Detection of Players on a Soccer Team

based on Informed Filters using Only Color Features

Takuro Oki

and Ryusuke Miyamoto

Department of Fundamental Science and Technology,

Graduate School of Science and Technology, Meiji University, Kanagawa 214-8571, Japan

Department of Computer Science, School of Science and Technology, Meiji University, Kanagawa 214-8571, Japan

1 INTRODUCTION

Semantic analysis of sports videos has become an ac-

tive research topic. Player detection on the ﬁeld is

a particularly important technique for various appli-

cations that are essential for semantic event detection

and tactical analysis, such as calculations of the dis-

tance covered by a player during a soccer match. Tra-

cab(ChyronHego, 2003) is one of the most famous

systems that can visualize the statistics of players’

performance during a match. However, the current

detection and tracking systems used for Tracab are

very large and expensive, so they are only found at

large stadiums. Many major teams require this sys-

tem when they play a match.

To solve this problem, we tackled this task by

using a simple monocular camera and developed a

highly accurate soccer player detection method using

only color features(Miyamoto and Oki, 2016). This

method is based on a simple sliding window algo-

rithm, but it does not use background subtraction or

inter-frame difference. This is because they are not

appropriate for moving cameras, though our system

has to operate properly for aerial photographs taken

by drones.

In our previous work(Miyamoto and Oki, 2016),

we tried to ﬁnd all humans on the ﬁeld including

coaches and referees. However, for team tactics and

player activity analysis, it is more useful to detect

only players that belong to a certain team. There-

fore, in this paper, we improve the previous method

(Miyamoto and Oki, 2016) and try to enable soccer

players to be accurately detected on the basis of their

teams.

2 INFORMED HAAR-LIKE

FEATURES

Informed Haar-like features (Zhang et al., 2014) en-

ables accurate human detection by representing the

object boundary properly. They possess two unique

points: a well-designed feature pool for construction

of a classiﬁer and computation of features using bi-

nary and ternary template models. A binary model

computes feature values using two types of rectan-

gles as coefﬁcients: −1 and +1. Its basic idea is

the same as Haar-like features proposed by (Viola

and Jones, 2001). A ternary model is applied to rep-

resent more complex geometric conﬁgurations than

the binary model and has three types of rectangle as

weights: −1, +1, and 0.

Filtered Channel Features(Zhang et al., 2015) us-

ing a feature pool including more complex templates

has topped the state-of-the-art accuracyfor human de-

tection and outperformed recent schemes based on

deep learning.

3 DETECTION OF PLAYERS ON

A SOCCER TEAM BASED ON

INFORMED FILTERS USING

ONLY COLOR FEATURES

Our previous proposal(Miyamoto and Oki, 2016) can

accurately detect people using only color features if

they are appropriately selected, but does not use his-

tograms of oriented gradients. This method targets

all people shown in the image without considering to

which team they belong. However, to obtain more

useful information for tactical analysis, players be-

longing to a certain team need to be detected. Thus,

we construct a classiﬁer that enables detection of

players on a certain team. To construct a classiﬁer, we

generate training samples and deﬁne samples that in-

clude a target team’s player as positive and the others

as negative. In addition, the goal keepers are excluded

from the detection target. To calculate feature value,

we generate templates and used only color features

like in our previous work(Miyamoto and Oki, 2016).

Oki, T. and Miyamoto, R.

Detection of Players on a Soccer Team based on Informed Filters using Only Color Features.

In Extended Abstracts (icSPORTS 2016), pages 27-28

Figure 1: Detection results in the same frame.

4 EVALUATION

In the experiments, miss rate vs false positives per im-

age was measured as a detection error tradeoff (abbr.

DET) curve to evaluate the classiﬁcation accuracy. In

addition, processing speed was also measured. For

this evaluation, 2000 images were randomly selected

for training samples from the PETS2003 dataset, and

500 images not included in training samples were also

randomly selected and used as test images.

Fig.1 shows detection examples. Our proposal can

extract all players belonging to a certain team in some

frames.

Fig.2 shows DET curves. In the ﬁg.2, “Liverpool”

and “Fulham” represent the results for both teams.

“Liverpool” results shows that the miss rate was about

3.0% at 0.1 FPPI, and “Fulham” results shows that

the miss rate was about 5.0% at 0.1 FPPI. These re-

sults shows that our proposed method can achieve an

acceptable detection accuracy.

-2

-1

-3

-2

-1

Miss rate (MR)

False positives per image (FPPI)

Liverpool

Fulham

Figure 2: DET curves for both teams.

5 CONCLUSION

In this paper, we have tried to extend our previously

proposed method(Miyamoto and Oki, 2016) to the

detection of soccer players that belongs to a speciﬁc

team. If features selection is appropriately operated,

experimental results using PETS2003 dataset show

that our proposed method can achieve high detection

accuracy. In the future, we will try to improve our

proposed method by adding an object tracking pro-

cess using time series information.

ACKNOWLEDGEMENTS

The research results have been achieved thanks to

“Research and Development of Innovative Network

Technologies to Create the Future”, the Commis-

sioned Research of National Institute of Information

and Communications Technology (NICT), Japan.

REFERENCES

ChyronHego (2003). Tracab optical tracking. http://

chyronhego.com/sports-data/tracab.

Miyamoto, R. and Oki, T. (2016). Soccer player detection

with only color features selected using informed haar-

like features. In Advanced Concepts for Intelligent

Vision Systems. to be published.

Viola, P. and Jones, M. (2001). Rapid object detection using

a boosted cascade of simple features. In Proc. IEEE

Conf. Comput. Vis. Pattern Recognit., volume 1, pages

511–518.

Zhang, S., Bauckhage, C., and Cremers, A. (2014). In-

formed haar-like features improve pedestrian detec-

tion. In Proc. IEEE Conf. Comput. Vis. Pattern Recog-

nit., pages 947–954.

Zhang, S., Benenson, R., and Schiele, B. (2015). Filtered

channel features for pedestrian detection. In Proc.

IEEE Conf. Comput. Vis. Pattern Recognit., pages

1751–1760.

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support