Visual Motion Estimation

Abstract

ContentsThis appendix presents the principle of the visual motion estimation technique used for SOLERO. Such a technique allows computing an estimate of the 3D camera motion between two stereo pairs acquisitions1. The six displacement parameters are computed on the basis of a set of 3D point-to-point matches established by tracking the pixels in the image sequence acquired by the robot. The following figure summarizes the approach: 1 At time t, a stereo pair is acquired and the Harris corner detector [29] is applied to extract interest points from both images. Then, the point-matching algorithm presented in [35] is used to find correspondences between the interest points in both images. Finally, the cloud of 3D points is obtained using stereovision and a second outlier rejection cycle is performed [36]. Stereovision is only computed for the interest points in order to reduce the computation time. 1 At time t+1, a new stereo pair is acquired and the Harris points are again extracted from both images. Then, the correspondences between the interest point extracted in the left images (acquired at time t and t+1) are searched using the same technique as presented in 1. 1 The stereovision is used to compute the cloud of 3D points at time t+1. 1 Finally, the six displacement parameters between t and t+1 are computed using the least square minimization technique presented in [28]. The approach is robust and provides accurate motion estimates. A precision of 1% can be attained after a 100-m traverse. More details about the visual motion estimation technique can be found in [36]. In particular, the error model for the six displacement parameters is presented. This uncertainty model is used in Chapter 5 for sensor fusion.