Visual Odometry

Problem

Mobile real-time systems require an estimate of their movement. This can be determined from odometry or different kind of sensors, including cameras. The change in the camera image between successive frames allows to determine the motion with five degrees of freedom from mono-cameras (i.e. scale cannot be recovered from images alone) or six degrees of freedom from stereo-cameras with known relation between the cameras (i.e. pre-calibrated) and synchronized capturing of images.

Method

The stereo ego-motion method works on successive images of the left camera of a synchronized and rectified stereo camera system. It is based on identifying image features that are likely to be found again in successive images. The well known Harris corner detector is used for this purpose. Correspondences between corners in one image with corners in the next (left) image are determined by Rank correlation, which can be fast computed and is known to be tolerant against geometric and radiometric outliers within the correlation window. Thus all corners of one image are compared to all corners of another image and vice versa (for consistency checking). Robustness against correspondence outliers are gained by reconstructing all points three dimensionally using the disparity image of a stereo algorithm and enforcing the rigidity constraint of corresponding point pairs. This is done by comparing the relative distance of two points as reconstructed from one stereo image with the relative distance of point pairs as reconstructed from the next stereo image. If the relative distances differ, than at least one correspondence is wrong. A special algorithm applies the constraint to all points for finding a large group of points with consistent distances.

After this outlier detection, motion computation is straight forward and works by computing the motion in closed form using a singular value decomposition, followed by a non-linear optimization that minimizes the re-projection error.

The detailed description of the method has been published (Hirschmueller et al., 2002; Hirschmueller, 2003).

Results

The stereo ego-motion method is used for real-time applications in mobile robotics, where a synchronized stereo camera is available, e.g. crawler.

Publications