Design And Implementation Of Human Behavior Recognition System Based On Pose Estimation And Graph Convolutional Neural Network Construction

In many applications such as human-computer interaction,video surveillance,and health care,the importance of detecting and analyzing human behavior has become increasingly prominent,and people have paid more and more attention to the research on identifying and modeling human behavior.In order to solve many problems in the RGB image behavior recognition algorithm,such as the influence of external environmental factors,the color change of the image and the influence of noise,this thesis studies the behavior recognition algorithm on the data of the skeleton sequence.With the widespread application of lowcost,high-precision depth cameras and the simpler and simpler methods for precise positioning of human skeletons,skeleton-based behavior recognition has gradually become the focus of research.During the process of motion recognition using the skeleton,neither the ambient light nor the background will interfere with the result.In addition,the simplicity and ease of use of skeleton data can effectively reduce the demand for computing resources.In this thesis,the human skeleton data is used to process it,and it is applied to the improved network model,so as to realize the recognition of human movements.This thesis conducts a series of studies to explore how to obtain more comprehensive feature information,extract its depth features,and how to better utilize the information of skeleton data.The main research contents are as follows:(1)A Global Adaptive Graph Convolutional Network(GA-GCN)based on global relationship is proposed,which can effectively capture spatial features and global dependencies between joints,so as to achieve more accurate recognition effects.In order to better identify features in the domain,the network not only improves on the relational connections of the original physical structure,adding new relational connection edge connections,but also introduces the spatial attention module to more accurately capture discriminative features in the domain.On the Kinetics dataset,the proposed network improves the accuracy of Top-1 by 1.4% and the accuracy of Top-5 by 2.3% compared with the baseline network,which proves the effectiveness and competitive advantage of the algorithm.(2)A Spatio-Temporal Attention Mechanism and Global Adaptive Graph Convolutional Network(STAM-GAGCN)based on spatio-temporal attention mechanism is proposed.It uses the spatiotemporal attention mechanism to effectively capture and extract spatiotemporal features through multi-scale temporal convolution technology,and can effectively search for important features from the two dimensions of time and space,and better realize the behavior recognition of the human skeleton by fusing the temporal and spatial attention modules together.In this thesis,the relevant information of joints and bones is extracted from the original skeleton data,so as to construct a dual-stream network model,which can effectively weighted and fuse the prediction results of the two to more accurately predict and classify the behavior.On the Kinetics dataset,compared with advanced methods,there is a considerable improvement compared with advanced methods,compared with the baseline network,the accuracy of Top-1 is improved by 3%,and the accuracy of Top-5 is improved by 3.4%,which proves the effectiveness and competitive advantage of the algorithm.(3)A skeleton-based human behavior recognition system is realized.First,the overall design of the system is carried out,and then the design of each module is carried out,and finally the functions of the system are realized and displayed.The system converts the video sequence into a skeleton sequence through the camera collection or video file upload,and then completes the whole process of recognition through the encapsulated behavior recognition algorithm.