Real-time Recognition Of Abnormal Behaviors For Intelligent Monitoring

In recent years,video surveillance has played a very important role in urban safety supervision and security.Surveillance video often records some abnormal behaviors.These abnormal behaviors will bring great harm to society.Traditional video surveillance uses staff to monitor the monitoring screen,which not only wastes a lot of money and labor,but also misses some abnormal events due to the fatigue of the monitors,and brings great harm to society.Therefore,it is necessary to study the real-time recognition of abnormal behavior based on intelligent monitoring,which can improve the safety factor of public areas,promote the harmonious development of society,and help to build a country ruled by law.In response to actual problems,this article proposes to divide abnormal behaviors into static abnormal behaviors and dynamic abnormal behaviors.Static abnormal behavior can be judged by a single frame image,for example,site workers do not wear safety helmets.Some intermediate states of dynamic abnormal behavior are not different from normal behavior.The accuracy of judging the category of dynamic behavior through a single frame image is low,so it is necessary to analyze the video sequence,such as fighting,climbing,etc.Aiming at the above-mentioned two behaviors with different attributes,this paper respectively proposes an abnormal behavior recognition algorithm based on static characteristics and an abnormal behavior location recognition algorithm based on spatio-temporal characteristics.The contents of this paper are as follows:1.For behaviors that can be described by static features,for example,the behavior of not wearing a mask during the epidemic and workers not wearing helmets on site,Chapter 3studies the static feature abnormal behavior recognition algorithm based on image target detection,using the YOLOv3 target detection algorithm to detect different abnormal behavior states,and in view of the problem that the YOLOv3 misses more small targets in the monitoring screen,two improvements are proposed: First,the target variable proportion data enhancement algorithm is used in the training phase to generate more small targets,so that the neural network can learn more features of small targets,to improve the detection ability of small targets;second,using multi-scale pooling layer,the features are stitched after the largest pooling cores of different sizes.This method can improve the expression ability of features,cope with the difference of targets brought by different positions at different moments in the video sequence,make the algorithm more sensitive to small targets.When the two methods are used at the same time,the robustness of the algorithm is better.Due to the continuity of this type of abnormal state behavior,the real-time performance of the algorithm can be guaranteed through uniform sampling.2.For dynamic behaviors that need to be described by spatio-temporal features,such as calling,smoking,climbing in certain situations and fighting,a 3D convolutional neural network based on spatio-temporal features is required.The existing behavior recognition algorithm based on 3D convolutional neural network can only realize the recognition of behavior category,but cannot realize the position of the behavior subject.This article draws on the idea of target detection,studies the abnormal behavior recognition algorithm based on temporal and spatial characteristics,and proposes the algorithm of abnormal behavior location recognition algorithm based on 3D convolutional neural network,the spatio-temporal features are extracted through the three-dimensional convolution network,and finally the abnormal behavior is recognized on the feature map by the detector.The algorithm can identify and locate the category and location of the behavior,and can identify and locate the behavior in the same time domain.However,considering a single neural network often does not perform well when processing multiple tasks,this paper proposes an abnormal behavior location recognition algorithm based on the fusion of 2D and 3D features.The algorithm uses 3D channels to extract spatiotemporal features,2D channels to extract human features,and then uses Channel fusion and attention mechanism reshape the features,and finally identify and locate the behavior on the feature map.The above two algorithms can guarantee the real-time performance of the algorithm through uniform sampling,and the accuracy of the latter is better than the former.