Group behavior detection in volleyball matches based on multi-scale transformer and dynamic attention

With the rapid development of sports video intelligent analysis technology, existing volleyball group behavior detection models still have various difficulties in multi-action concurrent modeling, dynamic interaction relationship characterization, and complex offensive and defensive rhythm understanding. Problems include insufficient semantic granularity, unstable spatiotemporal correlation, and poor perspective adaptability, as well as difficulty in meeting the needs of intelligent tactical analysis in real games. Therefore, this study proposes a volleyball group behavior detection algorithm framework that integrates multi-scale Transformer and dynamic attention mechanisms. This framework constructs multi-layer representations of individual actions, collaborative relationships and formation structures through three types of semantic tokens, and combines adaptive dynamic attention to achieve precise focus on key action chains and high-impact collaboration patterns. Experimental results demonstrated that the model achieved F1 scores of 95.59% and 93.38% on the VD and VTE datasets, respectively, with recognition latency of 17.4 ms and 18.2 ms. Among the selected comparison methods, ST-GCN, ConvLSTM, and Video Swin, the proposed model exhibited superior stability in key action recognition under occlusion, multi-person interaction modeling, and cross-view adaptability, achieving the highest overall score of 9.12. Across six view offset levels ( ± 5, ± 10, ± 15, ± 20, ± 25, and ± 30) and three illumination intensity conditions, the average accuracy remained above 87.5%. Compared to traditional spatiotemporal modeling approaches, the multi-scale Transformer simultaneously models local actions and overall formation relationships, while its dynamic attention mechanism enhances key action detection under complex occlusion and view variation scenarios, improving group behavior recognition stability while maintaining low latency. The proposed model has advantages in dynamic scene understanding, interactive relationship modeling, and robust performance. This paper can provide key technical support for intelligent tactical analysis systems, team collaboration evaluation, and sports big data platforms.