| With the popularity of cameras in daily life,video data has an explosive growth trend.It is no longer sufficient to rely on manual screening of abnormal behaviors in video surveillance.Therefore,the recognition of abnormal behavior based on deep learning has received people’s attention.In recent years,video-based abnormal behavior recognition methods have made many achievements,but such algorithms often use the entire video region to learn features.The in-frame information contains dynamic environment and complex background,which increases the processing cost of the detection system,and will learn many redundant features.Based on the excellent modeling ability of graph neural network on unstructured data,this thesis studies the recognition of abnormal behavior based on skeleton data,proposes two improvement schemes for the existing methods based on graph convolution network,and designs and implements an indoor human abnormal behavior recognition system based on this.The main work is as follows:(1)A Spatial temporal attention graph convolution network is proposed for human abnormal behavior recognition.In view of the weak effect of attention module in existing models and the insufficient feature extraction in time domain and spatial domain modeling,an embeddable spatial temporal attention module(STA)is used to extract spatial and temporal features at the same time,and a multi-branch time-domain convolution module is used to obtain features of different scales,so as to enhance the network’s ability to recognize abnormal behaviors;At the same time,the time convolution and space convolution modules of STGCN are improved to increase the flexibility of graph convolution network.The experimental results show that the accuracy of NTU RGB+D dataset X-SUB and X-VIEW is improved by 5.2% and 4.5% respectively compared with the baseline method.(2)A graph convolution based on dynamic topology is proposed for human abnormal behavior recognition.To address the issue of traditional GCN based methods using fixed graph topologies that make it difficult for the model to extract connections between distant joints,dynamic topology based graph convolution is used to dynamically learn different topologies and effectively aggregate joint features from different channels for abnormal behavior recognition.On the basis of the channel topology refinement graph convolution network,this thesis first improves the feature transformation method,enhances the feature fusion,and then improves the dynamic graph topology.The model constructs more behavioral features with the representation power of skeleton collaborative relationship.The dynamic topology graph realizes more effective skeleton dependency modeling,thus achieving better recognition performance.The experimental results show that the accuracy of NTU RGB+D dataset X-SUB and X-VIEW is 92.7% and 97.2% respectively.(3)An indoor abnormal behavior recognition system is designed and implemented.Based on the research of human abnormal behavior recognition model,the DT-GCN model proposed above is encapsulated,and a human abnormal behavior recognition system is designed and implemented.The system obtains video data by receiving user upload video or calling system camera,and converts video data into skeleton data through HPE algorithm Open Pose,and uses encapsulated DT-GCN for abnormal behavior recognition,Interact with users through the UI built by Py Qt5,use My SQL database to store relevant information,and alert users through SMS,email and other methods if abnormal behavior is identified. |
