Volleyball Motion Analysis Model Based on GCN and Cross-View 3D Posture Tracking

Abstract

The tracking of motion targets occupies a central position in sports video analysis. To further understand athletes’ movements, analyze game strategies, and evaluate sports performance, a 3D posture estimation and tracking model is designed based on Graphical Convolutional Neural Network and the concept of “cross-vision". The outcomes revealed that the loss function curve of the 3D tracking model designed for the study had the fastest convergence with a minimum convergence value of 0.02. The average precision mean values for the four different publicly available datasets were above 0.90. The maximum improvement reached 21.06% and the minimum average absolute percentage error was 0.153. The higher order tracking accuracy of the model reached 0.982. Association intersection over union was 0.979. Association accuracy and detection accuracy were 0.970 and 0.965 respectively. During the volleyball video analysis, the tracking accuracy and tracking precision reached 89.53% and 90.05%, respectively, with a tracking speed of 33.42 fps. Meanwhile, the method's trajectory tracking completeness was always maintained at a high level, with its posture estimation correctness reaching 0.979. Mostly tracked and mostly lost confirmed the tracking ability of the method in a long time and cross-view state with high model robustness. This study helps to promote the development and application of related technologies, promote the intelligent development of volleyball in training, competition and analysis, and improve the efficiency of the sport and the level of competition.