Machine Learning Based Automatic Effective Round Segmentation Method for Table Tennis

Table tennis is a popular sport worldwide, especially in China. To win in a game, most amateurs, professional athletes, and coaches have to watch and analyze competition videos to accumulate experiences and lessons for improving skills. However, most frames in a table tennis competition video are useless for technical and tactical analysis, requiring the video to be segmented into pieces and to keep pieces with effective playing actions for analysis, which are called effective rounds. Nowadays, this kind of segmentation works are performed manually, which is expensive and inefficient. To solve this problem, we propose and implement an automatic effective round segmentation method for table tennis based on machine learning techniques. In this method, we design an action recognition module and a key frame discrimination module to identify specific key frames of the video stream. Based on these key frames, we design an automatic segmentation module to combine effective frames as effective rounds of a competition video. To validate the proposed method, we build a human action dataset, a table tennis key frames dataset, and a competition video dataset, and conduct extensive experiments. The evaluation results with high-performance indicators demonstrate the effectiveness and efficiency of the proposed method.