Multiview Human Body Tracking of Hurdle Clearance: A Case Study
Tomasz Krzeszowski
1
, Krzysztof Przednowek
2
, Krzysztof Wiktorowicz
1
and Janusz Iskra
3
1
Faculty of Electrical and Computer Engineering, Rzeszow University of Technology, Rzeszow, Poland
2
Faculty of Physical Education, University of Rzeszow, Rzeszow, Poland
3
Faculty of Physical Education and Physiotherapy, Opole University of Technology, Opole, Poland
Keywords:
110 m Hurdle Race, Human Motion Tracking, Multicamera System, Particle Swarm Optimisation.
Abstract:
This initial research is a case study that uses a multiview human body tracking method as a tool to measure
hurdle clearance kinematic parameters. This study is conducted on a hurdler representing a high sport level,
who is a participant in the European and World Championships and the Olympic Games. The video recordings
were made under simulated starting conditions of a 110 m hurdle race. Kinematic parameters are estimated
based on the analysis of images from a multicamera system. The images were recorded with a resolution
of 1920x1080 and with a frequency of 100 Hz. The proposed method does not use any special clothes,
markers or other estimation support techniques. The parameters of the hurdle clearance were compared with
the parameters obtained from ground truth poses. Mean Absolute Error and Mean Relative Error were used as
the quality criteria.
1 INTRODUCTION
Hurdling is a group of athletic events in which tech-
nical preparation is very important. The hurdle race
technique involves running over 10 hurdles that are
from 0.84 to 1.07 m high, depending on the event.
In these races, the technique evaluation is mainly
focused on certain phases of human motion (Iskra,
2012). The existing studies are devoted to the kine-
matic analysis of the so called ”hurdle clearance”
(
ˇ
Coh, 2003;
ˇ
Coh et al., 2008). For example, the pa-
per by
ˇ
Coh (
ˇ
Coh, 2003) analyses selected kinematic
parameters (e.g., height of center of mass, angle of
placement of a leg) that describe the Colin Jackson’s
hurdle clearance technique. Meanwhile, the paper by
(Salo et al., 1997) contains a three-dimensional (3D)
biomechanical analysis of sprint hurdles. To esti-
mate the parameters (e.g., take-off distance, horizon-
tal velocity), two cameras (25 Hz) with ”Kine anal-
ysis” software were used. The main objective here
was to determine and compare selected biomechani-
cal parameters in two groups of men and two groups
of women at different competitive levels.
There are a number of computer vision methods
that play an increasingly important role in supporting
sports training. For example, (Reyes et al., 2016) de-
veloped an algorithm that processes underwater video
sequences for swimmer detection and tracking based
on light absorbance in conjunction with compres-
sive sensing concepts. The developed algorithm was
tested on two video sequences. Meanwhile, motion
detection and tracking methods have been used to
analyse athletics videos (Ramasso et al., 2009; Pana-
giotakis et al., 2006). Another solution that uses com-
puter vision techniques is a system for tracking play-
ers in indoor team games, such as in handball (Per
ˇ
s
and Kovacic, 2000). Indoor sports have also been
analysed by (Kim and Cho, 2016), who proposed a
robust multi-object tracking algorithm for acquiring
object oriented multi-angle videos. In this algorithm,
multiple camera images are integrated using homog-
raphy based transformation in order to cover large ar-
eas of interest. A motion tracking of a tennis racket
using a markerless system with a monocular camera
has been presented by (Elliott et al., 2014). Further-
more, (Sheets et al., 2011) used a markerless motion
capture system to evaluate kinematic differences at
the lower back, shoulder, elbow, wrist, and racquet
between the flat, kick, and slice serves. In this study,
seven male players were tested on an outdoor court
in daylight conditions. A method to identify sports
players in videos has been proposed by (Hamatani
et al., 2016), where the identification was achieved by
motion feature matching between (unknown) players
in videos (the features were obtained from estimated
postures in the videos) and wearable sensors whose
Krzeszowski T., Przednowek K., Wiktorowicz K. and Iskra J.
Multiview Human Body Tracking of Hurdle Clearance: A Case Study.
DOI: 10.5220/0006498400830088
In Proceedings of the 5th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2017), pages 83-88
ISBN: 978-989-758-269-1
Copyright
c
2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
IDs were already known. The experimental results
showed that the proposed method successfully iden-
tified 10 players with 72% accuracy. Cheng et al.
(Cheng et al., 2004) have shown that the proposed
motion descriptor can successfully classify the fol-
lowing four sports types: sprint, long-distance run-
ning, hurdling and canoeing. Their experimental re-
sults were obtained using video material from the
1992 Barcelona Olympic Games. Finally, Zhang et
al. (Zhang et al., 2017) proposed an algorithm to track
player actions from a sports video sequence. Their
method combines a particle filtering and mean shift
in order to effectively trace a fast-moving target.
This study proposes a multiview markerless
method of human body motion tracking to estimate
hurdle clearance parameters. In our analysis, nine
distance parameters and eight angle parameters are
taken into account. These parameters are estimated
based on the analysis of image sequences captured
with a multicamera system. The tracking system that
we have developed does not use any special clothes,
markers or other techniques supporting estimation.
This study is a continuation of research presented in
our previous paper (Krzeszowski et al., 2016). To
the best of our knowledge, the method of multiview
markerless tracking is used for the first time for mea-
surement of hurdle clearance kinematic parameters.
2 METHODS
2.1 Multiview Human Motion Tracking
The purpose of human body tracking is to estimate a
pose that reflects as closely as possible a real pose. It
should be noted that capturing the 3D position of a
human body is a very difficult task (John et al., 2010;
Kwolek et al., 2012). The main problems include:
high dimensional search space, image noise, the large
variability in the appearance of the tracked humans
and environment, the complexity of human motion
and the fact that particular parts of the body are of-
ten obscured. Although these issues can be solved in
many different ways, the most common methods use
simplified human body models (Deutscher and Reid,
2005; John et al., 2010; Kwolek et al., 2012; Krzes-
zowski et al., 2016), uniform background (Deutscher
and Reid, 2005), and also properly selected clothes
of the tracked human body in order to facilitate the
determination of distinctive features. In the process
of tracking, a particle filter algorithm (Sidenbladh
et al., 2000) or its modified versions are frequently
used (Deutscher and Reid, 2005). However, these al-
gorithms require a significant number of particles in
Figure 1: 3D human body model (left), hierarchical struc-
ture (right).
order to find the correct solution, which directly im-
pacts on the time needed for computations. Therefore,
in the human body motion tracking process, particle
swarm optimisation algorithms (Kennedy and Eber-
hart, 1995; John et al., 2010; Kwolek et al., 2012;
Krzeszowski et al., 2016) are mostly used because
they enable a more effective exploration of the search
space.
A 3D model is used to determine the human body
pose; that is, the position and orientation in space as
well as the angles between the joints. The model that
is used in this research is based on the kinematic tree
structure consisting of 11 segments, each of which is
represented by a truncated cone (Kwolek et al., 2012;
Deutscher and Reid, 2005), see Figure 1. The space in
which the model operates is determined by the num-
ber of degrees of freedom (DoF). Each segment in-
cludes up to three DoFs that define its orientation; the
only exception is the pelvis, which contains three ad-
ditional segments defining the model translation. The
model used in this paper includes 24 DoFs. In our
method both the model configuration and the pose of
a human body in the first frame of a sequence of im-
ages are selected manually.
The particle swarm optimisation algorithm (PSO)
(Kennedy and Eberhart, 1995; Krzeszowski et al.,
2016) is used in the motion tracking process. In our
method, the position of a particle represents the hypo-
thetical state (pose) of an athlete. The best solution is
selected based on a fitness function value. The fitness
function determines the degree of similarity between
the real and the estimated human pose. In this study,
the fitness function consists of two components. The
first is determined on the extracted human silhouette,
whereas the second is based on the edge distance map
(John et al., 2010; Krzeszowski et al., 2016). The
value of the function for the cth camera is calculated
using the following equation:
f
c
(x) = 1 (a f
c
1
(x) + b f
c
2
(x)) (1)
where x is the human body pose (the position of a par-
ticle) and a, b are experimentally chosen weighting
factors. The f
c
1
(x) function describes the degree of
overlap of the rendered 3D model with the extracted
silhouette, and f
c
2
(x) is determined by comparing the
3D model edges with the image, including the map
with pixel distances from the nearest edge. The fit-
ness function for all cameras is determined according
to the following equation:
f (x) =
1
C
C
c=1
f
c
(x) (2)
where C = 3 is the number of cameras.
2.2 Data Acquisition
The proposed method was applied for two image se-
quences. The recorded athlete was a participant in the
European and World Championships and the Olympic
Games. The data were registered in the athletics hall
with a tartan track. Throughout the research, the
sequence of passing the third hurdle in the regula-
tion conditions of 110 m race (hurdle height: 1.067
m, distance between the hurdles: 9.14 m) was cap-
tured. The sequences, in the form of color images
of size 1920x1080, were captured with three Basler
Ace acA1920-150uc cameras with the frequency of
100 Hz. Figure 2 illustrates how the cameras were
arranged in the athletics hall. The parameters of the
cameras have been estimated using the TSAI calibra-
tion method (Tsai, 1987).
2.3 Evaluation of the Parameters
Five key points (P
1
P
5
) were analysed in three phases
of hurdle clearance. The analysis included 17 parame-
ters, which are presented in Figure 3. The parameters
were selected based on the literature review (Iskra,
2012;
ˇ
Coh, 2003). The quality of the tracking was
evaluated based on the estimated pose and the ground
truth pose. The ground truth pose was obtained by
manually matching the 3D model to the athletes on
C3
C1
C2
RGB
Recording
equipment
RGB
RGB
hurdle
direction of the run
hurdle
hurdle
9.14 m
9.14 m
1.35 m
4.03 m
1.33 m
h = 1.32 m
h = 1.32 m
h = 1.32 m
5.20 m
4.30 m
Figure 2: Acquisition station, h is the height of the camera
from the ground.
those images that contained the five key poses char-
acteristic for hurdle clearance (Figure 3). The error
level was determined for each parameter. The param-
eters were estimated by the implemented algorithm
and they were then compared with the values of the
ground truth reference model (the model was man-
ually adjusted to the analysed images). The quality
criterion was defined for each parameter as:
e
n
= |
ˆ
X
n
X| (3)
MAE =
1
N
N
n=1
e
n
(4)
where e
n
is the absolute error, N is the number of al-
gorithm repetitions,
ˆ
X
n
is the estimated value (deter-
mined by the algorithm), X is the ground truth value,
and MAE is the mean absolute error. Moreover, the
mean relative error was calculated from the following
formula:
MRE =
1
N
N
n=1
e
n
X
· 100 (5)
3 EXPERIMENTAL RESULTS
An example of the tracking results for the selected se-
quence for three views is shown in Figure 4. The basic
statistics for the analysed parameters, the ground truth
value X, and the errors MAE and MRE are presented
in Table 1. These results were obtained for N = 10
repetitions of the tracking algorithm. The error anal-
ysis shows that among all of the distance parameters,
the estimation of CM height over hurdle in P
3
(h
3
) for
sequence 1 is determined with the greatest MRE er-
ror. This error is equal to 13.3% (MAE = 37.3 mm).
In the case of sequence 2, this error is smaller and
equal to 7.9% (MAE = 23.6 mm). Krzeszowski et al.
(Krzeszowski et al., 2016) used a monocular motion
tracking system, and this parameter was determined
with MAE = 30.5 mm and MRE = 8.3%. The small-
est MRE error for both sequences is obtained for the
CM distance from the hurdle in P
1
(w
1
). This error is
equal to 0.5% (MAE = 13.5 mm) in sequence 1 and
it is equal to 0.7% (MAE = 17.3 mm) in sequence 2.
In the work by (Krzeszowski et al., 2016), the errors
for w
1
were MAE = 25.8 mm and MRE = 1.0%.
By analysing the angular errors it can be seen that
the smallest MRE error was obtained for the angle
of the trial leg in P
2
(α
2
) and it is equal to 2.4%
(MAE = 1.6
) in sequence 1 and 4.1% (MAE = 2.9
)
in sequence 2, whereas the angle of inclination of the
torso in P
3
(γ
3
) is determined with the greatest er-
ror MRE = 40.1% (MAE = 16.4
) in sequence 1 and
MRE = 32.9% (MAE = 13.4
) in sequence 2. In the
Figure 3: Key points and parameters of hurdle clearance: P
1
is the braking point in take-off phase, P
2
is the propulsion point
in take-off phase, P
3
is the center of mass (CM) over the hurdle in flight phase, P
4
is the braking point in landing phase, P
5
is
the propulsion point in landing phase, h
1
is the height of CM, w
1
is the CM to hurdle distance, α
1
is the angle of the trial leg,
h
2
is the height of CM, w
2
is the CM to hurdle distance, α
2
is the angle of the trial leg, γ
2
is the angle of inclination of the
torso, h
3
is the height of CM (over the hurdle), β
3
is the angle of the lead leg, γ
3
is the angle of inclination of the torso, h
4
is
the height of CM, w
4
is the CM to hurdle distance, α
4
is the angle of the lead leg, γ
4
is the angle of inclination of the torso, h
5
is the height of CM, w
5
is the CM to hurdle distance, and α
5
is the angle of the lead leg.
Figure 4: Example of tracking results on sequence 1, frames #5, 19, 40, 57, 65; the green skeleton is the ground truth pose
and the white skeleton is the estimated pose.
work by (Krzeszowski et al., 2016), the estimation er-
rors for the mentioned parameters were MAE = 7.8
and MRE = 10.1% for α
2
, and MAE = 5.9
and
MRE = 12.1% for γ
3
.
It should be emphasised that in this paper the
errors were calculated for two particular sequences,
Table 1: Measured parameters and errors; units: h, w [mm], α, γ, β [
], MRE [%].
Sequence 1 Sequence 2
Parameter ¯x sd X MAE MRE ¯x sd X MAE MRE
P
1
h
1
961.7 8.2 953.5 9.2 1.0 966.2 20.3 959.1 14.8 1.5
w
1
2513 15.1 2502 13.5 0.5 2565 14.9 2550 17.3 0.7
α
1
66.4 2.7 62.4 4.0 6.4 64.6 3.7 61.4 4.2 6.9
P
2
h
2
1155 10.1 1159 8.7 0.8 1123 11.4 1130 11.5 1.0
w
2
1541 11.7 1525 16.2 1.1 1623 15.3 1605 18.4 1.1
α
2
67.5 2.1 67.2 1.6 2.4 69.8 2.5 72.0 2.9 4.1
γ
2
76.3 4.2 74.9 3.8 5.0 75.9 4.3 75.9 3.1 4.1
P
3
h
3
246.7 21.3 280.8 37.3 13.3 287.9 29.2 297.9 23.6 7.9
β
3
133.8 29.6 164.0 30.2 18.4 139.0 24.5 164.0 25.7 15.7
γ
3
57.2 5.4 40.8 16.4 40.1 53.9 8.2 40.8 13.4 32.9
P
4
h
4
1176 45.5 1099 80.1 7.3 1144 30.5 1100 43.2 3.9
w
4
1128 26.3 1175 46.9 4.0 1235 29.3 1296 60.8 4.7
α
4
73.2 6.9 70.4 6.0 8.6 74.7 3.9 68.0 6.8 10.0
γ
4
58.9 10.8 65.8 10.5 15.9 53.0 11.4 65.8 14.4 22.0
P
5
h
5
1118 84.9 1041 87.2 8.4 1094 40.3 1049 49.0 4.7
w
5
1670 71.1 1683 53.8 3.2 1777 52.4 1810 48.3 2.7
α
5
91.7 33.3 70.0 22.3 31.8 75.3 5.1 70.0 6.1 8.7
while in the work by (Krzeszowski et al., 2016) the
mean errors were obtained for 10 sequences.
4 CONCLUSIONS
This paper has proposed a multiview markerless
method to track human body motion. This method
was tested by experiments that were performed on
two image sequences of an athlete clearing a hurdle.
The estimated parameters were compared with the pa-
rameters obtained from ground truth poses. The er-
ror analysis indicates that the preliminary results are
promising but that further research is necessary. Con-
sequently, our further work will focus on a detailed
evaluation and improvement of the proposed method.
We will also concentrate on its use to support the tech-
nical preparation of hurdlers.
ACKNOWLEDGEMENTS
This work has been supported by the Polish Min-
istry of Science and Higher Education within the re-
search project ”Development of Academic Sport” in
the years 2016-2018, project No. N RSA4 00554.
REFERENCES
Cheng, F., Christmas, W., and Kittler, J. (2004). Periodic
human motion description for sports video databases.
In Proceedings of the Pattern Recognition, 17th Inter-
national Conference on (ICPR’04), volume 3 of ICPR
’04, pages 870–873, Washington, DC, USA. IEEE
Computer Society.
ˇ
Coh, M. (2003). Biomechanical analysis of Colin Jackson’s
hurdle clearance technique. New Studies in Athletics,
1:33–40.
ˇ
Coh, M., Dolenec, A., Toma
ˇ
zin, K., and Zvan, M. (2008).
Dynamic and kinematic analysis of the hurdle clear-
ance technique. In
ˇ
Coh, M., editor, Biomechanical
Diagnostic Methods in Athletic Training, pages 109–
116. University of Ljubljana.
Deutscher, J. and Reid, I. (2005). Articulated body motion
capture by stochastic search. Int. J. Comput. Vision,
61(2):185–205.
Elliott, N., Choppin, S., Goodwill, S. R., and Allen, T.
(2014). Markerless tracking of tennis racket motion
using a camera. Procedia Engineering, 72:344–349.
The Engineering of Sport 10.
Hamatani, T., Sakaguchi, Y., Uchiyama, A., and Higashino,
T. (2016). Player identification by motion features in
sport videos using wearable sensors. In 2016 Ninth
International Conference on Mobile Computing and
Ubiquitous Networking (ICMU), pages 1–6.
Iskra, J. (2012). Scientific research in hurdle races. AWF
Katowice.
John, V., Trucco, E., and Ivekovic, S. (2010). Markerless
human articulated tracking using hierarchical particle
swarm optimisation. Image and Vision Computing,
28(11):1530–1547.
Kennedy, J. and Eberhart, R. (1995). Particle swarm opti-
mization. In Proc. of IEEE Int. Conf. on Neural Net-
works, volume 4, pages 1942–1948. IEEE Press, Pis-
cataway, NJ.
Kim, Y. and Cho, K. S. (2016). Robust multi-object
tracking to acquire object oriented videos in indoor
sports. In 2016 International Conference on Infor-
mation and Communication Technology Convergence
(ICTC), pages 1104–1107.
Krzeszowski, T., Przednowek, K., Wiktorowicz, K., and
Iskra, J. (2016). Estimation of hurdle clearance pa-
rameters using a monocular human motion tracking
method. Computer Methods in Biomechanics and
Biomedical Engineering, 19(12):1319–1329. PMID:
26838547.
Kwolek, B., Krzeszowski, T., Gagalowicz, A., Woj-
ciechowski, K., and Josinski, H. (2012). Real-
time multi-view human motion tracking using particle
swarm optimization with resampling. In Perales, F.,
Fisher, R., and Moeslund, T., editors, Articulated Mo-
tion and Deformable Objects, volume 7378 of Lecture
Notes in Computer Science, pages 92–101. Springer
Berlin Heidelberg.
Panagiotakis, C., Grinias, I., and Tziritas, G. (2006). Auto-
matic human motion analysis and action recognition
in athletics videos. In 14th European Signal Process-
ing Conference, pages 1–5.
Per
ˇ
s, J. and Kovacic, S. (2000). A system for tracking
players in sports games by computer vision. Elek-
trotehni
ˇ
cni vestnik, 67(5):281–288.
Ramasso, E., Panagiotakis, C., Rombaut, M., Pellerin, D.,
and Tziritas, G. (2009). Human shape-motion analy-
sis in athletics videos for coarse to fine action/activity
recognition using transferable belief model. Elec-
tronic Letters on Computer Vision and Image Anal-
ysis, 7(4):32–50.
Reyes, C. E., Mojica, E. F., Correa, C. V., and Arguello, H.
(2016). Algorithm for underwater swimmer tracking
using the HSV color model and compressive sensing.
In 2016 IEEE Colombian Conference on Communica-
tions and Computing (COLCOM), pages 1–5.
Salo, A., Grimshaw, P. N., and Marar, L. (1997). 3-D
biomechanical analysis of sprint hurdles at different
competitive levels. Medicine and Science in Sports
and Exercise, 29(2):231–237.
Sheets, A. L., Abrams, G. D., Corazza, S., Safran, M. R.,
and Andriacchi, T. P. (2011). Kinematics differences
between the flat, kick, and slice serves measured us-
ing a markerless motion capture method. Annals of
Biomedical Engineering, 39(12):3011–3020.
Sidenbladh, H., Black, M. J., and Fleet, D. J. (2000).
Stochastic tracking of 3D human figures using 2D im-
age motion. In European Conf. on Computer Vision,
pages 702–718.
Tsai, R. (1987). A versatile camera calibration technique
for high-accuracy 3D machine vision metrology using
off-the-shelf tv cameras and lenses. IEEE Journal on
Robotics and Automation, 3(4):323–344.
Zhang, Y., Feng, S., Sun, X., and Yang, H. (2017). Re-
search on tracking algorithm for fast-moving target in
sport video. Journal of Computational and Theoreti-
cal Nanoscience, 14(1):230–236.