Research On Anomaly Detection Algorithm In Video Surveillance

With the development of our society,more and more social security problems also arise.As an important issue of public security,how to avoid and deal with public e-mergencies timely has been extensively studied by the academia and industry,resulting in a variety of security mechanisms and the corresponding technology.Video surveil-lance technology as a widely used regulatory technology,in lots of applications played a crucial role.However,the current general surveillance technology still has lots of disadvantages.As an important subject of computer vision,the analysis of surveillance video con-tent has always been the concern of the academic community.The analysis of video con-tent mainly includes target detection,target recognition,target tracking and other major areas.This dissertation mainly focuses on anomaly detection in video surveillance.Firstly,the basic concepts and basic theories of anomaly are briefly introduced,after which the research status and mainstream algorithms at home and abroad are summa-rized and analyzed.Then,this dissertation proposes the algorithm for global and local anomaly detection and analyzes the implementation and test results using this scheme.The research content and innovation of this dissertation are as follows:1.In this dissertation,we propose a local anomaly detection framework based on muti-feature extraction.The traditional method of anomaly detection generally extracts the characteristics of the anomaly,which will cause the unreasonable use of the mo-tion information and even the loss of detail,which results in a high false alarm rate for anomaly detection.In this dissertation,the local anomaly is further divided into velocity anomaly,appearance and location anomaly by constructing the feature mod-el of local anomaly,and the feature extraction and detection are carried out for each type of anomaly,and finally the unified anomaly score is obtained.Experiment results demonstrate that this classification is reasonable and has its advantages.2.In this dissertation,a foreground instance extraction scheme based on split merge clustering algorithm is proposed.The traditional local anomaly detection feature extraction scheme usually divides the video sequence into fixed-size regular-shaped patches(two-dimensional)or cuboids(three-dimensional),which can cause erroneous segmentation of the target and increase the false positive rate.In this dissertation,we decompose the foreground regions that do not conform to the sample model in each iter-ative process,and combine the similarity regions to achieve the effective segmentation of the foreground and extract the foreground of single or multiple overlapping targets,after which we extract features in these areas.This scheme has a great influence on the feature differentiation degree,so it is also a key factor to determine the effectiveness of the whole algorithm.3.NSH algorithm is improved in this dissertation.Although the original NSH algorithm can be used to detect abnormal data,but it is less efficient and less robust in finding the optimal parameters.The INSH algorithm proposed in this dissertation achieves better classification results than the original algorithm by redefining the ob-jective function and redesigning the convex hull solution under the basic framework.4.In this dissertation,we propose a global anomaly detection algorithm based on global kinetic energy.Most of the traditional global anomaly detection algorithms are complex and the detection results are not satisfactory.In this dissertation,a glob-al anomaly detection algorithm based on global kinetic energy difference is proposed by analyzing the characteristics of large-scale motion.The algorithm realizes the glob-al anomaly detection by constructing the global kinetic energy and the kinetic energy difference.The experiment proves that the method is effective and efficient.