Aggregation of Tennis Groundstrokes on the Basis of the Choquet Integral and Its Generalizations

This paper focuses on the aggregation of two tennis groundstrokes: Forehand and backhand, which are performed in every match and training session. Recognition of tennis movements is very challenging task. Therefore it is important to correctly classify the proper patterns of various tennis players. This kind of study may be a part of training or development of a player’s skills. In this study, Support Vector Machines, Multilayer Perceptron, and Spatial-Temporal Graph Convolutional Neural Networks are applied to find a match and put the move trajectories into the proper classes. The images obtained from three dimensional data recorded using the Vicon optical motion capture system are the input data for the classifiers. The images containing forehand and backhand shots are divided into two phases: The preparation and the shot together with the racket swinging as the finishing element of the move. Due to the problems with strict classification of these data, the input is also fuzzy. In order to improve the accuracy of the forehand and backhand recognition, the non-typical Choquet-like integral aggregation functions are applied as well as traditional aggregation operators like median, or voting. In particular, the so-called pre-aggregation operators with specific t-norms, overlap functions and modification of the shape of the function under the integral sign give the best results, reaching accuracy at the level of 98.88%, which is better than the above mentioned individual classifiers.