Research on Upper Limb Rehabilitation Action Recognition Method of Unsupervised Contrast Learning Based on Time-domain Multi-scale Feature Fusion

Vision-based action recognition technology plays a major role in human-robot interaction during the upper limb rehabilitation training of stroke patients with the aid of an upper limb rehabilitation robot. To address the problems that supervised learning requires a large amount of manual labeling, which is time-consuming and inefficient, in this paper, we present an upper limb rehabilitation action recognition method of unsupervised contrast learning based on multi-scale feature fusion and we improves SimCLR by replacing the single-scale CNN with Inception V1 which contains the multi-scale feature fusion module and increasing the temporal dimension based on the idea of the expanded convolutional neural network to fully extract the Spatio-temporal feature information of the actions. The experiment results show that unsupervised contrast learning based on time-domain multi-scale feature fusion achieves recognition accuracy of 92.59% for 8 upper limb rehabilitation actions without extensive manual labeling. It can help patients recover their motor functions through upper limb rehabilitation training.