Sequential Monocular Road Detection by Fusing Appearance and Geometric Information

This paper presents a monocular road detection algorithm for outdoor scenes in the driving process of mobile robots, which is an essential problem for autonomous navigation. Both appearance and geometric information are fused for more robust and accurate detection. This paper makes use of the image sequence captured during the driving process. The road segmentation is formulated into a maximum a posteriori estimation problem. The probability distribution is constructed by the image likelihood potential, the spatial smoothness potential, and the temporal smoothness potential. The image likelihood potential is decomposed into the probabilities according to road appearance model, geometric information, and edge preference. The spatial smoothness penalizes uneven segmentation in the image space. The temporal smoothness penalizes inconsistent segmentation during the image sequence. After the road segmentation in image space, a postprocessing procedure is performed in the Bird's Eye View for smoother and more accurate road detection. The postprocessing is formulated into a parameterized optimization problem considering structure, appearance, and geometry preferences. Experiments are conducted to show the effectiveness of the proposed method.