Conventional stereo systems for navigation use three basic steps. First, matches between pixels in the left and right images are computed. Second, the matches are converted to points in 3-D space. Finally, the 3-D points are arranged into a map which is used for evaluating possible vehicle paths and for selecting the safest path.
This conventional approach has been demonstrated successfully in a limited number of systems [14][9]. However, this approach suffers from several limitations that compromise its use in full-fledged navigation systems, especially in areas that cannot be easily accessed, e.g., as in planetary exploration scenarios. In particular, one major limitation is the need for precise camera calibration.
We set out to develop an alternative more general and robust approach by eliminating the conversion to Cartesian coordinates and the building of a dense elevation map. More precisely, our approach is to initially compute some of the parameters of the stereo system by matching natural features in the environment. These parameters are used at run-time for computing stereo matches between the two images and for deriving geometric cues such as the elevation of the terrain at the matched pixels. The parameters are also used for projecting candidate paths into the images. The candidate paths are evaluated by comparing them with the distribution of elevation values in the image. An optimal vehicle path is selected based on the result of the evaluation. The image processing and the path evaluation are fast enough that the vehicle path can be updated at a rate high enough to permit continuous motion at moderate speed.
The remainder of this section is organized as follows; in Section 2.1, we describe our algorithm for stereo matching; in Section 2.2, we describe our approach to computing relative elevations from stereo matches. In Section 2.3, we describe how the stereo system is integrated into a complete driving module, and, particular, how paths can be evaluated based on the result of stereo matching; finally, in Section 2.4, we discuss the performance of the system based on experiments on the HMMWV testbed. We also discuss our current work and future plans, including moving the system to the Ratler.