Robust Regression for the Detection of Independent 3D Motion by a Binocular Observer

method is proposed for the visual detection of objects that move independently of the observer in a 3D dynamic environment. Many of the existing techniques for solving this Aproblem are based on 2D motion models, which is equivalent to assuming that all the objects in a scene are at a constant depth from the observer. Although such methods perform well if this assumption holds, they may give erroneous results when applied to scenes with large depth variations. Additionally, many of the existing techniques rely on the computation of optical flow, which amounts to solving the ill-posed correspondence problem. In this paper, independent 3D motion detection is formulated using 3D models and is approached as a problem of robust regression applied to visual input acquired by a binocular, rigidly moving observer. Similar analysis is applied both to the stereoscopic data taken by a non-calibrated stereoscopic system and to the motion data obtained from successive frames in time. Least Median of Squares (LMedS) estimation is applied to stereoscopic data to produce maps of image regions characterized by a dominant depth. LMedS is also applied to the motion data that are related to the points at the dominant depth, to segment the latter with respect to 3D motion. In contrast to the methods that rely on 2D models, the proposed method performs accurately, even in the case of scenes with large depth variations. Both stereo and motion processing is based on the normal flow field, which (in contrast to the optical flow field) can be robustly computed from the spatiotemporal derivatives of the image intensity function. Although parts of the proposed scheme have non-trivial computational requirements, computations can be expedited by various ways which are discussed in detail. This is also demonstrated by an on-board implementation of the method on a mobile robotic platform. The method has been evaluated using synthetic as well as real data. Sample results show the effectiveness and robustness of the proposed scheme.