Research On Human Pose Action Recognition Oriented To Human-Machine Collaboration System

With the development of information science and technology,human-machine collaboration technology has become a research and application hotspot.The new generation of human-machine collaboration system can give full play to the expertise of humans and robots,which greatly promotes the improvement of production efficiency.The key to realize the human-machine collaboration system is to let the robot understand the human behavior.Therefore,this paper focuses on the human pose action recognition technology in the human-machine collaboration system,and proposes a pose estimation method and an action recognition method respectively.The specific work content is as follows:(1)A set of multi-view human special operation behavior database was established.Since there is currently no special action dataset for the factory environment,a set of multi-view and multi-modal camera sensor platform is built with reference to the composition structure of the mainstream general action recognition dataset.On this basis,we designed common human actions in the background of the manufacturing industry,and created a special database in the factory environment to provide corresponding support for subsequent behavior recognition research in specific scenarios.(2)A human pose estimation method based on attention multi-resolution network is proposed.The network improves the multi-resolution feature fusion approach,introduces the attention mechanism to evaluate the importance of different branch channel information,and increases the weight of the more important semantic representation channels,so that the network can reasonably fuse high resolution and low resolution features to obtain higher spatial positioning accuracy.In addition,the multi-context attention residual unit is used in the initial stage to increase the receptive field and improve the ability of the network head to learn multi-resolution information.(3)A human action recognition method based on adaptive joint correlation graph convolutional network is proposed.The network improves the fixed graph topology of human joint by introducing an adaptive mechanism,avoiding the limitation of traditional manual setting of joint connection methods.In addition,the joint correlation graph convolution module is used to capture the potential dependence between each joint,which enhances the network's extraction of long-distance joint features.This paper conducts experiments on the proposed pose estimation method and action recognition method on related datasets,and the experimental results can prove the feasibility of the method.