Motion-based video segmentation using fuzzy clustering and classical mixture model

Motion-based segmentation plays an important role in dynamic scene analysis of video sequences. We present a scheme for extracting moving objects. First, three different resolutions of the dense optical flow fields are calculated using a complex discrete wavelet transform. Surface fitting of all levels of these vectors is then performed over the affine parametric motion model. Next, the clustering by the competitive agglomeration algorithm is applied in the parameter space of the coarsest level. The results of this step are the optimum number of clusters and the center of each cluster. Using information from the previous level, the parameter spaces of the following levels are then segmented using the classical mixture model and the expectation-maximization algorithm. Finally, the individual moving object and background are represented in layers. Experimental results showing the significance of this proposed method are provided