Research On Abnormal Event Detection In Video Surveillance

Abnormal event detection in video surveillance is an important part of the intelligent video surveillance system.In recent years,this issue has been attentioned in many fields,such as image processing,machine learning and data mining of video.It is important to research on the abnormal event detection in theoretical and practical values.The process of abnormal event detection includes pretreatment of image,representation of basic events,construction of abnormal event detection model and judgment of abnormal events.This paper makes a deep research and analysis on the existing detection methods.There are some problems of existing methods for basic events representation.They ignore the detection of the global abnormal event and make a insufficient use of the feature descriptors.To solve these problems,some improved algorithms are proposed in this paper.These algorithms can improve the performance of detection and reduce the computational complexity.The main contents of this paper are as follows.On the one hand,we introduce the abnormal event detection algorithm based on the hierarchical feature representation.We found that this method has some defects.First,the calculation method of spatio-temporal interest points contains much noise.Second,there is a high computational complexity of the GPR model selected.In order to overcome these defects,we make propose improved algorithms to overcome these defects.Spatio-temporal interest points represent local events as low-level features.And ensembles of multiple spatio-temporal interest points represent global event as high-level features.The clustering and regression models are respectively used to judge the local and global anomalies.To solve the first defect,a noise reduction method is put forward.The visual background extractor is used to calculate the foreground mask.Next,the foreground mask is used to filter the spatio-temporal interest points.For the second defect,we propose the least square method to model the global events,which reduces the computational complexity of the model on the premise of guaranteeing the basic detection rate.On the other hand,we introduce the abnormal event detection algorithm based on cells.We find some shortages caused by rough feature descriptors.For example,the motion feature descriptor cannot detect the abnormal velocity direction and it is not easy to distinguish between normal and abnormal texture by rough texture feature.To fetch up these shortages,we make some improvement with experimental verification.The basic thought of this algorithm is as follow.First,each input frame is split into non-overlapping cells.Next,cells are modeled by several low-level visual features of themselves.Then,the corresponding classifiers are established.F inally,abnormal events cell-based are determined.For the problem of coarse motion feature descriptor,a feature called HOG3 D is proposed.The HOG3 D feature,based on the foreground object under 3D polar coordinates,can detect the foreground object's velocity and direction at the same time.In order to solve the problem of rough texture feature descriptor,the uniform pattern of LBP feature is used as the texture feature to effectively judge the abnormal texture of the foreground object.