Stereo Camera-Based Road Surface Profile Estimation for Moving Vehicles

Abstract

Road surface profile (RSP) estimation is to estimate the longitudinal shape of a road surface. It is useful to recognize imperfections of a road surface (e.g. bumps and potholes) in advance and has been therefore employed inactive body control systems to reduce shock to the ego-vehicle. Recently, several stereo-based RSP estimation methods have been proposed because the sensor can reconstruct a 3D shape of a road surface and measure its elevation quite accurately at an affordable price. However, the performance of the existing stereo-based RSP estimation method is limited due to noisy depth estimation, the disturbance of obstacle regions, inconsistent reference grids, etc. In this thesis, I propose a new framework for temporally consistent RSP estimation using a stereo image sequence. The proposed method recognizes non-obstacle regions using free space estimation results, which is used to reconstruct a road surface in the 3D space, i.e., digital elevation maps (DEM), without the disturbance of obstacle regions. Furthermore, the proposed method updates DEM and its reference grid over multiple frames using ego-motion estimation results from visual odometry (VO) and by considering geometric relationship between a ego-vehicle, camera, and road surface. Consequently, the experimental results in real-world driving scenes verify that these schemes provide temporally consistent RSP estimation results. In addition, I propose the methods to improve the performance of VO and free space estimation, i.e., robust and fast free space estimation by optimization-based non-parametric road surface modeling and integral disparity histogram-based techniques and accurate ego-motion estimation by the three point-based direct VO and stereo scale-guided VO. Finally, the enhanced algorithms are integrated with the proposed RSP estimation framework and produce the superior results.