Research on optimization of human posture estimation and sports training feedback system based on deep learning

As the requirements for accuracy and efficiency in sports training continue to increase, the application of traditional posture estimation algorithms in high-dynamic and complex training environments faces problems such as image redundancy, noise interference, and stability in long-term training. To solve these problems, this study proposes a sports training system based on an optimized human posture estimation algorithm, aiming to improve the accuracy and stability of athletes in training. By introducing optimization strategies such as deep image processing, enhanced feature extraction, focusing mechanism, and feature fusion, the joint estimation accuracy and system stability are improved. Based on tennis training data, the experiment analyzes the performance of different algorithms in joint angle similarity, long-term similarity, and system stability. The results show that the optimization algorithm improves joint angle similarity by about 10%, long-term stability by 12%, and system stability by 8%. These experimental results verify the advantages of the optimization algorithm in sports training, especially in complex movements and multi-person training scenarios, and can provide athletes with high-precision and real-time feedback.