Tennis Multivariate Time Series Clustering

In tennis there are two basic shots (forehand and backhand), which are two of key elements to win points. Sophisticated equipments, such as motion capture systems, enable one to record both the tennis player's movements and tennis racket. The 3D data may be used to define the perfect shot model or to give the directions to the player how to reach to this model and which aspects of impact should be improved. Clustering analysis can result in understanding the phases of a tennis player move and, as a consequence, the improvement of his/her play. Using the memberships obtained in the fuzzy clustering process one can evaluate the quality of a player's move and potentially estimate the player's progress. The main objective of this study is to apply the fuzzy c-means algorithm utilizing the dynamic time warping-based distance to cluster analysis of tennis shots. Both shots were taken into the consideration. The analysis consists of forty moves. Based on the 3D data of the tennis racket positions, the clustering was performed for subsequent two, three, and four clusters. The obtained results clearly show that clustering with two clusters is the most appropriate for analysing tennis shots. The model of a perfect shot was obtained. It is universal and does not depend on the player's height. Based on the model, it is possible to deduce technical differences in the players' shots. This analysis gives the directions for improvements of the shot technique. The advantage of the clustering of our approach is that we can get information to what degree the athlete should still correct his/her shots. The information is given to what extent the stroke is correct in relation to the ideal model.