YOLOv8-DTW-based pose recognition and training evaluation for free climbing actions

In rock climbing pose recognition and training evaluation tasks, traditional methods often show limited keypoint localization accuracy in complex environments, weak stability in evaluation, and insufficient real-time responsiveness to match the training rhythm. Therefore, based on the YOLOv8s-Pose algorithm, this study addresses key challenges in rock climbing scenarios and integrates three core modules: ① The C2F-GhostNetV2 lightweight feature extraction module, which reduces computational load while enhancing the capture of fine-grained posture features; ② The CARAFE adaptive upsampling technology, which improves the positioning accuracy of small-scale keypoints; ③ The Convolution Block Attention Module, which focuses on the features of the core exertion parts in rock climbing. Combined with the hierarchical optimized fast dynamic time warping algorithm, an integrated model with high-precision posture recognition and low-latency training evaluation is constructed. Experimental results show that, in rock climbing scene tests, the model achieves a keypoint detection accuracy of 98.6% and an area under the receiver operating characteristic curve of 0.984, which significantly outperforms mainstream models. In complex scenarios, the decline in action standard recognition accuracy is only 6.8%, and the model shows much higher stability than traditional models. The average real-time response time is only 25.4 ms, and 52.3% of the responses fall within the range of 0–20 ms, meeting the fast-paced demands of rock climbing training. In addition, the model achieves a pose estimation F1 score of up to 0.98, and the pixel error of core keypoints is reduced to 3.1, showing clear advantages in fine-grained localization accuracy. Overall, the proposed model achieves full-process coverage from rock climbing pose keypoint detection and recognition to training evaluation tasks. It provides reliable technical support for intelligent rock climbing training guidance and offers a new framework for pose recognition and training evaluation tasks in sports scenarios.