Emotion recognition in dance therapy driven by DanceEmoNet: a deep learning model based on facial expression and pose estimation.

With the widespread application of dance therapy in mental health interventions, precise, real-time identification of emotional fluctuations has become a significant challenge. Existing emotion recognition methods predominantly rely on unimodal information, making it difficult to fully capture the complexity and temporal dependencies of emotional changes. To address this, the DanceEmoNet model is proposed, integrating facial expression recognition and pose estimation techniques to improve the precision of emotion recognition through multimodal feature fusion. The model employs YOLOv11 for face and pose detection, utilizes a hybrid TriBAN network and CNN-LSTM model for feature extraction and temporal modeling, and performs feature fusion via the GCNC module. Experimental comparisons with several benchmark models demonstrate that DanceEmoNet achieves better results across multiple metrics, exhibiting faster inference speed (e.g., reduced per-frame latency) and lower computational cost (e.g., fewer FLOPs), with overall performance gains ranging from 5 to 10%. The experimental results confirm the clear strengths of DanceEmoNet in capturing complex emotional changes and dynamic dance movements, indicating its practical applicability for real-world deployment.