Research on an Improved RTMPose Model for Evaluating Dance Standardization Scores

Human pose estimation is a critical technology in computer vision that enables machines to understand and interpret human movement. One important application of this technology is the standardized analysis of dance movements. In this paper, we improve the RTMPose model by optimizing keypoint representation through convolutional layers and Large Kernel Gated Attention Units (LGAG), transforming 2D pose estimation into a coordinate classification task. The LGAG selfattention mechanism enhances keypoint detection accuracy. Additionally, the loss function based on SimCC, combined with soft label encoding, further optimizes the model's performance. Comparative experiments conducted on the COCO, MPII, and our collected standard dance datasets demonstrate that our model significantly improves human pose estimation performance and effectively achieves standardized dance movement scoring. This highlights its potential for various applications in sports, health, education, and human-computer interaction.