Automatic Recognition of Joint Points in Sprinting Sports Based on Video Images with Biological Visual Attention Mechanisms

Along with the rapid development of artificial intelligence and computer vision technology, automatic recognition of athletes’ joints based on video images has become an important field of research in sports science and intelligent training. The fast dynamic characteristics and background complexity in sprinting require the recognition system to reach a higher level of precision and stability. This paper proposes a deep neural network model that integrates the biological visual attention mechanism with optical flow modeling, and the overall structure involves three modules: video preprocessing, attention-guided recognition, and joint post-processing. In the process of feature extraction, the model introduces spatial and channel attention mechanisms, and also combines with optical flow network to model inter-frame dynamic information, which greatly enhances the efficiency of capturing key regions in high-speed motion. The experimental results show that the model achieves 92.3% accuracy in the self-constructed sprinting dataset for the COCO—Pose 2017 dataset Achieved a PCK@0.2 accuracy of 89.7%, and its inter-frame stability FSI value is 92.3%, and its FSI value is 92%. Its inter-frame stability FSI value is 0.85, which is significantly better than a group of mainstream methods such as OpenPose, YOLO—Pose, etc. The image recognition method that integrates the attention mechanism and optical flow perception has stronger robustness and structural consistency in dynamic scenes, which paves an effective technological path for the intelligent sprinting posture analysis and training system.