Colour co-occurrence and edge based background modelling with conditional random field shadow removal

Good foreground detection is important as the basis to many other video processing algorithms. Current approaches are able produce accurate detection but under tight conditions and constraints. Thus, we have developed an algorithm that proved to be robust to shadow and afterimage noise, colour similarity between foreground and background and slight movements of the moving foreground. Our approach to weaken the constraints is by fusing gradient based background modelling, a colour co-occurrence based approach, and a 3D conditional random field shadow and afterimage removal. Our results show improvement over existing algorithms in terms of robustness and accuracy. Our method is suitable to be implemented alongside other video processing techniques such as people counting, face recognition, etc.