Capitalizing on the boundary ratio prior for road detection

Most of the color based road detection methods use a lower-center region as a “safe” road reference to construct appearance models for the road. However, its utility critically relies on the pose of vehicles. In case of involving non-road pixels, color models trained by using samples from this region often yield erroneous results. To address this problem, we propose a novel color-based road detection method based on a boundary ratio prior, with which we are able to infer the confidence of a certain image region belonging to the road class. Specifically, the boundary ratio prior is defined as the ratio of the length of the coincident boundary of this region and the image bottom to that of the region boundary itself. The calculation of this prior model is realized by a graph based geodesic distance measure. Moreover, the conventional illumination invariance space is integrated to calculate the distance metric of two neighboring nodes in the graph in order to make our approach robust to shadows. Experiments on multiple datasets demonstrate that the proposed approach is efficient and more robust than the existing methods based on the lower-center region prior.