This paper proposes a novel method for detection and segmentation of foreground objects from a video which contains both stationary and moving background objects and undergoesboth gradual and sudden “once-off” changes. A Bayes decision rule for clas