In this study, Gaussian mixture model preprocessing and histogram of orientation gradient (HOG) and histogram of optical flow direction (HOF) action feature extraction techniques are used to significantly optimize the accuracy of dance performance action recognition. In addition, the VNect pose estimation algorithm and dual-stream 3D-CNN-based skeleton action recognition are used to improve significantly the accuracy and efficiency of recognition. In terms of music feature recognition, this study utilizes short-time time domain and frequency domain analysis techniques and BP neural networks to effectively achieve accurate recognition and sentiment classification of music features. Analyzing a dataset containing 6000 training videos, 800 validation videos, and 1200 test videos, the study reveals that musical elements, melody, rhythm, emotional tone, and song style significantly impact the stage performance effect. The study not only highlights the importance of traditional Chinese music in dance performance to enhance the artistic effect and deepen the emotional expression, but also opens up new perspectives and methods for disseminating and developing traditional culture.