechanical con-

straints in order to improve the estimation of body

motion in the horizontal plane. Such studies can be

divided into two categories:

In the ﬁrst category, the joint angles are estimated

in an state estimation approach, assuming the known

IMUs mounting orientations. In (Lin and Kuli

2012), with the assumption of constant joint accel-

eration and using Denavit-Hartenberg (DH) conven-

tion, the angle related to each possible degree of

freedom(DOF) at hip and knee joints is estimated in

an EKF. Here the accelerometer measurement model

follows a similar approach in (Cheng and Oelmann,

2010) and (Dejnabadi et al., 2006).

In the second category, the state vector contains the

relative orientation and the position of the segments.

In (Kok et al., 2014), an optimization approach incor-

porates the biomechanical constraints together with

the biases of inertial sensors and the error of limited

DOF on knee joint. Other than body pose, the body-

IMU calibration parameters are simultaneously esti-

mated. However, (Kok et al., 2014) didn’t present any

evaluation of accuracy of such parameters. Moreover,

this approach can not be used in a real-time applica-

tion, as it requires a batch of observations.

In the same category of methods, (Luinge et al., 2011)

proposed the Kinematic Coupling(KiC) algorithm, at

ﬁrst for a hinge joint; Assuming A and B are IMUs,

which are mounted on the adjacent segments of the

joint m, from the coupling concept the following con-

straint is deﬁned:

∆

P =

lmB −

lmA (1)

,where

∆

P is the relative position of B wrt. A and

lmA,

lmB are the distance vector between joint m

and the respective IMUs, all deﬁned in the global co-

ordinate system, G, which is typically aligned with

gravity and magnetic north. In this case the EKF state

vector contains relative positions and velocities of two

adjacent segments of the joint, the error in orientation

of the segments and gyroscope biases. In this work we

propose a similar approach to (Luinge et al., 2011),

in order to estimate the relative velocity and position

of a leg segments, in addition to hip and knee joint

angles. Using the biomechanical constraints and re-

sult of body-IMU calibration, the number of required

states is less in comparison to (Luinge et al., 2011),

plus there’s no need for extra measurements, such as

magnetic ﬁeld.

SENSORNETS 2020 - 9th International Conference on Sensor Networks

2.3 Exercise Identiﬁcation

It is common that trainers initially monitor trainees

and instruct them based on their health conditions,

while they are performing a strength exercise for the

ﬁrst time, so that they can independently perform the

exercise later. Possibility of injuries are normally high

in this phase, as the correct way of the exercise can’t

be carefully controlled, considering body character-

istics and abilities of individuals. For example, in the

rehabilitation, where range of motion is limited in dif-

ferent stages of recovery. It is thus very important to

provide a personalized monitoring application.

Exercise identiﬁcation is the process of identify-

ing the start and stop time of one repetition in an ex-

ercise, which could be composed of multiple smaller

components known as motion primitives. Usually a

repetition in an strength exercise comprises the se-

quence of increasing and decreasing velocity. There-

fore, zero velocity crossing(ZVC) approach is one of

the most optimal approaches to ﬁnd the motion prim-

itives, where the velocity of signal changes the sign.

(Fod et al., 2002) applied ZVCs, in order to detect the

motion primitive of two DOFs.

Hidden Markov Model(HMM) is a stochastic ap-

proach, which considers a signal as unobservable se-

quences of Markov states. At each time point the sys-

tem undergoes a state transition, which is deﬁned by a

probability in a transition matrix. In (Janus and Naka-

mura, 2005), a template free approach is proposed,

where the data is windowed and probability density

function of each window was used in a HMM to de-

tect different states of the movements. The segment is

identiﬁed, where the transition between states occurs.

HMM is more used in the template based ap-

proaches. In (Lin and Kuli

c, 2013) a two-stage ap-

proach is proposed to reduce the computational cost

by reducing the number of times that HMM should

run. This is achieved by ﬁrst scanning the observa-

tion signal for the candidate segments using ZVC or

velocity peaks in joint angle.

There are other learning based classiﬁers, which

are used for motion identiﬁcation such as Convolu-

tional Neural Network(CNN) (Um et al., 2017) or

Support Vector Machine(SVM) (Morris et al., 2014).

However, they require enough labelled training data

to increase the accuracy.

An alternative approach, which doesn’t require

any training data is template matching.

Dynamic Time Warping(DTW) is a popular tem-

plate matching algorithm, which creates a matrix of

the distances between each point of the observed sig-

nal and the template. Then this matrix is searched for

a warping path which leads to a minimum distance.

This can be used to identify an exercise using a tem-

plate of motion data, performed and captured with the

supervision of a trainer. However, it can be only ap-

plied in ofﬂine scenarios, due to its expensive compu-

tational cost, especially in higher dimensions. (Saku-

rai et al., 2007) proposed an online approach, which

addresses the problem of subsequent matching using

DTW. This algorithm is fast which means the process-

ing time of current observation point doesn’t depend

on the past data length. It only requires a single ma-

trix to ﬁnd the matching subsequent. In (Sakurai et al.,

2007), an experiment on joint position data from op-

tical motion capture system is presented. Here, this

algorithm was evaluated for real time motion identiﬁ-

cation based on IMU measurements.

3 PROCEDURE

3.1 Body-IMU Calibration

The calibration procedure is based on the method in

(Salehi et al., 2015), here using two different types of

movements: A) random movements in all directions,

B)separate movements on each DOF. Additionally, in

order to avoid problems such as outlier and distorted

measurements, caused when using accelerometer and

magnetometer, a new approach is proposed, speciﬁ-

cally for movements of type B, in order to estimate

the relative orientations, which are required for cali-

bration procedure (see Figure 1). This is explained in

the following section:

3.1.1 Relative Orientation Estimation

Considering a hinge joint, n, the measurements of gy-

roscopes (

ω), mounted on its two adjacent segments

(B,C), provide useful information in order to iden-

tify the joint axis ~r

. As the rotation is limited to

only one direction i.e. joint axis, any difference be-

tween the angular velocities of the two segments, in

the plane perpendicular to the rotation, is in contrast

to that limited degree of freedom. It is also obvious

that the difference between the angular velocities of

the segments in the direction of the joint axis is the

joint angular velocity. These facts are deﬁned in the

equations 2 and 3.

k − k

k = 0 (2)

= ω

− ω

(3)

,where · is scalar product. Thus the joint axis can

be estimated, using a set of measurements from two

gyroscopes on the joint segments, in an optimization

Validation of a Low-cost Inertial Exercise Tracker

Figure 1: The computation graph of body-IMU calibration procedure.

problem, with equation 2 as the cost function. The

relative segment orientation is obtained from integra-

tion of equation 3. Though this approach is explained

for knee joint, the same can be applied for hip, if the

movements are kept limited to only one DOF at a time

(see Figure 1). Therefore type B movements provides

optimal measurements for both joint axes and relative

orientation estimation, which are the prerequisites for

body-IMU calibration algorithm.

3.2 Body Pose Estimation

In this work an extended Kalman ﬁlter is designed to

estimate the knee and hip joint angles, as well as the

legs’ segments positions. The assumptions here are:

(1) a simple biomechanical model for leg with rigid

segments connected via frictionless joints: hip(m),

knee(n), (2) at least one IMU sitting on each rigid

segment that should be tracked: (A on pelvis, B on up-

per leg, C on lower leg), (3) forward kinematics equa-

tions, (4) constant acceleration for integration dura-

tion(sampling time).

The state vector, Equation 4, includes the

velocities(~v) and positions(

P) of the upper and

lower segments in addition to the hip rotation

quaternion(~q

), and the knee joint angle(

), all of

which are wrt pelvis coordinate:

~x = [

] (4)

Kinematic Process Model is deﬁned in Equation 5

(note: tilde superscripts indicate the measurements

and the vector signs is removed for simplicity).

k−1

˜a

k−1

∆T (5a)

k−1

˜a

k−1

∆T (5b)

k−1

∆T +

˜a

k−1

∆T

(5c)

k−1

+ ∆T

k−1

˜a

k−1

∆T

(5d)

= exp(R

k−1

−

k−1

(5e)

= θ

k−1

+ ∆T (

k−1

−

k−1

) (5f)

,where

˜a

k−1

= R

k−1

˜a

k−1

− ˜a

k−1

(6a)

˜a

k−1

= R

k−1

˜a

k−1

− ˜a

k−1

(6b)

Here, accelerometer( ˜a) and gyroscope(

ω) measure-

ments are the control inputs. R

is obtained from

using the quaternion to rotation matrix conver-

sion and R

is calculated from

and

using the

axis-angle to rotation matrix conversion. ∆T is the

sampling time.

Observation Model is deﬁned based on assumption

1; the segments pelvis and upper leg connected in

hip joint (Equation 7a) and upper and lower legs con-

nected in knee joint (Equation 7b):

= −R

lmB +

lmA (7a)

= −R

lmC +

lmB) +

lmA, (7b)

3.3 Exercise Identiﬁcation

In this work a template based exercise identiﬁcation

algorithm is proposed, which deploys the recorded

supervised exercise motion as the template for later

identiﬁcation of correct performance. Here, we ap-

plied a streaming subsequence matching method,

SPRING (Sakurai et al., 2007). For further optimiza-

tion of execution time, a preprocessing stage, includ-

ing motion primitive detection and a feature extrac-

tion step, are presented in this section.

3.4 Preprocessing

3.4.1 Motion Primitive Detection

As the strength exercises mostly contain the periodic

pattern, where the velocity of motion increases and

decreases sequentially, we chose ZVC method to de-

tect the motion primitives in the template as well as in

the streaming signal. Therefore the motion primitive

is detected, where the sign of derivative is changing

on the dominant DOF of motion signals e.g. positions

or joint angles. The derivative is calculated here using

a sliding window of 3 samples. The dominant DOF

SENSORNETS 2020 - 9th International Conference on Sensor Networks

100

is selected on each dataset by ﬁnding a dimension of

template signal which has the highest value.

3.4.2 Feature Extraction

The features consist of velocity, variance and mean of

each DOF in motion primitive. Therefore, for each

motion primitive tens of input samples are reduced to

3 features for each dimension. This, as it is shown

in the experimental results, has increased the speed of

identiﬁcation.

4 EXPERIMENTAL RESULTS

4.1 Body-IMU Calibration

In order to evaluate the proposed method, we cap-

tured a dataset from 7 subjects, 2 women and 5 men,

each performing 3 trials using the IMU harness of the

tracking suit, presented in (Salehi et al., 2014), and an

optical reference system, the NaturalPoint OptiTrack,

with 12 Prime 13 cameras, operated with the Motive

software (Optitrack, 2019). The test setup is shown

in Figure 2. In this experiment, each IMU was rigidly

connected with a rigid body marker. The IMUs were

interconnected via textile cables. In order to reduce

artefacts due to movement of the garment, the IMU-

marker-sets were strapped ﬁrmly on the pelvis and

one leg. We also used straps in order to attach marker

clusters on anatomical landmarks around the hip and

knee joints, from which we determined the joint cen-

ter of rotations.

In each trial, the subjects performed movements

of type A and B. In order to assess the repeatability

of the process each subject has done three trials.

Data in both optical system and IMUs captured

with 50Hz rate. The IMU measurements then were

downsampled to 5Hz. A hand-eye calibration was

used to transform the optical system coordinate to

the IMU coordinates. Note that errors due to marker

positioning are present, however, similarly for all the

tested methods. The data from type B movements

was ﬁrst used to estimate the joints’ axes, by applying

method described in 3.1.1. The calibration algorithms

were applied to both types of data. The following

describes the preprocessing and result analysis of

each one:

A. Calibration with Random Movements in All Di-

rections. For this type of movements, the global IMU

orientations were calculated based on a similar ap-

proach in (Harada et al., 2007), using IMU measure-

ments.

Figure 2: The red arrows show three IMUs which are

mounted on 1.pelvis (not visible), 2.upper leg and 3.lower

leg.

Figure 3 illustrates the overall RMSE in all the

four segments for all the trials, when applying the dif-

ferent calibration methods. This shows that the pro-

posed method provides more accurate results in 80%

of the trials, while a slightly worse performance can

be observed in the rest. The average error over all the

trials in the proposed method is 16.9±0.8cm, which is

lower than the ones of the Seel et al. method with

error of 18.13±0.7cm. The more detailed result of

proposed method is presented in Table 1. It can be

observed that the errors in segments lmA and lmB are

higher than the ones in lnB and lnC, as the movements

of these segments around the hip joint are more lim-

ited than the ones around the knee. This was also

proved theoretically, in the observability analysis in

(Salehi et al., 2015): conditions 2 and 3, which are

related to the angular velocity variations.

B. Calibration with Separate Movements on Each

DOF Both approaches are applied on the dataset,

which includes separate movements in each DOF.

Here, in order to estimate the relative orientations of

the segments, we applied the method in 3.1.1. Figure

4 illustrates the overall RMSE in all the four segments

for all the trials, which shows the proposed method

has more accurate results in all the trials with the av-

erage error of 15.1±0.6cm. The detailed result of pro-

posed method in Table 1.

It is shown that the proposed method with the

dataset of type B movements results better than the

type A, while the Seel et al. method results worse.

This proves the ﬁrst condition of observability in

(Salehi et al., 2015). Moreover, during the experi-

ment with the random movements in order to excite

all the segments simultaneously the subjects have per-

formed fast and hardly controlled movements, which

led to high deviation of accelerometer measurements

Validation of a Low-cost Inertial Exercise Tracker

101

Table 1: The positions of each joint m,n with respect to each IMU A,B,C estimated by method in (Salehi et al., 2015) from 7

subjects performing type A movements.

cm s1 s2 s3 s4 s5 s6 s7

lmA 12.5±3.37 18.57±4.3 15.6±1.19 15.99±2.14 15.02±3.31 16.89±0.7 13.68±5.1

lmB 24.39±0.9 22.28±4.93 19.72±2.58 23.34±2.81 25.92±3.97 24.8±3.16 31.18±3.79

lnB 9.98 ±6.37 16.12±0.52 10.26±3.33 9.35±2.01 12.49±4.62 10.12±1.85 7.23±3.87

lnC 14.86±0.47 12.85±1.57 9.5±1.34 9.48±1.15 11.54±2.86 15.24±5.94 11.86±5.23

Table 2: The positions of each joint m, n with respect to each IMU A,B,C estimated by method in (Salehi et al., 2015) and

Section 3.1.1 from 7 subjects performing type B movements.

cm s1 s2 s3 s4 5 s6 s7

lmA 23.13±3.26 18.18±3.53 14.12±1.69 14±3.07 13.51±4.07 17.33±0.82 15.2±0.73

lmB 14.26±0.87 17.03±4.98 19.69±3.04 18.42±2.88 16.49±7.61 20.21±2.23 24.79±4.37

lnB 13.5±4.93 14.66±3.5 7.86±1.55 6.4±2.65 13.89±5.54 8.2±1.88 3.61±1.92

lnC 13.72±0.22 11.71±3.23 9.93±0.93 10.17±2.51 8.88±5.49 13.34±3.14 11.94±3.86

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Test number

0.12

0.13

0.14

0.15

0.16

0.17

0.18

0.19

0.2

0.21

0.22

Position error[m]

Seel method

Proposed method

Figure 3: Results on measurements from random move-

ments: RMSE of the body-IMU calibration using Seel et

al.(green) and the proposed method(black), with respect to

optical tracker. Test 1,2,3 is with subject 1, test 4,5,6 with

subject 2, etc.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

Test number

0.12

0.13

0.14

0.15

0.16

0.17

0.18

0.19

0.2

0.21

0.22

Position error[m]

Seel method

Proposed method

Figure 4: Results on measurements from the movements on

each DOF separately: RMSE of the Body-IMU calibration

using Seel et al.(green) and the proposed method(black),

with respect to optical tracker.

from gravity. Commonly in most of the orientation

estimation ﬁlters, these measurements are considered

as outliers and the estimated orientation in the pres-

ence of such measurements is prone to error. This has

led to higher error in the proposed method than when

using type B measurements.

Table 3: Error of estimated upper and lower segments posi-

tions using the proposed method for squat exercise.

PB PC

cm [RMS, STD, Max] [RMS, STD, Max]

s1 [8.47, 10.02, 4.10] [8.98, 8.43, 36.16]

s2 [7.72, 5.13, 16.53] [7.69, 3.87, 14.20]

s3 [6.05, 4.44, 15.30] [9.12 5.49, 17.68]

s4 [6.57, 4.93, 19.92] [7.51, 6.14, 24.64]

s5 [8.91, 5.10, 18.02] [7.84, 4.14, 16.96]

s6 [7.20, 4.03, 13.73] [5.47, 3.55, 12.84]

s7 [6.40, 4.65, 15.27] [7.42, 5.15, 18.53]

Table 4: Error of estimated upper and lower segments posi-

tions using the proposed method for abduction and adduc-

tion exercise.

PB PC

cm [RMS, STD, Max] [RMS, STD, Max]

s1 [5.19, 3.38, 10.57] [10.53, 6.91, 21.40]

s2 [4.83, 2.97, 25.61] [9.96, 6.14, 24.84]

s3 [4.60, 1.99, 8.52] [10.63, 5.66, 21.18]

s4 [7.16, 3.87, 13.84] [12.31, 7.14, 24.65]

s5 [6.10, 3.64, 12.40] [12.43, 8.58, 27.58]

s6 [7.74, 4.58, 15.56] [16.28, 11.0, 34.62]

s7 [8.09, 5.95, 19.45] [16.33, 12.84, 42.45]

4.2 Lower Body Pose Estimation

The pose estimation approach in Section 3.2 was eval-

uated for lower body movements. The state vector is

initialized using the Equations 7a and 7b. The initial

global orientations and pelvis orientation were esti-

mated using a similar approach to in (Harada et al.,

2007). With the setup similar to Section 4.1, esti-

mation of a leg segments positions captured by both

SENSORNETS 2020 - 9th International Conference on Sensor Networks

102

Table 5: Joint angle identiﬁcation without ZVC.

s1 s2 s3 s4 s5 s6 s7

Accuracy[%] 99 99 99 99 99 99 99

Precision[%] 100 100 100 100 100 100 100

Time[s] 4.4 3.1 3.0 2.1 2.8 4.5 2.7

Table 6: Joint angle identiﬁcation with ZVC.

s1 s2 s3 s4 s5 s6 s7

Accuracy[%] 99 77 99 99 98 99 98

Precision[%] 100 6 100 71 100 80 40

Time[s] 0.4 0.2 0.2 0.1 0.1 0.1 0.2

tracking suit and optical system. This experiment was

carried out with 7 subjects each performing squat and

hip abduction/adduction exercises.

The results for squat and abduction/adduction ex-

ercises are presented in Tables 3 and 4 respectively.

The average error for all subjects in estimation of

lower leg position, is higher than upper leg, as it con-

tains more error related to knee joint angle plus the

calibration error of IMU position in particular lnC.

4.3 Exercise Monitoring

The algorithm was evaluated using the result of body

motion tracking in section 4.2. The ﬁrst squat which

was performed following the instruction of a supervi-

sor, was used as a template to identify the next repeti-

tions. After squats the subjects have performed other

movements, Figure 5.

As the joint angle is commonly used in the move-

ment identiﬁcation, here this signal was used as a

baseline to evaluate the exercise identiﬁcation for

when using ZVC for feature extraction. As it can be

seen from Tables 5 and 6, applying ZVC has a high

impact on the execution time.

For position signal, the average accuracy and pre-

cision are higher than these values for the joint angles.

This implies that the position signals in comparison to

the joint angles, contain more information related to

type of exercise. For more detailed comparison see

Tables 6,7.

The last experiment was done by using the pelvis

orientation as an additional input, as the correct per-

formance of the squat exercise is highly depends the

pelvis movements. The result is presented in Table 8.

It was noticed that in only one test, where the subject

didn’t follow the instructed exercise, the sensitivity is

lower in comparison to the previous experiment.

Table 7: Position identiﬁcation with ZVC.

s1 s2 s3 s4 s5 s6 s7

Accuracy[%] 99 99 99 98 98 99 99

Precision[%] 100 83 100 100 100 100 100

Sensitivity[%] 25 100 60 20 28 20 20

Time[s] 0.6 0.2 0.1 0.1 0.1 0.2 0.1

Table 8: Position plus quaternion identiﬁcation with ZVC.

s1 s2 s3 s4 s5 s6 s7

Accuracy[%] 99 99 99 98 98 99 99

Precision[%] 100 83 100 100 100 100 100

Sensitivity[%] 25 100 60 20 14 20 20

Time[s] 0.9 0.3 0.2 0.1 0.1 0.2 0.2

5 CONCLUSIONS

In this work the application of exercise monitoring is

presented in order to validate a low-cost inertial track-

ing suit. This has been achieved following three dif-

ferent processes: ofﬂine body-IMU calibration, on-

line body pose estimation and online exercise identi-

ﬁcation. It was shown that using the previous body-

IMU calibration method for random movements, de-

grade the results due to the presence of outliers espe-

cially from accelerometer measurements. Therefore

an optimal orientation tracking method was proposed,

which can be realized using only gyroscope measure-

ments, while the user performs separate movements

around each DOF. Using dataset from this type of

movements, the result showed overall improvement,

especially for the leg segments, which can not provide

enough useful movements for calibration. The body

pose estimation approach is based on EKF, where

only inertial measurements are contributing as control

inputs. In the proposed approach the lack of a reli-

able reference measurement for the horizontal plane,

0 10 20 30 40 50 60 70

Time[s]

-0.2

-0.15

-0.1

-0.05

0.05

0.1

0.15

0.2

Position[m]

Squat Identified

Figure 5: Online squat exercise identiﬁcation. The red ar-

rows show the timestamps when the squats are identiﬁed.

Validation of a Low-cost Inertial Exercise Tracker

103

e.g. magnetic ﬁeld, was compensated by modelling

the joints’ constraints in observation model and using

the body-IMU calibration results, i.e. joint axes and

positions. In order to monitor the strength exercises,

a personalized identiﬁcation approach was proposed,

which doesn’t require a large labelled training dataset.

The idea is to use a template signal captured, where

users are instructed to perform the movements cor-

rectly according to their ability and health conditions.

Therefore, an online template matching algorithm is

optimized and applied to estimated position, which

led to improved accuracy and execution time. The ex-

perimental results of this validation showed relatively

good results, considering high intensity of the move-

ments. For further improvement in future , in order to

compensate for the intensive dynamic movements, an

outlier rejection approach can be implemented. Ad-

ditionally performance of EKF can be improved by

adaptive tuning of the noise covariances.

REFERENCES

Bleser, G., Steffen, D., Reiss, A., Weber, M., Hendeby, G.,

and Fradet, L. (2015). Personalized physical activity

monitoring using wearable sensors. In Smart health,

pages 99–124. Springer.

Chardonnens, J., Favre, J., Cuendet, F., Gremion, G., and

Aminian, K. (2013). A system to measure the kine-

matics during the entire ski jump sequence using iner-

tial sensors. Journal of biomechanics, 46(1):56–62.

Cheng, P. and Oelmann, B. (2010). Joint-angle measure-

ment using accelerometers and gyroscopes?a survey.

IEEE Transactions on instrumentation and measure-

ment, 59(2):404–414.

Dejnabadi, H., Jolles, B. M., Casanova, E., Fua, P., and

Aminian, K. (2006). Estimation and visualization of

sagittal kinematics of lower limbs orientation using

body-ﬁxed sensors. IEEE Transactions on Biomedi-

cal Engineering, 53(7):1385–1393.

Fod, A., Matari

c, M. J., and Jenkins, O. C. (2002). Auto-

mated derivation of primitives for movement classiﬁ-

cation. Autonomous robots, 12(1):39–54.

Harada, T., Mori, T., and Sato, T. (2007). Development of

a tiny orientation estimation device to operate under

motion and magnetic disturbance. The International

Journal of Robotics Research, 26(6):547–559.

Janus, B. and Nakamura, Y. (2005). Unsupervised proba-

bilistic segmentation of motion data for mimesis mod-

eling. In ICAR’05. Proceedings., 12th International

Conference on Advanced Robotics, 2005., pages 411–

417. IEEE.

Kok, M., Hol, J. D., and Sch

on, T. B. (2014). An

optimization-based approach to human body motion

capture using inertial sensors. IFAC Proceedings Vol-

umes, 47(3):79–85.

Lin, J. F. and Kuli

c, D. (2012). Human pose recovery us-

ing wireless inertial measurement units. Physiological

measurement, 33(12):2099.

Lin, J. F.-S. and Kuli

c, D. (2013). Online segmentation of

human motion for automated rehabilitation exercise

analysis. IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 22(1):168–180.

Luinge, H. J., Roetenberg, D., and Slycke, P. J. (2011).

Inertial sensor kinematic coupling. US Patent App.

12/534,526.

Morris, D., Saponas, T. S., Guillory, A., and Kelner, I.

(2014). Recoﬁt: using a wearable sensor to ﬁnd, rec-

ognize, and count repetitive exercises. In Proceedings

of the SIGCHI Conference on Human Factors in Com-

puting Systems, pages 3225–3234. ACM.

Optitrack (2019). Optitrack ﬂex 13 (accessed 9/17/2014).

Reiss, A., Hendeby, G., Bleser, G., and Stricker, D.

(2010). Activity recognition using biomechanical

model based pose estimation. In 5th European Con-

ference on Smart Sensing and Context (EuroSSC),

pages 42–55.

Sakurai, Y., Faloutsos, C., and Yamamuro, M. (2007).

Stream monitoring under the time warping distance.

In 2007 IEEE 23rd International Conference on Data

Engineering, pages 1046–1055.

Salehi, S., Bleser, G., Reiss, A., and Stricker, D. (2015).

Body-imu autocalibration for inertial hip and knee

joint tracking. In Proceedings of the 10th EAI Interna-

tional Conference on Body Area Networks, BodyNets

’15, pages 51–57, ICST, Brussels, Belgium, Bel-

gium. ICST (Institute for Computer Sciences, Social-

Informatics and Telecommunications Engineering).

Salehi, S., Bleser, G., and Stricker, D. (2014). Design and

development of low-cost smart training pants (stants).

In 4th International Conference on Wireless Mobile

Communication and Healthcare, At Athen, Greece,

pages 39–44. IEEE.

Seel, T., Raisch, J., and Schauer, T. (2014). Imu-based

joint angle measurement for gait analysis. Sensors,

14(4):6891–6909.

Um, T. T., Babakeshizadeh, V., and Kuli

c, D. (2017). Ex-

ercise motion classiﬁcation from large-scale wearable

sensor data using convolutional neural networks. In

2017 IEEE/RSJ International Conference on Intelli-

gent Robots and Systems (IROS), pages 2385–2390.

IEEE.

Yan, X., Li, H., Li, A. R., and Zhang, H. (2017). Wear-

able imu-based real-time motion warning system for

construction workers’ musculoskeletal disorders pre-

vention. Automation in Construction, 74:2–11.

Zhu, R. and Zhou, Z. (2004). A real-time articulated human

motion tracking using tri-axis inertial/magnetic sen-

sors package. IEEE Transactions on Neural systems

and rehabilitation engineering, 12(2):295–302.

Zinnen, A., Blanke, U., and Schiele, B. (2009). An analysis

of sensor-oriented vs. model-based activity recogni-

tion. In 2009 International Symposium on Wearable

Computers, pages 93–100. IEEE.

SENSORNETS 2020 - 9th International Conference on Sensor Networks

104