Spatio-Temporal Recognition and Correction of Martial Arts Actions Using a Two-Stream Convolutional Network

Existing models are difficult to effectively capture the spatio-temporal features of martial arts actions, resulting in low recognition accuracy and efficiency. This study improves the Two-Stream Convolutional Network (TSCN) to recognize deviated joint actions and assist trainers in correction, stores data based on Hadoop Distributed File System (HDFS), and processes data in a distributed manner on the Spark platform. The self-attention mechanism is applied in the spatial stream of TSCN to enhance the network’s attention to joints and action details, and the temporal feature extraction module based on optical flow is applied in the motion stream to capture martial arts motion trajectories and improve action recognition accuracy. When the action detection score is low, the OpenPose model is used to generate a sequence of skeleton points, and Fast Dynamic Time Warping (FastDTW) is used to achieve trajectory matching. Finally, correction prompts are generated based on the deviation results to assist trainers in adjusting martial arts actions. The results show that the improved TSCN model has a sensitivity (Sn) of 0.92 and a specificity (Sp) of 0.94 for action recognition; after posture correction, the average elbow error is reduced from 4.68[Formula: see text]cm to 2.65[Formula: see text]cm, and the martial arts performance of beginners increases by 18.57%. The improved model can precisely recognize complex martial arts actions, assist teaching, and improve the standardization of students’ actions.