Research On Monitoring Video Abnormal Event Detection Model

With the strengthening of the public safety awareness and the development of science and technology,the intelligent video surveillance system will become the mainstream of video surveillance system development.The monitoring video abnormal event detection technology as a key part of the intelligent video surveillance system,it can independently perform intelligent analysis of various motion events in the monitoring video,once the occurrence of abnormal event is detected,it will send alarm to the management in real time,with the advantages of intelligent,real-time and so on,monitoring video detection technology has high academic research value and commercial application prospects in the field of public safety.This article study the monitoring video abnormal event detection technology from two aspects of motion feature extraction and model building.The main tasks of this article are as follows:1.In the monitoring video,the same moving object in different image areas,the optical flow value of the pixel point will change according to the position of the pixel point,based on this problem,a monitoring video abnormal event detection method based on multiple instances and time series was studied.In this method,for the input monitoring video stream firstly divided into multiple instances to form different instance streams,and the motion features are extracted based on the optical flow values of the pixel points in the spatio-temporal blocks from different instance streams;in the modeling phase,the motion features of internal continuous frames in the spatiotemporal blocks are first considered as a time series,and time series algorithm is used for modeling and prediction,then all prediction intervals are merged,finally abnormal events is judged by the relationship between the actual feature value of the image block and the prediction intervals.The experiments show that this method has a good detection effect on the local and global abnormal events in the monitoring video.2.The applicability of the artificial feature under different monitoring video scenes is different,when video scene is changed,feature selection experiments are needed to select the most suitable feature,based on this problem,a nonparametric abnormal event detection method based on deep learning was studied.In this method,for the input monitoring video stream,firstly,the histogram bimodal method is used to count the optical flow value of each pixel in the video stream to obtain a partitioning threshold that can distinguish between an active pixel point and a background pixel point,and then the active area of the spatio-temporal blocks is screened through the threshold to remove unnecessary background information,finally use the deep learning network to automatically extract corresponding depth semantic features from the filtered blocks;in the modeling phase,first,a fixed-size dynamic dictionary set is maintained by vector merging,then clustering the dictionary set and establishing the corresponding codebook and the judgment threshold by clustering center,finally abnormal events is judged by the similarity between the motion features and codebooks.The experiments show that this method has a good detection effect for crowded and sparse monitoring video scenes.