Keypoint Detection Method for Single Person Gymnastics Actions Based on Multi-Scale Incremental Learning

Keypoint detection of human body is a hot research area in computer vision. At present there exist some problems for keypoint detection in gymnastics actions, such as insufficient detection accuracy and lack of capability to detect detailed body parts. In order to improve the detection accuracy, this paper proposes a multi-resolution network that has a larger receptive field in the shallow layers and can utilize high-resolution channel to enhance the extraction of detailed features. To achieve the detection of keypoints of hands and feet, an incremental learning network is designed. The network fuses the shallow features of the multi-resolution network and computes deep features using a gymnastics actions self-built dataset, so that the detection ability of keypoints on hands and feet is improved. Finally, the output results of the two sub-networks are concated. Computer simulations demonstrate that the multi-resolution network achieves an accuracy rate of 94.4% on the COCO2017 keypoint detection dataset, and the incremental learning network can accurately detect keypoints of detailed body parts with fewer training data. © 2024 Chinese Institute of Electronics. All rights reserved.