Research on Basketball Player Tracking Method Based on Multi-Feature Fusion

Multi-object tracking is a hot topic in the field of computer vision, which is widely used in intelligent transportation, intelligent surveillance, autonomous driving, and other fields. With the rapid development of the sports industry, the problem of athlete tracking in sports has gradually attracted attention. Unlike pedestrian tracking in ordinary surveillance scenarios, athlete tracking in sports scenarios faces great challenges due to frequent occlusion, large pose changes, and the similar appearance of athletes. This article takes basketball player tracking as the research object, focusing on the methods of athlete detection and tracking. To address the problem of low accuracy due to the similarity of athletes’ appearance, this paper proposes a basketball player tracking algorithm based on multi-feature fusion. First, a multi-branch re-identification model for athletes is designed. The global branch learns the global semantic information of the athletes; the local branch horizontally partitions the body parts of the athletes to learn fine-grained features; the feature deletion branch deletes the most activated part of the features, forcing the network to learn the remaining significant features. Then, the features of the three branches are merged to improve the discriminability of the athletes. Finally, experiments are conducted on the dataset, and the results show that the proposed method achieves good performance.