Vision-Based Probabilistic Map Estimation with an Inclined Surface Grid for Rough Terrain Rover Navigation

This paper describes a novel map representation called Inclined Surface Grid (ISG) maps that provides the ability of modeling traversability of an environment. In order to accomplish safer navigation autonomously with a planetary rover by avoiding hazardous areas, appropriate traversability of the terrain has to be evaluated. In the ISG maps, the patch on each grid stores not only traversability but more detailed information — height, slopes (roll and pitch angle) and roughness. The ISG maps have higher degree-of-freedom knowledge about the surface compared with two-dimensional grid or elevation maps and have very good performance to analyze traversability. In order to assure the reliability of surface estimation in far range areas where the point cloud density decreases, an expanding sampling area approach is introduced to homogenize the point density in a grid by increasing the sampling points. Moreover, probabilistically updating using a stereo vision-specified error model is introduced to estimate the information of each patch. Thus, probabilistic updating allows the ISG maps to yield few failed patches. In the experiment, the resulting geometrical information of our ISG maps converged into an accurate geometry of terrain more than the simple overwriting approach. It is shown that our approach can also create the appropriate traversability map in rough terrain.