Fast abnormal event detection from video surveillance

Video surveillance systems are becoming increasingly important both in private and public environments to monitor activity. In this context, this paper presents a novel block-based approach to detect abnormal situations by analyzing the pixel-wise motion context, as an alternative for the conventional object-based approach. We proceed directly with event characterization at the pixel level, based on motion estimation techniques. Optical flow is used to extract information such as density and velocity of motion. The proposed approach identifies abnormal motion variations in regions of motion activity based on the entropy of Discrete Cosine Transform coefficients. We aim at a simple blockbased approach to support a real-time implementation. We will report successful results on the detection of abnormal events in surveillance videos captured at an airport.