Computer-assisted Table Tennis Posture Analysis using Machine Learning

Abstract

Manual assessments for table tennis players, done in person or virtually, can be tedious, inefficient, and error-prone. Existing machine learning software tries to eliminate these gaps; however, its capability is only limited to one technical skill at a time. In this study, a software was developed to help assess the key technical skills of a table tennis player: (1) upper body position (leaning and not leaning), (2) lower body position (knees bending and straight), (3) basic hand strokes (forehand and backhand), and (4) footwork (side-to-side and in-and-out); these four would be used as performance metrics for the video input. Datasets of five OpenTTGames videos depicting professional player's postures and three videos from YouTube portraying amateur player's postures were extracted into frames, resulting to 32,395 frames. Posture detection was first carried out on the extracted frames using OpenPose library, generating a total of 519,320 key points. Then, various machine learning models were trained using the key points for posture analysis, and their performances were compared and benchmarked. The models with the highest accuracies would be integrated into one assessing model. Among Backpropagation, SVM-Linear, SVM-RBF, and SVM-Polynomial models, the SVM-RBF model yielded the highest in all performance metrics: 95.03% for upper body, 95.28% for lower body, 95.72% for hand stroke, and 92.78% for footwork. These results indicated that the software successfully assessed the player's posture, providing relevant data for coach's assessment of the player's performance. This software will help coaches and players analyze and evaluate their performance for improvements in lacking areas.