TC-YOLO: An Improved Tai Chi Movement Pose Estimation Algorithm Based on YOLOv8

At present, the dissemination of Tai Chi is predominantly facilitated through offline instructional methods combined with video practice, lacking effective feedback on movement poses and exhibiting relatively low efficiency. This article proposes a novel algorithm named TC-YOLO for pose estimation based on YOLOv8. TC-YOLO enhances the efficiency through utilizing real-time detection of key points in Tai Chi practitioners’ movements to provide more accurate and intuitive feedback for instructional evaluation and pose correction. Focusing on the demonstration of the Essential Eighteen Movements of Chen-style Tai Chi as the research subject, Tai Chi movement dataset comprising 3,688 images is constructed. To enhance model efficiency, the backbone network is reparameterized through the Reparametrized C2f (RC2f) module, which optimizes feature extraction process by allowing more effective information flow and reducing computational complexity. Furthermore, a simplified neck network structure is designed to provide effective information transmission and multiscale feature fusion, thereby enhancing detection accuracy. Experimental results show that the TC-YOLO algorithm achieves a mAP of 97.2% and a recognition speed of 146.1 FPS with lower parameters and computational cost on the self-built dataset, which is better than YOLO-Pose and other models. TC-YOLO could make an instructive contribution to the field of sports science through the analysis of Tai Chi movement poses, promoting the inheritance and global dissemination of Tai Chi.