Adaptive road detection towards multiscale-multilevel probabilistic analysis

Vision-based road detection is a challenging problem because of the changeable shape and varying illumination. Though many efforts have been spent on this topic, the achieved performance is far from satisfactory. To this end, this paper formulates a Bayesian method which simultaneously explores the multiscale-multilevel clues that are considered to be complementary. Two contributions are claimed in this proposed method. 1) By computing the prior distribution in super-pixel-level with a novel Laplacian Sparse Subspace Clustering and observation likelihood in pixel-level with statistical color similarity, the posterior probability of road region can be effectively inferred. 2) To ensure the adaptivity of road model in various conditions, a multiscale strategy is presented to fuse the detection results of different scales. Experimental results on several challenging video sequences verify the superiority of the proposed method compared with several popular ones.