Detect Anomaly Event In Video And Study Sparse Representation In Feature

With the development of computer technology and innovation of video monitor technology,smart video monitor has been increasing attention.Video anomaly detection,as an important part of smart video monitor,charactered with real-time,smart and economic,have great academic value and excellent business outlook in public safety assurance.It can analyze what happens in video monitor.Once something abnormal happen,it will give the alarm and shorten the time to solve problem.This dissertation mainly focuses on two aspects of the abnormal event detection.One is the presentation of primitive events,and the other is the construction of anomaly detection model.The details of this dissertation are as follows:1.Considering the advantages of spectral clustering,we proposed a novel algorithm for video abnormal event detection based on spectral clustering.The algorithm consists of training and testing.During the training phase,optic flow features are extracted to describe the local volumes and taken as the atoms to build a graph The local fetures whose dimensionality is reduced by Laplacian Eigenmap(LE)is clustered by an adaptive clustering algorithm and corresponding cluster centers are obtained as code word to construct code book.During testing phase,the previous 8-neighborhood information,and similarity between the extracted local features of the test samples and the code book are used to detect the events.The results show that the proposed method has good detection peformance in abnormal events.2.Considering the geometric structure feature,we proposed an abnormally detectection model based on graph regularization.The video is partitioned into sub-blocks from which the local feature is extracted,and then A graph regularization based sparse coding model for detection is constructed and a dictionary for local features of normal behavior characteristics is studied in training stage The reconstruction error of local features of the test video is obtained based on the model and used to detect the event based on the its difference to the predefined threshold.The results show that the proposed is effectin monitoring abnormal events.3.Considering the powerful ability of gaussian mixture model(GMM)to categorize data and the efficient information expression of sparse representation,we proposed a novel algorithm for video abnormal event detection based on GMM and sparse representation.GMM is used to establish the similarity matrix as the dictionary in the training The reconstruction error of nuclear vector,which is consists of the kernel function calculating from the local features and mean matrix and the prefinded threshold are compared to determine detect the event.The results show that the proposed works well in monitoring abnormal events.