“…The detected change image was then analyzed to find if the changes correlate to the prediction of Bayesian background model [78] or are they likely to be foreground. The regions classified Relatively stationary pixels [17], [18], [37], [88]− [90] Temporal filter extracts statistical representative of background Sliding temporal window, change from reference, spatial smoothing Simple, computation efficient, online learning Cannot deal with highly dynamic backgrounds, e.g., wakes Spatio-temporal filtering approaches [22], [39], [40] Background is modelled as low spatial frequency component, albeit with temporal variation Sliding temporal window and fixed spatial window (blobs or neighborhood) are used for filter parameters' update Simple, online learning, computation efficient, robust to small dynamics and illumination variation Cannot deal with highly dynamic backgrounds, e.g., wakes GMM [12], [77] Intensities at a pixel as mixture of Gaussian distributions Multi-step approaches [13], [21] Combination of more than one technique More robust and versatile, often made adaptive and capable of dealing with wakes Complicated, computation intensive, slow due to frequent feedback -feed-forward steps as foreground were then used with color-based segmentation approach to further strengthen the foreground estimation and thus contribute to more robust background detection. The background thus determined was used to update the reference stationary image and the Bayesian background model.…”