Research On Anomaly Detection Technology In Surveillance Video By Sparse Combination Learning

With the continuous improvement of the security awareness and public safety requirements,it has become increasingly common to install surveillance cameras in public places.At the same time,the number of video data has increased exponentially.However,among the data recorded by surveillance cameras,only a small part of frames that recorded abnormal events are concerned by people,and most of the other data can be regarded as redundant data.Therefore,after obtaining a large number of video,how to obtain effective information from it has become an urgent need to be solved.Since the surveillance camera records a large number of consecutive video frames,therefore,manual inspection need take a lot of time to view video by frame,and the manual inspection has the problems of low efficiency and unstable reliability.The longtime work will also cause visual fatigue,and further reduce work efficiency and the detection accuracy.In order to better check video,surveillance video anomaly detection technology arises at the moment.Compared with the manual inspection,the surveillance video anomaly detection technology in this dissertation can automatically find the abnormal events in video,and can ensure the reliability of detection.Sparse coding is a common and effective algorithm in video anomaly detection.The sparse combinatorial learning(SCL)algorithm base on the principle of sparse coding,and uses multiple fixed combinations to solve the problem that sparse coding is too time-consuming to find the appropriate base vector.The SCL can get the accuracy of not less than sparse coding in video anomaly detection.This dissertation studies the abnormal events detection in surveillance video based on the sparse combinatorial learning algorithm and the distribution characteristics of the optical flow in video.For the training video,we provide a partitioning algorithm by analyzing the distribution of the optical flow magnitude in each region.In each partition,we extract the variable scale 3D-HOF and the entropy of optical flow orientation,and learn sparse combination set from the features.For the test video,we reconstruct data by using the corresponding sparse combination in each partition.The anomaly events are determined based on the size of the reconstruction cost,and sparse combinations are updated by the data that are determined as normal.Finally,the surveillance video anomaly detection system is presented.The main work of this dissertation is as follows:Firstly,we provide a video partition algorithm based on similarity of optical flow magnitude distribution probability.First of all,all frames are divided into the same size blocks in training video,and the histograms of the optical flow magnitude are counted in each block.Then we use K-Medoids algorithm to cluster the histograms of the optical flow magnitude,and the video is partitioned by the clustering result.The influence of perspective deformation on the detection results is alleviated by partition algorithm.Secondly,we present the variable scale 3D-HOF feature based on the distribution of the optical flow magnitude.The variable scale 3D-HOF can solve the problem that the optical flow in 3D-HOF is unevenly distributed in each magnitude bin,and it can more accurately extract moving information.At first,the optical flow magnitude is divided into several variable scale intervals.Then,according to the distribution of the optical flow magnitude,different optical flow magnitude scales are set in each variable scale interval,so that the distribution of the optical flow magnitude in the histogram is more uniform.Then,we present an anomaly detection algorithm that supporting online update for combination set.In training process,we train the sparse combination set by the variable scale 3D-HOF features and the information entropy of the optical flow,and the detection accuracy is improved.In detection process,we add the online update for the combination set to increase the robustness of sparse combination learning algorithm,and the ability of the adapt to the dynamic changes of data is enhanced.The effectiveness of the algorithm is finally verified by experiments.Finally,we design and implement an anomaly detection system for surveillance video based on the basis of the above work.First,we introduce the system design and present the system function.Then,we introduce the system function based on the system design.Finally,the experimental results of the system are evaluated.