A Study on Improved YOLOv7-pose and Bytetrack for Dual Person Tracking

This study proposed a joint network that integrates person re-identification (Re-ID) and object detection tasks to address the challenge of identity changes in traditional Chinese medicine massage. The algorithm combined a CBAM attention mechanism with YOLOv7 and incorporated a Re-ID branch into the YOLOv7-pose detection algorithm. By leveraging feature matching in the Bytetrack tracking algorithm, the Re-ID features compensate for missing appearance features and optimize the tracking strategy, reducing identity switches. The experimental results demonstrated that the proposed algorithm achieved significant improvements. It successfully captured all the data in the videos, outperforming the Deepsort and Strongsort algorithms with failure to capture rates of $82.3{{\% }}$ and $72.9{{\% }}$ , respectively. The target ID change rate was reduced to $2.1{{\% }}$ , making data collection less challenging than the original Bytetrack algorithm. The algorithm ensured consistent identification of physiotherapists and patients after interactive behaviors, ensuring data consistency during the data collection.