A method for segmentation and motion estimation of multiple independently moving objects

The aim of this work is to design a robust method for online estimation of object motion and structure in a dynamic scene. A method for segmentation and estimation of object motion and structure from 3-d points in a dynamic scene, observed by a sequence of stereo images, is proposed. The proposed method is a recursive one where the estimation problem is solved by state estimation of a dynamic system. The method is inspired by a method proposed in [1]. State estimation will be handled by a Rao-Blackwellized particle filter, where the object motion part of the state is represented by a set of weighted samples. Conditioned on the motion, the object structure is represented by Kalman filters. The segmentation problem, defined as the task of segmenting the scene into independently moving objects, is solved by keeping track of which motion samples are compatible with which part of the object structure. By clustering motion samples according to similar object structure compatibility, a segmentation will result where the set of motion samples is divided into subsets, each representing a distinct object motion. The feasibility and performance of the proposed method are investigated by running the method on synthetic and real scenes. Results from these experiments indicate that the method works well for the considered scenes.