Dance Action Recognition Technology Based on RTDETR Network and PoseConv3D

Abstract

To address low recognition accuracy in dance action recognition due to high dynamics and complex spatiotemporal features, this study proposes a hybrid model combining enhanced object detection and spatiotemporal modeling. First, a lightweight Re-PConv Block (RPCB) module is designed to improve the feature extraction and efficiency of the RTDETR detector. Second, the enhanced detector is integrated with a PoseConv3D network equipped with a Convolutional Block Attention Module (CBAM) to jointly optimize spatiotemporal representations of dance movements. Experimental results show that the improved detection method increases precision and recall by 12.73% and 15.92% on COCO2017, and by 10.71% and 13.75% on Pascal VOC, compared to other methods. In action recognition, the proposed model achieves gains of 7.45% in Top-1 accuracy and 3.41% in Top-5 accuracy on FineDance, with recall and F1 score rising by 7.95% and 6.07%, and inference speed leading by 31.0 fps. On Kinetics, Top-1 and Top-5 accuracy improve by 6.81% and 4.25%, with recall and F1 score increasing by 7.14% and 7.30%. The innovation lies in using RPCB for lightweight, enhanced detection and CBAM for strengthened spatiotemporal modeling, forming an end-to-end, accurate, and real-time framework for dance action recognition, offering an effective solution for digital and intelligent dance analysis.