Fusion and improvement of stacked hourglass network and SlowFast network for rock climbing action key point detection

Human action key point detection is a core technology for motion action analysis and behavior recognition. In complex sports scenes such as rock climbing, its detection accuracy and robustness face severe challenges due to problems such as limb occlusion and dynamic action changes. To this end, the study proposes a rock climbing action key point detection model that integrates the improved stacked hourglass network and SlowFast network to improve detection performance in complex scenes. Based on MPII, COCO, and self-built rock climbing data sets, testing was conducted in an NVIDIA RTX 4090 hardware environment and 200 iterations of experimental settings. The results showed that the Percentage of Correct Keypoints value of the knee keypoints of the proposed model reached 87%, which was significantly higher than that of RLE CCV2021 (79%) and Mobilenetv2 CVPR2018 (80%). In the rock climbing scene of the COCO data set, the [email protected] value of each key node was better than that of SCNet_101 and ResNet-50. In practical applications, the model performed well on climbing specific datasets, with a confidence level of 0.77–0.92 for key action recognition, and could effectively cope with dynamic action and complex background interference. This study has verified the effectiveness of the proposed model, which can provide accurate data support for scientific training, action specification guidance and sports injury prevention in rock climbing. It contributes to the intelligent and scientific development of rock climbing and has important practical application value and promotion significance.