Vision-based obstacle detection and path planning for planetary rovers

NASA's next-generation Mars rovers are capable of capturing panoramic stereo imagery of their surroundings. Three- dimensional terrain maps can be derived from such imagery through stereo image correlation. While 3-D data is inherently valuable for planning a path through local terrain, obstacle detection is not fully reliable due to anomalies and noise in the range data. We present an obstacle-detection approach that first identifies potential obstacles based on color-contrast in the monocular imagery and then uses the 3-D data to project all detected obstacles into a 2-D overhead-view obstacle map where noise originating from the 3-D data is easily removed. We also developed a specialized version of the A* search algorithm that produces optimally efficient paths through the obstacle map. These paths are of similar quality as those generated by the traditional A* , at a fraction of the computational cost. Performance gains by an order of magnitude are accomplished by a two-stage approach that leverages the specificity of obstacle shape on Mars. The first stage uses depth-first A* to quickly generate a somewhat sub-optimal path through the obstacle map. The following refinement stage efficiently eliminates all extraneous way-points. Our implementation of these algorithms is being integrated into NASA's award-winning Web-Interface for Telescience.