Athletes' action recognition and performance prediction based on TS-GCN-SAM

This study aims to meet the modern competitive sports' demand for dynamic adjustment of athletes' movements and provides a more efficient and accurate solution for athlete performance prediction. It proposes a new prediction model based on the Two-Stream Graph Convolutional Network (TS-GCN) combined with the Spatial Attention Module (SAM). First, the OpenPose technology is used to extract human joint information from video data to generate skeletal feature sequences. These feature sequences are then input into the TS-GCN model, where the spatial stream network and the temporal stream network extract the static and temporal features of the movements, respectively. The spatial stream network focuses on the static information in video frames, while the temporal stream network captures the dynamic changes and motion trajectories of the movements. To enhance the model's ability to focus on key features, the SAM is further introduced. It dynamically generates attention maps through global average pooling and max pooling, highlighting the features of key joints while suppressing noise interference. The model's performance is verified on the public Kinetics-700 dataset, where the TS-GCN-SAM model achieves Top-1 and Top-5 classification accuracies of 33.6% and 56.1%, respectively, significantly outperforming traditional methods. Additionally, the TS-GCN-SAM model's performance under different learning rates and training epochs shows that a lower learning rate (0.001) can converge faster and reach a higher accuracy (88.2%), further verifying the model's stability and efficiency. In summary, the proposed TS-GCN-SAM model improves the accuracy of movement recognition and provides coaches with real-time training feedback tools based on athlete performance predictions, helping them quickly adjust training plans and optimize training content.