Research On Anomaly Detection In Videos Of Crowded Scenes

Automatic analysis of densely crowded environments such as subways, railway stations and airports has been a recent interest in computer vision. The task of anomaly detection is to detect deviations from extremely crowded environments. Since anomaly detection provides much more valuable hints than normal behaviors in crowded scenes, the research on anomaly detection is imperative. Unlike anomaly detection from non-crowded scenes, a crowed environment requires monitoring an excessive number of individuals and their activities using video surveillance devices. As a result, computational approaches of anomaly detection in densely crowded scenes may face more difficulties both in scene modeling and abnormal behaviors detecting.In this paper, we present a novel spatio-temporal framework for modeling the intrinsic structure of a crowded scene and detecting abnormal activities in it. Specifically, we first extract depth information using a stereo video camera to capture pedestrian behaviors along with the normal direction of the camera. Then after dividing the hypervolume into a number of local blocks, we use 3D gaussian distributions and Gaussian Mixture Models (GMM) over a set of depth optical flows to model our motion patterns in each block, respectively. Next, we construct a set of correlation co-occurrence matrices to model the local spatio-temporal contexts using Markov Random Field (MRF). Statistical deviations from two level image pyramids are finally detected as abnormal events.Our method has the following advantages:1) Unlike the existing methods considering only 2D information from surveillant video frames, we integrate the depth constraints along the depth-axis and propose a novel spatio-temporal motion pattern representation to represent local motions within local spatio-temporal position in crowded scenes.2) We present a spatio-temporal context model to reveal the relationships among the motion patterns. Since a crowded scene is actually composed of a large number of activities together with their interrelations, our context model is suitable to describe the interactions between spatially or temporally adjacent activities and enforce local consistency on them.3) We use image pyramid to reduce the interference caused by partial body actions of pedestrians. These will assist in modeling crowded scenes and detecting abnormal motion patterns in a more effective and robust approach.Experiments on a new depth image dataset composed of four crowded scene categories show that our spatio-temporal framework offers promising results in real-life crowded scenes with complex activities.