Automatic swimmer tracking using video sequences: Application to performance analysis
Djamel-Eddine Benarab, T Napoleon and A Alfalou
Higher Institute for Electronics and Digital Training, France
: J Comput Eng Inf Technol
Abstract
In order to improve the performance of professional swimmers, we developed, in collaboration with the FFN (French Swimming Federation), an automatic tracking approach using video sequences. To do this, we proposed a new 8K shooting system adapted to the aquatic environment and allowing a pool calibration using the DLT technique (Direct Linear Transform). This establishes the link between pixel and metric coordinates, which allows, among others, to extract the concerned lane to carry out the different treatments. In order to initialize the tracking, it is necessary to localize the swimmer. For this, we applied a contrario technique to detect his movement in the lane. Then, the initial reference (swimmer’s head) of our tracking system is detected using the proposed scaled composites JTC which is based on a pre-generated database. Afterwards, we implemented and adapted several tracking techniques, namely those based on the NL-JTC correlation, color histograms, Local Binary Patterns (LBP) and Histograms of Oriented Gradient (HOG). Given the various limitations of these techniques, we proposed an optimized approach based primarily on data fusion. This approach is called dynamic fusion and it combines NL-JTC correlation plane and color scores plane in order to generate a richer description of the target (form + color). This approach has shown very good results in the cases of visible target but it is still sensitive to occlusions. To solve this problem, we improved it by tracking simultaneously the head and the swimsuit of the athlete. This multi related targets approach enables, through a complex decision criterion, to find the occluded zone based on the visible one. Finally, a swimmer performance study is conducted and the results allowed validating the proposed system. In particular, we were interested in cyclical, intra-cyclical and instantaneous speed measurements, to study and improve the swimmers’ performance.
Biography
Email: djamel-eddine.benarab@univ-lorraine.fr