Anomaly Detection And Localization In Video Surveillance By Deep Neural Network

Owing to the popularity of intelligence and people's attention to public safety,people have an urgent need for an intelligent monitoring system with real-time massive video data processing capacity.Video anomaly detection technology has been the research emphasis of image processing and machine vision,and is playing an increasingly important role in social life.Due to the requirement of the video monitoring system ability of vast video monitoring data processing in real-time,different kinds of convolutional neural network are used in this paper to detect the anomaly in crowded scenes.The main research works are as follow:Firstly,the novel two stream spatio-temporal convolution neural networks model is proposed to detect and localize the abnormal behavior in video sequence of crowed scene.Considering the typical position of camera and large number of background information,we introduce a novel spatio-temporal cuboid of interest detection method with varied-size cell structure and optical flow algorithm.Then,the two stream 3D convolution neural networks is used to describe the same behavior in different temporal-lengths.That method not only ensures that the mostly information of the behavior in spatio-temporal interest cuboids could be captured,but also insures the information unrelated to the major behavior in the cuboid could be reduced.Secondly,a novel anomaly detection and location method based on sparse reconstruction cost and spatial-temporal convolutional neural network is proposed to improve the adaptive ability of anomaly data and reduce the limitations of application.The three-dimensional convolution neural network is used to describe interest cuboids extract from foreground extraction temporal and spatial features of.In order to deal the large number of features of spatial-temporal interest cuboids,we apply the AP clustering method in dictionary learning,and the representative feature of the training sample is add to the dictionary,which greatly reduces the dictionary dimension and reduces the memory requirement of sparse representation.In the test stage,considering the similarity among normal samples,we apply sparse reconstruction cost with AP clustering to reduce the computational cost.Finally,in order to detect and locate the anomaly in real time,an end-to-end SSD real-time anomaly detection and location algorithm is proposed.This method completes one-step implementation of anomaly detection by setting default box on six different convolution feature maps to achieves anomaly classification and more accurate and complete boundary box.The algorithm can process nearly 58 video frames per second,which meets the real-time requirements of anomaly detection.