The Study On Human Abnormal Behavior Recognition Algorithm Based On Deep Learning

With the improvement of people's safety awareness and the need to build safe campus and a safe society,more and more surveillance cameras are used on various occasions,but traditional monitoring equipment can only play the role of video recording and does not have intelligent analysis functions.It is necessary to manually view the video content in real-time to capture abnormal events and emergencies.The rise of artificial intelligence technology has made monitoring equipment intelligent,but the current behavior recognition algorithm has a huge amount of calculation and needs to run on the server.Therefore,this article focuses on how to reduce the number of network calculations while improving network accuracy,so that the network can efficiently and accurately identify abnormal targets in videos on small processors such as ordinary computers and single-board computers.In order to improve the accuracy of network recognition,the sub-sampling algorithm is used to obtain global video information during video preprocessing,and a new recognition network with high parameter utilization called RD3 DNet is designed to explore the structure and training parameters of the network.The influence of accuracy rate obtains a set of optimal structure parameters and training parameters of the network.Then,the feasibility,Top5 analysis,feature map analysis,and advanced network comparison analysis were performed on the test set.The accuracy rates of 95.4% and 73.5% were obtained on the UCF10 and HMDB51 data sets,respectively,compared with R3D(85.8% and 54.9%),Fast3D(89.5% and 55.4%),TSN(94.2% and 69.2%)and other Advanced networks have improved to varying degrees.The comparison shows that the network designed in this paper has certain advantages in the accuracy of recognition of human behavior.To reduce the amount calculation and meet the real-time requirements,this article optimizes the network structure,the decomposition method of the convolution kernel,group convolution and channel pooling,etc.,and conducts detailed experimental exploration to analyze the improvement method for the accuracy of network,the influence of the amount calculation and parameter.The accuracy of the test on the UCF101 and HMDB51 datasets reached 96.31% and 74.12%,which were 0.91 percentage points and 0.62 percentage points higher than before the improvement,and the number of parameters and calculations were significantly reduced to 6.83 M and 19.23 G FLOPs,respectively.Compared with the 28.15 M and 42.42 G FLOPs before optimization,this is a reduction of 75.73% and 54.67%.The tested reasoning speed reached 91.73 VPS,1467.68FPS(not including data loading),which is about twice as fast as advanced network.This article uses the optimized RD3 D algorithm to perform knowledge transfer on the self-built data set,and then transplants the network to simple intelligent monitoring for indoor abnormal behavior testing.The test results are consistent with actual actions,and the rapid and accurate identification of abnormal behaviors is achieved,research purposes.This paper studies the behavior recognition algorithms,solve the problems of low accuracy and a large amount of calculation,promotes the application of intelligent surveillance,has potential practical application value,and promotes the development of computer vision.