Real-Time Aerobics Pose Estimation and Motion Trajectory Optimization Using Enhanced YOLOv7 with CA Attention and ASPP

Abstract

Aiming at the high dynamic characteristics of aerobics, this study proposes a real-time pose capture and motion trajectory optimization method based on the YOLOv7-Pose algorithm. The method is evaluated on the CAF-3D and FitMotion-VIS datasets. By improving the keypoint detection head of YOLOv7, combined with the CA attention mechanism and the atrous spatial pyramid pooling (ASPP) structure, the accuracy of human keypoint detection is significantly improved (the mAP of the verification set reaches 95.7%, outperforming OpenPose and AlphaPose). At the same time, the dynamic time warping (DTW) algorithm and a multi-objective trajectory optimization strategy are introduced to solve trajectory matching issues caused by varying action speeds, and TensorRT is used for accelerated deployment to achieve real-time performance of 84 FPS. Experiments show that the system maintains high robustness under complex illumination, multi-person occlusion, and dynamic motion, with the position error of keypoints reduced to less than 3%. These results provide reliable technical support for applications in sports training, rehabilitation evaluation, and other real-world scenarios.