Research On Anomaly Detection And Tracking In Crowded Scenes

Anomaly detection and object tracking of surveillance video in crowded envi-ronments has been a recent interest in computer vision.In the recent years,with the widespread of monitoring equipments,the related methods of anomaly detection and tracking has been investigated and a significant progress has been made-However,due to the complexity of the crowded scenes,such as the existing noise,varying viewpoint,clutter background and even illumination changes,most current methods do not work well on crowded scenes.In this paper,we explore anomaly detection and object tracking from the crowded scenes.We present a novel spatial-temporal framework for modeling the crowded scenes and detecting abnormal activities.We first extract depth information using a stereo camera,then we divide the hypervolume representation into a number of local blocks and further model the motion patterns for each block.Next,we construct a set of correlation to model the local spatio-temporal contexts using Markov Random Fields.Statistical deviations are finally detected as anomaly events.Then in the research of object tracking in the crowded scenes,we focus on the d-ifficulty of occlusions,which always lead to the failure of tracking and occur frequent-ly in crowded scenes.To achieve this,we hypothesize that when a scen is densely packed,individual movements will be restricted into a coupled pattern and thereby makes themselves relatively regular to be described.Moreover,members in a crowd-ed scene are clustered into groups,which essentially constituting scene contexts for a tracking object since the similar motion.Inspiring by the observations,we integrate the motion analysis method to track arbitrary object in crowded scenes.We model the ap-pearance of a target through the analysis of its motion patterns and verify if occlusions happened by context modeling in the crowded scene.Then we segment the attention-al regions which have similar motions with the target as an auxiliary for tracking the target robustly.By this way we can track a target for a long-term in crowded scenes.Experiments on a new depth image dataset composed of four crowded scene cat-egories show that our spatio-temporal framework offers promising results in real-life crowded scenes with complex activities.The experiment results on two real-life crowd-ed scenes also show the effectiveness of the proposed tracking method in crowded scenes.