A Simulation Study of Athletic Error-Motion Recognition Based on Visual Image Technology

Abstract

This study addresses common error motions in track and field by integrating visual image processing with deep learning techniques to develop a simulation framework for image-based feature extraction and classification. We first describe the characteristics and challenges of recognizing erroneous motions, then detail our methods for image preprocessing, keypoint detection, and spatiotemporal feature fusion based on convolutional neural networks. Experiments on a self-built track-field video dataset compare several mainstream models in terms of recognition accuracy and real-time performance. Results show that our method achieves an average recognition accuracy of 92.4% on typical error motions (e.g., hurdle-leg deviation, false start), with a response latency under 50 ms—meeting the requirements for online monitoring and feedback. Finally, performance evaluation on the simulation platform demonstrates the system’s robustness and scalability under varying illumination and occlusion conditions, providing effective technical support for training and competition monitoring. We also discuss future directions for integrating wearable sensors.