Object segmentation in image sequences using motion and color information

grantor: University of TorontoAccurate segmentation of moving objects in image sequences is of paramount importance for many object based multimedia applications. In this thesis, we present a novel automatic, multi-frame, region-feature based object segmentation technique. It combines the advantages of feature based methods and gradient based methods. Salient region features are extracted from the first two frames of an image sequence and are tracked over a number of frames. Trajectory clustering is then performed to group these features into putative objects, from which a set of motion models are estimated. Final segmentation result is obtained by region classification based on these motion models. The proposed technique uses both static and motion information to precisely localize object boundaries. It provides reliable and coherent interpretation of the scene over time by exploiting temporal information from several frames. Experimental results on a variety of image sequences clearly show its advantages over traditional techniques.M.A.Sc