A Method for Detecting Multiple Independently Moving Objects from the Sequences Acquired by Active Stereo Cameras and Estimating the Cameras' Egomotion

This paper proposes a modified RANSAC based method that can detect multiple independently moving objects from the video sequences acquired by active stereo cameras, and can estimate the egomotion of the stereo cameras. We first compute 3D optical flows from consecutive frames, where dense 3D optical flows, which are needed for the subsequent egomotion estimation based on the modified RANSAC, are obtained by the process proposed in this paper. From the dense flows, three 3D optical flows are randomly selected for computing an egomotion (model). The validity of that egomotion model is checked using the weight obtained by calculating a consistency between that model and each of the dense flows. If it turns out that the estimated egomotion model is valid, the label that identifies this model is assigned to the inliers, which are consistent with this model, and then the labeled inliers are deleted. The random choice of three flows and its subsequent processes are repeated till one of the termination con itions is met. Using the inliers present in the static background, the real egomotion is computed. Experiments using synthesized and real stereo sequences demonstrate the effectiveness of the proposed method.