Deep Reinforcement Learning-based combat recognition of traditional Chinese Sanda under artificial intelligence technology

In the modern sports ecosystem, the digital preservation of traditional Chinese Sanda requires precise methodology to capture its complex technical lineage. This study proposes an innovative combat recognition framework for traditional Chinese Sanda based on Deep Reinforcement Learning. The framework employs a two-stream three-dimensional Convolutional Neural Network for spatiotemporal feature extraction, integrates a Convolutional Neural Network to process spatial information, and utilizes the optical flow method to capture temporal dynamic features. At the decision optimization level, the Proximal Policy Optimization algorithm enables intelligent decision-making. A multi-objective reward function is formulated to comprehensively optimize technical execution accuracy, tactical sequence consistency, and quantified martial arts movement characteristics. The experimental results show that the proposed method achieves a Sanda action recognition accuracy of 89.7% ± 1.3% on the self-constructed dataset. The highest accuracy obtained through five-fold cross-validation reaches 92.3%, representing an improvement of 15.6% points over the Support Vector Machine method. In the action category recognition with tactical consistency evaluation task, an average F1-score of 83.6% is obtained. This study provides a novel technical pathway for the digital preservation and intelligent training of traditional martial arts. It also extends research directions in sports action analysis and artificial intelligence applications in cultural heritage preservation, demonstrating substantial theoretical value and practical significance.