Video Anomaly Detection Technology Based On Instance Domain Correlation

With the deployment of video collection devices such as surveillance cameras and the development of video storage devices,the number of video content that can be retained and detected has grown explosively.The video anomaly detection task of automatically analyzing massive videos and obtaining abnormal video frames from them has attracted more and more attention from researchers.Due to the difficulty in obtaining abnormal samples and the high cost of frame level labeling for videos,the video anomaly detection field mainly uses two modeling methods: unsupervised modeling using only normal sam-ples and weakly supervised modeling using only video level labeling.However,no matter which modeling method is used,most of the work for video frame anomaly determina-tion only depends on the pixel anomaly within the video frame or the optical flow feature anomaly of objects within the frame.However,such exception definition only focuses on the specific semantics of video frames and ignores the correlation information between frames.This paper takes video frames as examples in multi instance learning methods,proposes a method to determine the correlation exception based on the similarity of video frame instance global correlation features and instance neighborhood correlation features,and designs a deep learning model based on weak supervision and unsupervised modeling methods,Excellent results have been achieved in many open data sets.The main research contents of this paper are as follows:1.To solve the problem that the definition of anomaly depends on pixel anomaly and optical flow anomaly and ignores the correlation information between frames,this paper uses the multi instance learning training mode,extracts the whole domain correlation features and neighborhood correlation features of the instance through attention model and Gaussian fitting,and uses the correlation information between frames by defining the anomaly through the similarity of the two domain correla-tion features.And the neighborhood suppression module is designed to realize the self supervised iterative learning of global correlation features and neighborhood correlation features.2.To solve the problem of low utilization rate of abnormal samples in the weakly su-pervised modeling mode,based on the definition of correlation anomaly,this paper designs a Top-k preselection strategy for the norm features of the global correlation features of normal frame instances and abnormal frame instances to optimize the self-monitoring training process,and constructs a hinge constraint mechanism to constrain the global correlation features of instances more finely.3.To solve the problem of missing abnormal sample supervision signal in unsuper-vised modeling,this paper designs unsupervised pseudo anomaly construction method based on the definition of case domain correlation anomaly,and constructs an opti-mization model of pseudo anomaly simulation abnormal supervision signal through normal samples.And design a pseudo anomaly feature conversion model based on comparative learning to balance the interference of normal video frames in the pseudo anomaly video,and design support vector instance strategies for global cor-relation features to extract support vector instances of positive anomaly sample boundaries.Finally,excellent results have been achieved in several public video anomaly detection data sets.