Tracking a moving hypothesis for visual data with explicit switch detection

The use of support vector (SV) methods has been successful in many areas involving pattern recognition. Video surveillance requires pattern recognition algorithms that are efficient in their operation, and requires the use of online processing for the detection and identification of events, objects, and behaviours. To successfully use SV methods in video surveillance, on-line training methods must be employed; NORMA is one such training method. A video surveillance system represents a dynamic system with non-stationary characteristics. It is the purpose of our work to enhance NORMA to better adapt to sudden changes (switches) in the surveillance environment. We show that the decision hypothesis that NORMA generates is more accurate when a switch in the data is explicitly detected and managed. Our preliminary testing involves simulated data, real world benchmark data, and real video data captured from a digital camera.