Drivable road region detection based on homography estimation with road appearance and driving state models

Road detection is one of the key issues for autonomous driving. In this paper, we present a drivable road region detection method based on homography estimation with road appearance and driving state models. In the method, the planar road region is detected and objects inside the region are localized through a 2D projective transformation between the stereo image pair by computing the homography induced by the road plane dynamically. This method is mainly composed of three modules: 1) preliminary classification module, which selects the most appropriate classifier from the road appearance model to detect the preliminary road-like region; 2) feature-based detection module, which finds the correspondences of feature points on the road plane to estimate the homography for the first image pair, and then extracts the drivable road region; 3) area-based detection module, a nonlinear optimization process, uses the results obtained in module 2 as the initial values for the homography estimation as well as drivable road region detection of the subsequent image pairs with the driving state model based on sequential information. The combination of these three modules uses both image evidence and temporal information; meanwhile, an error correction mechanism is applied. Therefore, more accurate as well as robust estimation of the homography can be expected, and so is the drivable road region detection. Experimental results on real road scenes have substantiated the effectiveness as well as robustness of the proposed method.