Multi-scale Ill-structured Road Detection

Ill-structured road scenarios are complicated due to inhomogeneous road surface and the lack of clear boundaries. In this paper, we propose a novel road boundary detection approach which is based on the multi-scale detection scheme and the use of patch-wise boundary cues. A characteristic scale range of road boundaries is first defined and estimated as a priori knowledge. Then, the patches with high local curveness strength are selected at each characteristic scale as the candidates potentially straddling the road boundaries, and are further validated by several cues globally. The road boundaries are finally localized precisely within the regions delimited by the candidate patches. The proposed approach reaches real-time requirements and has been tested on thousands of image frames covering a variety of challenging road scenarios. Experimental results show the effectiveness and robustness of our algorithm.