Research On Abnormal Behavior Sensing Technology Of Surveillance Video Based On CRNN

With the rapid growth in the deployment of surveillance camera devices in urban environments in recent years and the proposed construction needs of a series of projects such as smart cities,human abnormal behavior sensing technology based on surveillance video has become a key direction of current research with high research value and broad application landing needs.However,due to the motion characteristics of a single human body or crowd and the complex and variable nature of the external environment,it makes the recognition and perception of human abnormal behavior with some challenges.At the current stage,the abnormal behavior sensing technology based on surveillance video is still stuck in identifying simple or single abnormal behaviors in brief videos,which cannot quickly and effectively process large and complex surveillance video data,and therefore cannot be well applied to production and work.As a classical network model,the main structure of CRNN model is a combination of Convolutional Neural Network and Recurrent Neural Network.In this paper,the traditional CRNN model is improved and innovated for current application needs,and multi-branch convolutional fusion neural network is proposed and used for the first time for fast and accurate perception of abnormal behaviors in surveillance video,and a system for fast perception and localization of abnormal behaviors in surveillance video and a statistical heat map program for abnormal events are developed.The experimental results show that the system and the program developed in this paper can effectively help the relevant people to trace and prevent the abnormal events.The main research contents and conclusions of this paper are as follows:1.In this paper,a temporal adaptive preprocessing method is proposed according to the problem of large span of different video data duration in UCF-Crime dataset,which automatically selects the appropriate frame extraction frequency according to the total number of video frames to reduce the redundant information and extracts the key frames containing effective information in the video.The experimental results show that the proposed preprocessing method can improve the accuracy of the algorithm for this dataset and accelerate the training speed of the network.2.The spatial and the temporal feature extraction parts of the multi-branch convolutional fusion neural network are compared in several sets of experiments under the same conditions to determine the specific structure and parameters of the overall network.Follow-up experiments show that the proposed fusion network can reduce the number of parameters and computational effort by nearly 50% while the accuracy is better than previous methods,which lays the foundation for the development of related applications.3.After the progress of the network model study,this paper extends the network structure in different directions according to the requirements of different levels,by changing the number of layers,branches,and input dimensions to improve the accuracy of the network or reduce the number of parameters and computation,which provides options for the network model for subsequent applications and programs with different device bases.4.Based on the above network model study,this paper deploys the network model into the application and develops a surveillance video abnormal behavior sensing system and abnormal event statistical heat map program with an easy-to-use interface,which is intended to improve the efficiency of abnormal event tracing and prevention for relevant law enforcement officers.The experimental results show that the system and program developed in this paper can quickly and accurately sense abnormal behaviors in surveillance video for a long period of time.