A Vision Transformer-Based Model for Basketball Tactics Recognition Using Swarm Intelligence

Abstract

In order to solve the problem of poor classification performance of traditional algorithms for basketball tactics, we propose a scientific training model for basketball tactics computer swarm intelligence algorithm. We designed a basketball tactical recognition model (TacViT) based on player trajectory data in NBA games. The TacViT model employs a Vision Transformer (ViT) as its backbone network. It utilizes a multi-head attention module to extract rich global trajectory features and integrates a trajectory filter to enhance the interaction of feature information between the court lines and player trajectories. The trajectory filter learns long-term spatial correlations in the frequency domain with logarithmic linear complexity, thereby improving the representation of player position features. We transformed the sequence data from the sports vision system (SportVU) into trajectory maps and constructed a basketball tactical dataset (PlayersTrack). The experimental results demonstrate that TacViT achieves an accuracy of 81.4%, which is 15.6% higher than the unmodified ViT-S model. Additionally, TacViT exhibits superior performance in precision, recall, and computational efficiency. The PlayersTrack dataset contains 10,000 trajectory images, each with a resolution of 256x256 pixels. The TacViT architecture introduces a novel trajectory filter module and a multi-head attention mechanism, which together enable efficient feature extraction. Key evaluation metrics include accuracy, precision, recall, and FLOPS. These results highlight the significant improvement in classification performance for basketball tactics recognition.