Trajectory Analysis-Based Anomaly Detection In Surveillance Video

Surveillance systems provide the capability of collecting information andenhancing safety. Vast amounts of video data render manual video analysis useless, soautomatic video analytics techniques become very important. Unfortunately, recentautomatic video analytic techniques has poor performance. More and more advancedsurveillance systems providing huge amounts of trajectory data of moving objects,such as pedestrians, vehicles, vessels and planes. Vessels’ unexpected stop, vehicles’over speed, pedestrians’ incorrect walk direction may indicate threats or dangers.Thus, there is strong demand for automated detection of abnormal behavior utilizingthe trajectory data.In this thesis, we propose a new approach based on trajectory analysis as aneffective and efficient way for the abnormal detection. Firstly，we translate theposition information of the sample point belongs to corresponding trajectory intospeed information and direction information. Then each trajectory is statisticalmodeled by means of the bivariate kernel density estimation, considering both thevelocity of moving object and the shape of trajectory, at the same time overcome thenoise problem in object tracking. In our work, we present a machine learningframework for abnormal detection, we first cluster the normal/abnormaltrajectories,then we detect the abnormal trajectory by classification. It is unfeasible tobuild a supervised training set, because what a normal/abnormal trajectory/behavior isunknown in priori. We employ an unsupervised training set made up of trajectoriesfor clustering, the only assumption is most trajectories in training collection is normal.The optimal number of clusters is obtained by maximizing the mutual information.Abnormal trajectories are detected by a measure based on Shannon entropy. Theproposed approach is evaluated through a variety of tests on both simulation and realdata Quantitative results for clustering and detection show that our technique whichusing kernel density estimation for statistical model, Bhattacharyya distances forsimilarity compute, mutual information for optimal number select combined withShannon entropy for detecting abnormal trajectory, behaves very well andoutperforms the state-of-art methods.