Lightweight pose detection for marathon sports based on improved Center Net algorithm

Abstract The human pose detection in marathon sports faces challenges such as large motion amplitude, multiple limb occlusions, and limited computing resources. Traditional detection models are prone to accuracy degradation and response delay in practical applications. Therefore, the study introduces an attention mechanism module based on multi-scale channel weighting and deformable convolution to enhance feature expression ability. The detection head is designed to be lightweight through channel by channel convolution mechanism and Squeeze channel compression mechanism. Finally, a lightweight Center Net pose detection model that integrates multi-module optimization is proposed. The proposed model achieved F1 values of 92.83% and 94.37% on the Human 3.6 M Human Motion Capture Dataset-Running Subset and the AI Challenger Human Keypoint Detection Dataset-Running Category, respectively, with an average response time of less than 0.65 s, significantly better than that of the other three advanced models. The joint prediction error under different running poses was less than 4.5°, and the missing rate of key points in multi-light and camera shake scenes was controlled within 5%, with a frame rate of up to 35FPS. The model performs well in accuracy, robustness, and real-time performance, which is suitable for human pose recognition and analysis tasks in intelligent terminals and sports scenes.