Efficient wavelet based detection of moving objects

Moving object detection in video sequences presents an important problem in computer vision. If a video sequence is generated by a stationary video camera, one usually attempts to build a statistical model of the background and an appropriate statistical test to classify pixels into foreground or background. This approach is efficient for many laboratory test sequences, but may render itself inadequate in real-life surveillance systems with many additional scene-, illumination- and camera-related phenomena. In this case better results can be obtained with frame differencing scheme, that is unfortunately prone to aperture problem and leads to inconsistent detections. In this paper, we propose a multiresolution frame differencing technique. Each frame is first decomposed into undecimated wavelet transform coefficients and after that, differencing scheme is applied on wavelet coefficients in several bands separately. These band-dependent motion detections alleviate the aperture problem and when fused, they produce more consistent moving object detection. The obtained detection results greatly facilitate later processing steps, like object tracking and recognition.