Detection of moving shadows using mean shift clustering and a significance test

An algorithm that discriminates moving objects from their shadows is presented. Starting from the change mask of an image sequence, first of all the changed area is devided into subrogions consisting of pixels with similar colour properties. This is done using the mean shift algorithm which is very powerful in non-parametric clustering of data. In a second step a significance test is performed to classify each image pixel inside the change mask into one of the classes foreground or shadow. To do this a straight-forward image model is used where the grey-level of a fore-ground pixel covered by a shadow is given by the product of the corresponding background pixels' grey-level and a constant value. Assuming that fore- and background images are corrupted by Gaussian white noise, a significance test is derived which classifies all pixels inside the change mask. In the third step global and local information from the first and second steps are combined. For each region inside the change mask it is examined if the majority of pixels survived the second step. If this is the case, the whole region is kept for the final moving object mask, if not the region is set to zero.