Due to the fact that most of the violations are instantaneous actions, they are easily obscured by redundant information from a large number of compliant actions, resulting in poor feature discrimination at different distances and low recognition accuracy. Therefore, a mixed attention mechanism is proposed to conduct research on the recognition algorithm of illegal movements in competitive sports. Background subtraction is used to locate the athlete's area, and Gaussian filtering and downsampling techniques are combined to construct a scale space to adapt to motion features at different distances. Hessian determinant is used to extract spatiotemporal and multi-scale features, and high-level features are fused through bidirectional transmission to improve feature discrimination. Spatial attention focuses on key parts of the limbs, temporal attention locks in keyframes of illegal actions, channel attention enhances effective feature channels, and the three work together to output enhanced features. Finally, the end-to-end training optimization model is used, combined with sequence distance judgment to distinguish similar actions and achieve the recognition of illegal actions. The experimental results show that the algorithm can effectively separate similar action features, with a recognition accuracy of over 99% for illegal actions, an average recognition accuracy of 99. 81% , which are better than the comparative algorithms and have greater application value.

hybrid attention mechanism, competitive sports, illegal actions, action recognition, spatial and channel attention