YOLOv7-Pose Based Human Pose Recognition Framework: Applied in the Field of Competitive Aerobics

In this study, we focus on human pose estimation in competitive aerobics—a sport that combines strength, speed, and flexibility. Athletes must perform continuous, complex movements such as squats, lunges, burpees, jumping jacks, push-ups, mountain climbers, and knee lifts within a short time frame while maintaining strict form and stability. Accurately capturing and recognizing these movements with computer vision is therefore crucial for training evaluation, judging assistance, and performance analysis. To this end, we propose an efficient pose-estimation framework based on YOLOv7-pose. Leveraging single-stage detection integrated with keypoint regression, YOLOv7-pose simultaneously performs object detection and pose estimation, demonstrating robustness against the rapid and variable motions typical of competitive aerobics. To evaluate the framework, we constructed the Aerobics-Pose dataset, which comprises 5000 multi-view, manually annotated frames covering 26 representative competitive-aerobics actions. Experimental results show that our method achieves mAP@50 of 90.19% and mAP@50-95 of 69.23% on this dataset, with an average per-frame inference time of 17.74 ms, outperforming existing approaches in both accuracy and speed. These findings confirm the practical value of YOLOv7-pose for competitive aerobics, offering reliable technical support for motion-quality monitoring, injury prevention, and intelligent judging systems during training and competition.