Self-Supervised Online Long-Range Road Estimation in Complicated Urban Environments

摘要

In this paper, we propose a long-range road estimation method for autonomousmobile robots in unstructured urban environments. Near-range road surface conditions are estimated by using remission value as reflectivity of a laser scanner. Graph cut algorithm is applied to estimate road region robustly also in complicated environments. Moreover, we propose a novel image segmentation method to estimate long-range road surface. A compact texture/color feature is integrated with level-set method to estimate precise road boundaries robustly. Our proposed image segmentation approach gives better performance compared with standard classification approach. Finally, we run our autonomous mobile robot in “Tsukuba Challenge 2009” and our university campus, and experimental results have shown a marked increase accuracy in road estimation over standard methods.