Identifying Student Body Expression Characteristics in Interactive Music Instruction Using HRNet

This paper addresses the recognition of body expression in interactive music instruction under instrument occlusion, rapid motion, and subtle hand-movement conditions. An improved HRNet-based recognition method is proposed to enhance fine-grained posture estimation and multimodal behavior-feedback analysis. A multimodal dataset was constructed by synchronously collecting video, touch and pressure signals, and sound-field data. In the model, standard convolutions in HRNet are replaced by involution operators to improve spatial feature representation while reducing computational complexity. A multi-scale feature fusion mechanism integrates macroscopic body movements with microscopic hand-motion information, and a Transformer-based hand module further captures detailed features such as knuckle positions and pressure points. Experiments on a self-built dataset show an overall keypoint-recognition accuracy of 92.3%, representing an 8.5% improvement over the original HRNet. Computational complexity is reduced by 44%, and inference speed increases by 69.8%. Ablation analysis verifies the effectiveness of the model in body pose estimation, fine hand-motion recognition, and cross-modal fusion, providing an engineering solution for real-time visual sensing in interactive teaching environments.