Research On Key Technologies Of Behavior Recognition Based On Graph Convolution And Pose Estimation

Human behavior recognition plays a major role in autonomous driving,intelligent monitoring,human-computer interaction,smart home,and sports.It has become a hot research topic in the field of computer vision.But due to the complexity of human clothing and human behavior,there are still huge challenges in feature extraction and designing efficient network models.This thesis conducts research in two aspects of network model structure and feature description.The main research contents are as follows:(1)Aiming at the low performance of the network model in the behavior recognition problem,an improved spatial-temporal graph convolutional neural network model(STE-GCN)is proposed.In this thesis,the spatial-temporal graph convolutional neural network is used as the baseline model.First,add the proportion coefficient that measures the depth and width of the network in the recognition effect of the network in the spatial-temporal graph convolutional neural network.And then add a specified coefficient that controls the amount of resources when the model is scaled.And then by setting different specified coefficients to simultaneously change the width and depth of the spatial-temporal graph convolutional neural network.According to the training results,the spatial-temporal graph convolutional neural network is further scaled.And finally the STE-GCN network model is obtained,thereby improving the accuracy of human behavior recognition.(2)Aiming at the problem of insufficient description of human behavior characteristics,a calculation method of human behavior characteristics based on angles and vectors is proposed.From the perspective of human posture,this thesis first divides human behavior into two stages,static posture and dynamic motion,and then uses the angle and vector to describe the characteristics of human behavior in these two stages.The angle between the body's tilted trunk and the ground normal vector and the angle between limbs express the characteristics of the static posture,and the information vector with the time dimension is used to express the characteristics of dynamic motion.Finally,it combines the spatial-temporal graph convolutional neural network model to recognize the human behavior,the recognition accuracy of the spatial-temporal graph convolutional neural network is higher.In the experimental part,this thesis conducts experiments on the UCF101 video sequence data set and the NTU-RGB+D three-dimensional bone point data set.At the same time,the pose estimation method is used to extract the coordinate position information of the human body when processing the UCF101 data set.The experimental results shows that the human body features extracted by the calculation method of human behavior features based on vectors and angles can clearly distinguish human actions,the improved spatial-temporal graph convolutional neural network has a high accuracy rate when performing human behavior recognition,the method proposed in this thesis has good robustness in human behavior recognition.