Wushu 3D Pose Estimation and Transformer-GRU Fusion Recognition Method

Abstract

In order to address the limitations of traditional Wushu 3D posture recognition, such as low efficiency and accuracy, the study extracts key joint points of the human body using the Mediapipe technique. The Transformer is utilized to capture dynamic correlations and long-term dependencies in martial arts action sequences. The Bidirectional Gated Recurrent Unit (BiGRU) establishes global spatiotemporal dependencies, while the Graph Convolutional Network (GCN) constructs the spatial structure of martial arts postures. On this basis, a three-dimensional pose recognition model for Wushu is constructed by integrating Depth-Separable Convolution (DSC). The results show that the proposed algorithm achieves optimal adaptation in the shortest time for Wushu 3D posture recognition, with recognition accuracy reaching 97.01% and a loss value of 0.04. The proposed model has an average error of only 1.99 mm per joint in Wushu 3D posture recognition, and the key point detection accuracy reaches 94.32%, both of which are better than those of the comparison models. Meanwhile, the recognition accuracies for the two martial arts scenarios, Taijiquan and Wing Chun, are 95.96% and 95.87%, respectively. Taken together, the model shows superior accuracy in recognizing 3D martial arts postures and demonstrates good operational stability, providing technical support for related fields.