Scene analysis for fire detection

Scene analysis is an important component of vision systems. This paper proposes a computationally efficient background modeling method based on a non-parameter statistical test. The image colors and temporal motion are integrated to capture both the temporal and spatial context of the video sequence. An updating scheme is applied for automatic maintenance of the background corners. An experiment designed for fire detection in video surveillance demonstrates that the model can solve the problems including change of illumination, short period of occlusion and small motion of background objects.