The Occlusive Basketball Player Detection Algorithm Based on Posture Recognition Assisted Feature Alignment

Abstract

A trajectory prediction algorithm for basketball players under complex occlusion conditions is proposed to address issues such as fast transition between attack and defense and severe occlusion in basketball. The complex occlusion video scene of basketball is taken as the research object. The spatial information and depth appearance features of athletes are fused. The proposed player detection algorithm enhanced accuracy on three different backbone networks, with an increase of 3.6%, 2.5%, and 2.9% compared to traditional methods. The maximum processing speed on the backbone feature extraction network-34 was 23.4 frames per second, which was significantly improved compared to other algorithms. The player re-identification algorithm achieved a rank-k accuracy of 0.7851 and a mean average precision of 0.445 in the top rank. For cumulative matching curves, the re-identification algorithm’s recognition accuracy was the highest in severely occluded environments. The occlusive basketball player detection algorithm based on the fusion of spatial information and deep appearance features was validated. The multi-objective tracking algorithm that combined computer vision and deep learning still lagged behind this research algorithm in higher order tracking accuracy by 0.027. These results have important application value for predicting the movement trajectory of basketball players in complex occlusion scenes.