Detection of Single Ball Contacts using a Radio-based Tracking

System

A Basis for Technical Performance Analysis

Nicolas Witt

, Matthias Völker

and Björn M. Eskofier

Digital Sports Group, Friedrich-Alexander-University-Erlangen-Nuernberg (FAU), Erlangen, Germany

Time-based Radio Location Group, Fraunhofer Institute for Integrated Circuits (IIS), Nuremberg, Germany

1 OBJECTIVES

Nowadays, many systems provide positional data of

players in major ball sports like soccer, basketball or

American football. Position data in soccer results

from video tracking during official games (Di Salvo,

2006) or from radio-based systems in training

sessions (Frencken, 2010). The existing systems are

used for a physical assessment of players. However,

higher rate tracking systems for players and

especially for balls could additionally enable a detai-

led technical assessment of athletes. There also have

been attempts to extract possession related actions

from official tracking data (Hoernig et al., 2016).

This work investigates the detection of single

ball contacts, distinguishing left and right foot, using

a radio-based real-time tracking system, (v. d. Gruen

et al., 2011) for proving its suitability for technical

performance analysis in training sessions and games.

2 METHODS

This section first describes the tracking system and

the test environment, followed by an overview on

the different scenarios and the gold standard that is

used for testing. Finally the detection algorithm used

for extracting ball contacts is explained as well as

the statistics to evaluate its performance.

2.1 Tracking System and Test Location

The RedFIR Real-Time Locating System (RTLS)

from the Fraunhofer Institute for Integrated Circuits

provides highly accurate and high frequent tracking

data that is based on time-of-flight and phase

measurements (v. d. Gruen et al., 2011 and

Mutschler et al., 2013). Small transmitters emit burst

signals, which are received by antennas around the

tracking area. The tracking data is computed from

time differences of arrival (TDoA) using hyperbolic

trilateration. The locating system is able to generate

an overall of 50,000 tracking samples per second.

This includes athlete tracking data with 200 Hz for

every transmitter (size: 50 x 33 x 6 mm) and ball

data with 2000 Hz for every ball. Each tracking

sample consists of position (x, y, z), velocity (v

, v

) and acceleration (a

, a

) values in the local

coordinate system. The system can simultaneously

track twelve balls and 126 player transmitters.

Figure 1: Mounting points of player and ball transmitters.

The tracked transmitters are attached to the

athletes’ left and right shins (see Figure 1) by

compression tubes. All experiments are conducted in

an indoor environment called L.I.N.K. (Eidloth et

al., 2014), a multipurpose hall for testing tracking

systems and application scenarios.

2.2 Datasets for Game and Training

Two different scenarios are regarded, to evaluate the

ball contact detection in training and game settings.

944 ball contacts are played in total during both

scenarios.

We record all sessions with a standard

camcorder at 50 fps. The ground truth for ball

contacts (point in time and left or right foot) is

derived from the video footage by a four eyes

Witt, N., Völker, M. and Eskoﬁer, B.

Detection of Single Ball Contacts using a Radio-based Tracking System - A Basis for Technical Performance Analysis.

In Extended Abstracts (icSPORTS 2016), pages 19-22

(analysts are certified soccer trainers) principle. A

frame by frame analysis, leads to a precise point in

time for every contact. The training and game

scenarios each consists of two different parts:

Training scenario: Two different exercises are

played by two players each. One exercise with 12

repetitions was a double passing exercise with a shot

on goal at the end. There are 70 ball contacts in the

ground truth for this exercise. The second recording

session is a continuous passing exercise around a

square (10 x 10 meters) containing 125 contacts.

Game scenario: Three sessions (each 4 minutes)

of FUNiño (Wein, 2016) are played on a field of 20

x 26 meters, with 404 contacts in total. On the same

pitch three sessions of a ball possession exercise are

recorded (3 vs. 3 players), resulting in 345 annotated

ball contacts. As teams could only score points after

eight consecutive ball possessions, this exercise

should enforce more duels, representing complex

situations for automatic contact detection. The game

scenario contains 749 contacts in total.

Figure 2: The two steps of the algorithm detected hits for

player 5 and 6. Feet are indicated by R and L.

2.3 Detection Algorithm

We use a straightforward algorithm to detect single

ball contacts (also referred as hits) in the tracking

data. The algorithm includes two major steps:

Step 1: Find threshold crossings of ball

acceleration in the x,y-plane to get possible times of

ball contacts (th

acc

= 48 m/s

was used).

Step 2: Assign the nearest foot to the ball at the

acceleration crossings that is closer than a certain

radius (we used r

near

= 0.6 m).

The first step detects possible hits where the ball

changes its direction and/or velocity. The second

step identifies the foot the ball was played with. It

also eliminates impossible hits, where no foot was

involved. Both steps are illustrated in Figure 2. The

threshold values are the results of a grid search, on

the game scenario, optimizing for the F1-score.

Together with the fact that there was a ball hit,

its time and location (x,y-coordinate) gets extracted.

The detected action (referred as event) also carries

the identification of the player as well as additional

attributes (e.g. foot identifier and peak acceleration).

Thus the event can be used for further higher-level

analyses. Examples are measuring the ball-

processing time as the time difference between ball

reception (first hit) and passing (last hit) in training

exercises. Also extracting possessions, passes and

shots during matches or find differences in ball

handling skills for the left and right foot become

possible.

2.4 Evaluation Measures

To evaluate the detection performance, recall and

precision rates are provided (as in Huang, 2011). To

compute these statistics, four different counts are

used besides the number of hits in the ground truth

(GT). The wrong detected hits (WD) are those that

are detected, but with a wrong assignment of the

foot, e.g. during a close duel. The false detected

contacts (FD) are not present in the ground truth, but

the algorithm detected a hit, e.g. a bouncing of the

ball near a foot. Missed thresholds can be a reason

for not detected hits (ND). False positives in the

computation of precision and false negatives for the

recall are computed as the sum of FD and WD. The

F1-score mentioned in the algorithm section was

computed as the geometric mean of precision and

recall.

3 RESULTS

The description of the results will be split according

to the training and the game scenario.

The training scenario results in Table 1 show an

optimal result for the double passing exercise. All 70

contacts were detected and also the correct foot was

identified for every hit.

Table 1: Results for the training scenario.

Exercise Recall Precision

Double passing 100 % 100 %

Passing (square) 87 % 97 %

In the passing exercise around the square, a

precision of 97 % is reached and a lower recall of 87

%. The low recall results from 16 contacts that were

not detected. These contacts were gently played with

the sole of the shoe, hardly changing the direction of

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

the ball. From the 125 contacts played, 107 were

correctly detected. For two contacts the wrong foot

was assigned. One contact was detected false.

The aggregated results for the three FUNiño

sessions in Table 2 show a precision of 87% and a

recall of 92%. From the 404 contacts in the ground

truth, 358 were detected correct. 36 hits were

detected false and for 16 the wrong foot was

assigned. 36 contacts were not detected. The results

for the ball possession session are slightly higher

with a recall of 93% and a precision of 90 %. The

algorithm correctly detected 313 from the 345

contacts in the ground truth. The numbers of false

and wrong detected hits are 24 and 10 respectively.

22 contacts were not detected. In total, the two

sessions show a precision of 89% and a recall of

93% for the whole game scenario, where 671 of 745

contacts were correctly detected.

Table 2: Results for the game scenario.

Session Recall Precision

3 x FUNiño

92 % 87 %

3 x Possession 93 % 90 %

Overall 93 % 89 %

4 DISCUSSION

As applications for using detected contacts for

training and game analyses can be different, the

results are discussed separately for the two

scenarios.

4.1 Training

The results show an optimal detection for the non-

continuous exercise (double passing with a shot on

goal). The reason for that are the straightforward

tasks without opponent intervention, so players

execute very clear contacts, also during dribblings.

For the passing around the square exercise there

were gentle hits that were not detected. In

accordance with soccer trainers these contacts are

not crucial for an assessment in training. For

automatic training applications, the high detection

performance enables a variety of automatic ball

handling analyses. Examples are rating passing

precision, speed of dribblings or proximity to the

ball during exercises, to get objective measures for

technical skills. Also simpler analyses for the

footedness of a player over training sessions become

easy.

4.2 Game

As can be seen from the game scenario results, a

number of hits are detected wrong. For higher level

analyses (e.g. passes and shots) where only an

assignment to a player, not to a certain foot is

necessary, this is not a problem. To further avoid

false detected contacts, adaptive thresholds could be

a possible improvement. Not detected hits mainly

appear during longer dribblings. The same holds for

wrong detected ones. Those ball contacts seem not

to be crucial in game analysis. It can be expected

that the great majority of higher level actions can be

automatically detected with the presented ball

contact detection as a basis. This would enable an

objective assessment of technical skills as well as

reducing the manual effort for annotations.

ACKNOWLEDGEMENTS

This contribution was supported by the Bavarian

Ministry of Economic Affairs and Media, Energy

and Technology as a part of the Bavarian project

'Leistungszentrum Elektroniksysteme (LZE)'.

REFERENCES

Di Salvo, V., Collins, A., McNeill, B., Cardinale, M.,

2006. Validation of Prozone: A new video-based

performance analysis system. Int. Journal of

Performance Analysis in Sport, 6(1), pp. 108–119.

Frencken, W.G.P., Lemmink, K.A.P.M., Dellemann, N.J.,

2010. Soccer-specific accuracy and validity of the

local position measurement (LPM) system. Journal of

Science and Medicine in Sport, 13, pp. 641–645.

Hoernig, M., Link, D., Herrmann, M., Radig, B., Lames,

M., 2016, Detection of Individual Ball Possession in

Soccer, Proc. of the 10th Int. Symposium on Computer

Science in Sports, pp. 103-107.

v. d. Gruen, T., Franke, N., Wolf D., Witt N. and Eidloth,

A., 2011. A Real-Time Tracking System for Football

Match and Training Analysis. In: Microelectronic

Systems: Circuits, Systems and Applications. Springer,

Berlin Heidelberg, pp. 193-206.

Mutschler, C., Ziekow, H., Jerzak, Z., 2013. The DEBS

2013 Grand Challenge. Proceedings of the 7th Int.

Conf. on Distributed Event-Based Systems. Airlington,

Texas, pp. 283-294.

Eidloth, A., Lehmann, K., Edelhaeusser, T., v. d. Gruen

T., 2014. The Test and Application Center for

Localization Systems L.I.N.K. Int. Conf. on Indoor

Positioning and Indoor Navigation, Busan, Korea.

Huang, Q., Cox, S., Yan, F., de Campos, T., Windridge,

D., Kittler, J. and Christmas, W., 2011. Improved

Detection of Single Ball Contacts using a Radio-based Tracking System - A Basis for Technical Performance Analysis

detection of ball hit events in a tennis game using

multimodal information, 11th Int. Conf. on Auditory-

Visual Speech Processing, pp. 123-126.

Wein, H., 2016. Spielintelligenz im Fussball: Kindgemaeß

trainieren, Meyer & Meyer. Aachen, 4

edition.

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