Abnormal Event Detection in Surveillance Videos using Spatio-temporal Compositions

Abstract

Anomalies are events and actions that occur with unexpected ap pearance or motion patterns compared to normal ones in a spe cific environment in a video sequence. Anomaly detection has ap plications in video surveillance, automated security systems and health-care setups etc. In this thesis, we utilize motion informa tion across spatial and temporal dimensions of a video sequence to detect anomalies. To capture the motion dynamics of surveillance videos, we propose a low dimensional descriptor (LDD) comprised of thresholded optical flow (OF) sum, joint entropy and variance features. In order to develop an enhanced representation of the raw optical flow, we present a novel 2D variance plane which encodes the local spatio-temporal variations around a pixel. We identify the salient 2D variance plane pixels by using an ant colony clustering algorithm. Subsequently, we construct a video descriptor based on histogram of swarm (HOS) using a novel predator prey algorithm. In order to detect fall events in vision based health-care setups, we present a heuristic approach for low frame per second (fps) videos. The proposed method detects prolonged inactivity by region based video analysis. The performance of raw OF based LDD and the 2D variance plane based HOS descriptors is evaluated on publicly available crowd anomaly datasets. Experimental results demon strate their ability to tackle the task of anomaly detection with high accuracy. Specifically, the LDD is able to perform in real time due to the simplicity of the extracted features. The heuristics based ap proach is evaluated on a new office dataset recorded at 1 fps. The proposed approach detects all the fall events in the dataset despite the challenging aspect of very low fps