Image Segmentation Based On Optical Flow

The recent development of computer has made it possible to research motorial image sequences, including multiple moving objects especially. It requests that the technique of image segmentation grows rapidly. Single immobile image was segmented to find the foreground objects from the background ago. But now image sequences including multiple moving objects need to be segmented and grouped into segments according to motorial characters of moving objects.Optical flow field contains motorial information of moving object, which supports the image sequences segmentation. The main interest of this paper is image segmentation based on optical flow field. Firstly, some known methods of optical flow estimation are integrated and a method based on gradient is confirmed to estimate optical flow; secondly, the transforming relationships from 3D moving field to 2D vector field are analyzed and three optical flow models are modeled, which are affine transform model, planar optical flow model and general quadric polynomial model respectively; lastly, a novel image segmentation method based on optical field is proposed. The method is a region growing method, which consists of three main operations: preliminary segmentation, main segmentation and supplementary segmentation. The preliminary segmentation is a procedure that divides the image pixels into regions whose pixels have approximate optical flow. The main segmentation makes regions grow progressively in virtue of the residual sum of squares(RSS) to fit the local optical flow with a optimal optical flow model, which is selected from three optical flow models. The supplementary segmentation merges insular pixels or regions. Simulation results show the method can segment image sequences with still background.