Human action recognition network containing hands based on NPoseC3D59

Résumé fourni par la source

With the intelligent development of machinery manufacturing, the importance of human–computer interaction has become more prominent, and recognizing human actions is a prerequisite for realizing the intelligence of human–computer interaction. Human limbs and subtle hand movements contain rich communication information, and combining the two in research can more accurately capture human behavior and understand human intentions. PoseC3D-based human action recognition models are favored for their excellent performance and low parameter count and can be friendly to add the key point model of the hand on top of only the body torso. Therefore, in this study, the NPoseC3D human action recognition model is constructed, and a human action recognition dataset containing the hand is built for effect validation. Firstly, to improve the recognition accuracy, the ReLU activation function layer, the BN layer, and the downsampling method in the original SlowOnly backbone are adjusted to retain more and more active information, and trained on the NTU RGB+D 60 publicly available dataset for validation, and the recognition accuracy is improved. Secondly, because there is no publicly available dataset containing hands, Whole59, an action recognition dataset containing hand gestures in industrial scenarios, is labeled, and 42 hand keypoint inputs are added to the NPoseC3D model, and trained and tested on the Whole59 dataset, to obtain a human action recognition NPoseC3D59 model containing hand, and the experimental results prove the effectiveness of the proposed method in this research.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.