Human Motion Detection in Swimming Motion Video Based on Multiscale Separation Spatio-Temporal Attention Mechanism.

Swimming motion video human motion detection is becoming increasingly important in sports training and event analysis. Existing methods are deficient in dealing with complex underwater environments and rapid changes in swimming movements, and the accuracy and real-time performance of motion detection are low. Therefore, the study proposes a human motion detection method for swimming motion video based on multiscale (MS) separation spatio-temporal attention mechanism (STAM). The encoder-decoder architecture extracts and fuses features of different scales in both spatial and temporal dimensions to realize automatic detection and precise localization of swimming motion. The experimental results indicated that the feature extraction accuracy reached 97.34% after 43 iterations, and the feature importance reached 0.982 after 40 iterations. In terms of recognition accuracy, the average accuracy of the model reached 94.02%, the recall rate was 93.09%, and the F1 score was 93.56%. Adaptive testing of movement changes showed that the detection accuracy generally remained above 89%, and the accuracy in slow and large-sized movements even exceeded 95%. In addition to increasing swimming action detection's precision and resilience, the work offers technological and theoretical backing for the creation of intelligent sports analysis systems.