Automatic Incident Detection in Video Suveillance

Abstract

A new generation of video based detection systems for use outside has recently been developed. They are intelligent detection systems which provide automatic surveillance through real-time video analysis and event detection. In doorway scenarios, the goal of such systems is to be able to automatically detect any incident occurring in front of a door. This thesis introduces several tools useful to create an automatic incident detection system in video surveillance. Every event occurring in the scene filmed by the camera is first detected. Then, classification techniques will be described so that a robust human classifier can be constructed; assuming the humans are the only objects able to cause an incident. A probabilistic video tracking algorithm will then be presented. It will be used to track the detected human beings over time in order to estimate their trajectories and know if an alarm should be given or not by the system. The theoretical backgrounds to these tools will be explained. Some results will also be shown. Toy data will first be used to validate the different tools separately. Global results on real data will be displayed at the end of the thesis.