Share this Abstract

Rate this Abstract

Login to allow rating


Login to allow views


Abstract Details

Abstract Title

Comparison of Two Methods For Tracking a Tennis Player Motion in Competition

Abstract Theme

Technology in sports

Type Presentation


Abstract Authors

Presenter Carina Pimentel Santamaria - University of Campinas (Faculty of Physical Education ) - BR
Heber Teixeira Pinto - University of Campinas (Faculty of Physical Education ) - BR
Afonsa Janaina Silva - University of Campinas (Faculty of Physical Education ) - BR
Cláudio L. R. Vieira - University of Campinas (Faculty of Physical Education ) - BR
Tiago. J. C Pereira - State University of Londrina (Physical Education Department ) - BR
Milton S. Misuta - University of Campinas (School of Applied Science ) - BR
Felipe A. Moura - State University of Londrina (Physical Education Department ) - BR
Ricardo M. L. Barros - University of Campinas (School of Applied Science ) - BR

Presentation Details

Poster Exhibition Site (Local): Gold - 3        Date: 1 September        Time: 8am to 7pm        Presenter: Carina Pimentel Santamaria

Abstract Resume

Nowadays one of the most used method to obtain information about a sport player motion in competition is the 2D automatic tracking based on video images. Such methods apply computer vision techniques
to extract the player position in function of time and allow the calculation of the total distance coverage by the athlete, its instantaneous and average speed among many other important variables
useful to characterize the player performance. After the pioneering studies in soccer, some recent papers used such 2D methods for tracking tennis players. However, no study was found in the
literature comparing the results of  such 2D tracking method with a more accurate and reliable 3D method in tennis. Therefore, this study intends to compare the results provided by a 2D and a 3D
player tracking methods during an official tennis match of the ATP 250 World Tour. The designated 3D method used four fixed cameras (JVC, 60 fields/second), located at higher positions of the bench
gymnasium. The cameras were calibrated (DLT) using 72 control points distributed in a volume of 11x11x2.6m in one of the court halves. The accuracy of the 3D method was assessed through a rigid bar
test (912.3 mm) was moved throughout the calibrated volume. The maximum error found was of 38mm, with a mean error of 3 mm and a RMSE of 12 mm. The 2D method used one fixed video camera (Casio, 30 Hz)
to capture and process the images. The camera was calibrated with a 2D method that used 6 control points measured at the plane of the court. The 2D tracking method was tested with three different
post-processing procedures. The first one tracked the player and delivered the results fully automatically (2D automatic). The second one applied the automatic tracking followed by a procedure using a
low-pass digital filter (Butterworth 2nd order, 0.375 Hz cutoff frequency), as described in the literature. The third procedure applied the automatic measures and correcting manually by a human
operator when, eventually, discrepancies were detected (2D manual correction). Taking the results provided by the 3D method as the ground truth, the results delivered by the three other 2D procedures
were compared. The worst results were obtained by the 2D automatic procedure, in x-coordinate, with a R²=0.83, and a maximum error of 1.19 m. The filtering procedure improved slightly the results with
R²=0.90 a maximum error of 1.14 m. The 2D manual correction procedure presented the better results, with a R²=0.96 and maximum error of 0.77 m. Although the regression analysis comparing 2D procedures
against the 3D method showed relatively high coefficients of determination (over 0.83), meaning that 2D methods can explain a high percentage of the variability, errors of over 1 meter in the player’s
position were found. Taken into account that distances between positions are used to calculate cumulative variables such as distance covered, the results have to be taken carefully and requires
further analysis to conclude about the equivalence of 2D and 3D methods.

Comment this abstract (0 comments)