Human Action Recognition Based On Deep Learning

Mechanical manufacturing is the foundation of national economy.With the increase of personalized demand,the production mode has changed accordingly.The demand for small batch and multi variety manufacturing increases,which requires higher flexibility and efficiency of mechanical manufacturing.The development of artificial intelligence technology provides a way of thinking and effective means for intelligent mechanical manufacturing.In the process of human-computer collaborative assembly,the flexibility and efficiency of product production are greatly improved through human-computer interaction.The premise of human-computer interaction collaboration is to recognize human behavior and judge behavior intention by intelligent robot.Based on video sequence,the research on human motion recognition is realized by constructing human body detection model,human posture estimation model and human motion recognition model.Finally,the performance of the model is tested by experiments.A human body detection model based on YOLO Nano network was established to locate the position of human body in the picture.Based on the YOLO Nano multi-objective detection model,the human body detection model is established by modifying the number of output channels of the network,while retaining the category of people.Compared with the original multi-objective detection model,the parameters and calculation amount of the network are reduced.On this basis,the anchor boxes of different sizes and quantities are preset by clustering method,which improves the performance of the model and improves the accuracy of human detection.The human posture estimation model based on SNHRNet network is established,and the position of human bone points is estimated based on the obtained human position.Firstly,based on HRNet human posture estimation network,NHRNet network is established by cutting the low resolution high-level features of the original model.Compared with the original model,the parameter quantity is reduced,but the attention mechanism is missing;On this basis,the SNHRNet human posture estimation model is constructed by integrating se module of attention mechanism to make up for the lack of attention mechanism for NHRNet network.The experimental results show that the accuracy of the constructed SNHRNet human posture estimation model is improved compared with the original model,and the parameter quantity is decreased,and the position of human bone point can be accurately estimated.The ST-GCN human body movement recognition model based on the new partition strategy is established,and the human body movement recognition is realized by learning and modeling the human bone point information.Firstly,based on ST-GCN network model,spatial graph convolution is constructed by sampling function and weight function;Secondly,a new partition strategy is proposed.By dividing the neighborhood set of root node into multiple subsets,more bone points can be associated and different weights can be given to different bone points.The process of constructing weight function can be simplified by using the new partition strategy;Finally,the ST-GCN human body movement recognition model is constructed under the new partition strategy.The experimental results show that the recognition rate of ST-GCN human body movement recognition model is improved compared with the original model,and the recognition rate of human motion based on 3D bone points is much higher than that of 2D bone points.Human action recognition experiments in real scenes are carried out.Based on the construction of human detection model,human posture estimation model and human action recognition model,six kinds of action videos were collected in indoor and outdoor scenes by using MV-EM130C camera and mobile phone respectively.The experimental results show that the human body detection model can accurately determine the position of the human body in the indoor scene with simple background and the high-resolution video captured by the camera.In the outdoor scene with complex background and the low-resolution video captured by the mobile phone,it is prone to false detection and inaccurate detection;The human pose estimation model can effectively locate the position of bone points in the case of human scale change and bone point occlusion,with high accuracy and good robustness;The human action recognition model based on two-dimensional bone points has high recognition rate for obvious and different actions such as push ups and jogging,and low recognition rate for actions with similar local movement and action changes such as clapping and shaking head.