| The monitoring system deployed in large shopping malls plays an huge role in ensuring the normal operation of the shopping malls and the personal safety of consumers.However,the monitoring personnel are prone to visual fatigue after long hours of staring at the monitoring screen,which may lead to the omission of abnormal behavior of pedestrians and seriously threaten the public safety of shopping malls.Although video analysis technology is gradually being promoted and applied in surveillance systems,it has not yet achieved ideal results in the handling of abnormal behavior in shopping malls.Therefore,this paper takes a large shopping mall as the background,takes double fighting,vandalism and falling behavior as the target to be measured,and designs a human abnormal behavior detection for surveillance video through the study of human joint points detection algorithm and human abnormal behavior recognition model,which cooperates with monitoring personnel to discover abnormal situations in a timely manner and has important public safety research significance and certain commercial application value.In order to obtain the behavioral joint points information in the video,the detection principle of the OpenPose human pose estimation algorithm is analyzed and the experimental comparison is made with the detection performance of Alpha Pose.For the consideration of algorithm processing speed,it is determined that this paper is based on OpenPose for the research of the joint points detection algorithm;To further improve the operating speed of OpenPose,the backbone network in the feature extraction stage of OpenPose is replaced with a Mobile Nets network based on deep separable convolution,the redundant convolutional structures in the feature refinement stage are merged and the six feature refinement stages are streamlined into three stages to obtain a model that can balance detection accuracy and speed.The experimental results show that the improved OpenPose model,with little decrease in detection accuracy,significantly reduces the computation of the model and significantly increases the running speed,laying a good foundation for rapid detection of human abnormal behavior.In order to realize the recognition of abnormal human behavior in video,this paper introduces the idea of graph theory and starts the research based on Spatial Temporal Graph Convolutional Networks.Due to the single way of feature update of nodes by graph convolution in the original network,it leads to the poor performance of the model for classification of behaviors.In order to improve the accuracy of recognition and increase the feature aggregation ability of the model for key nodes,the idea of Graph Attention Network is adopted,calculate the attention coefficients among nodes through the graph attention mechanism,highlight the nodes that play a key role in behavior discrimination,and make the model focus on their features adaptively;a human abnormal behavior recognition model is built,and experiments are conducted on the self-established skeletal behavior dataset to demonstrate the human abnormal behavior recognition model with the fused graph attention mechanism has higher recognition performance than the Spatial Temporal Graph Convolutional Networks,and also performs optimally in comparison with several other behavior recognition models based on skeletal nodes,proving the effectiveness of the model with the fused graph attention mechanism for the abnormal behavior recognition task.By using Visual Studio Code,Php Study and other platforms,the various functional modules of the human abnormal behavior detection software were designed and implemented;meanwhile,the human abnormal behavior detection method designed in this paper was tested on actual data,and the results proved that its accuracy and processing speed can meet the requirements for surveillance video detection,and can be used as an auxiliary monitoring means in the monitoring system of shopping malls. |
