Illumination-invariant road detection for monocular camera via patch propagation

Abstract

Road detection is one of the most crucial problems for advanced driving assistance system. In this paper, we propose an efficient illumination-invariant road area detection method via a vision-based approach using an onboard monocular camera on a vehicle. Accurately identifying road areas is challenging due to the road variability caused by illumination changes. To address this problem, we propose to use PCA based illumination invariant grayscale image computed from an input RGB image. We assume that road areas in the grayscale images have the similar texture appearance, and those areas are connected to each other. Based on these assumptions, we develop an efficient, yet robust patch based techniques for identifying seed road patches and propagating them to find the road area. We have tested our method with the KITTI benchmark, and compared the performance against three state-of-the-art techniques. We found that our method shows higher accuracy in a range of 7% to 23% points over the tested methods. Moreover, our method runs two orders of magnitude faster than other tested methods. These results are mainly achieved thanks to the usage of the illumination invariant grayscale images and our patch based approach, which is efficient and robust to various noise.