Drivable road region detection using homography estimation and efficient belief propagation with coordinate descent optimization

Road detection is one of the key issues for the implementation of intelligent vehicles. In this paper, we present a drivable road region detection method using homography estimation and efficient belief propagation. In the method, each pixel in stereo images is assigned a label by minimizing an energy function that accounts for the planar road region, which is defined by utilizing the 2D projective transformations of stereo information and the inference algorithm in binary piecewise Markov Random Field (MRF). The energy is minimized in coordinate descent iterations that alternate between optimizing the homography induced by the planar road plane and implementing efficient belief propagation to find the optimal binary labeling that segments the image into two non-overlapping road and non-road regions. In the optimization process, both image evidence and temporal information are used; meanwhile, an error correction mechanism is applied. Therefore, more accurate as well as robust detection of the road region can be expected. Experimental results on a wide variety of typical but challenging real road scenes have substantiated the effectiveness as well as robustness of the proposed method.

I. Introduction

Road detection has attracted a lot of interest on research of intelligent vehicles. Many researchers have been studying it for several decades and dramatic development has been accomplished. Among them, vision-based method is one of the most important branches in the field. Some methods use monocular camera to extract the road region by employing features with specific intensity, color and texture as visual cues on the road surface [1]–[5]. These methods are mainly supposed to be used in well-arranged environments such as urban streets and highways, since they have to depend on specific intensity patterns in the image, such as lane markings, pavement colors, etc. Others use binocular camera (or more cameras) for road detection by utilizing 3D structural information [6]–[9]. In practice, stable reconstruction of the 3D structure by computing the disparity map is difficult due to the requirement of solving the correspondence problem for every pixel. However, to most of the stereovision-based road detection systems when assuming the road to be detected is nearly flat, the complete 3D structure reconstruction is not necessary, but only the detection of planar region on the road plane where a vehicle can pass safely is needed. We have developed a method [10] that dynamically detects the planar road region through 2D projective transformations between the stereo image pair by computing the homography induced by the road plane. The road region is extracted using the SAD (Sum of Absolute Difference) matching technique with the transformed image and the reference image. It works well in many typical but challenging scenarios including both urban roads and rural roads. However, the results in the unhomogeneous road surfaces with complex intensity patterns, such as the road shown in Fig. 8(f), are not accurate enough since the SAD matching technique is sensitive to the intensity patterns of the image. To overcome the sensitivities brought by complex unhomogeneous road surfaces, in this paper, we formulate the road detection problem as a Markov random field (MRF) model and implement an efficient belief propagation approach to obtain the Maximum A Posteriori (MAP) estimation in the MRF, since the MAP-MRF approach has proven to be extremely successful for many vision applications [11].