Detection of abnormal behaviour in surveillance applications
Floris De Smedt, Toon Goedemé, Tinne Tuytelaars · Lirias (KU Leuven) · 2012
Cameras are present in most public places (train stations, public trans- port, parking lots, ...) with security in mind. These systems are used mostly in a passive way to be used as evidence after an incident. If these systems are used in an active way, human intervention is needed. In this Phd project we develop a method to recognise ”abnormal behaviour” in real-time in camera images. This can be used to support the human intervention by indicat- ing the presence of abnormal behaviour, to increase the efficiency. We are working on an object detector and tracker to follow people in the camera images. Existing techniques are hardened to be robust for the problems of realistic conditions. We will combine the information of multiple cameras to elliminate false detections. Based on the information of multiple frames, the activity of the persons can be determined. Also for this challenge, the existing technology has to be adjusted to the real world conditions. Knowledge representation will be used for the actual classification of whether or not a situation is abnormal. We create a set of rules wich are ”normal” in a given enviroment. Based on this rule-set a set of ”sanity checks” will be executed. A second, and more complex, task of KR is the creation of statemens over the detected activities (for example: ”a person has entered the bus and will walk to a free chair”). These can be used to determine ex- pected future activities of the computer vision part. All these information will be combined in a classificator, trained by machine learning algorithms which are capable of handeling probabilistic information. A real-time im- plementation will be created and tested on public bus transportation. We are using OpenCL to speed-up the object detection and are working on a demo to combine the detection results of cars with KR to verify they measure up to the traffic regulations.