A Comparative Study of Two Vertical Road Modelling Techniques

Abstract

Binocular vision combined with stereo matching algorithms can be used in vehicles to gather data of the spatial proximity. To utilize this data we propose a new method for modeling the vertical road profile from a disparity map. This method is based on a region-growing technique, which iteratively performs a least-squares fit of a B-spline curve to a region of selected points. We compare this technique to two variants of the v-disparity method using either an envelope function or a planarity assumption. Our findings are that the proposed road-modeling technique outperforms both variants of the v-disparity technique, for which the planarity assumption is slightly better than the envelope version.

Dense Stereo-Based Robust Vertical Road Profile Estimation Using Hough Transform and Dynamic Programming

“…In particular, in the case of the distant road surface, it is hard to accurately select ROIs because errors of the driving corridor prediction and triangulation dramatically increase. This might be the reason why Keller et al [10], [11] utilized this method only for road surfaces up to 30-40 m. Schauwecker and Klette [12] presented a robust estimator-based method. This method iteratively performs least squares estimation of a B-spline curve by growing the road surface region participating in the estimation process.…”

“…One is that the performance of the M-estimator is highly dependent on the initial estimate, which is difficult to correctly establish in the case of a large proportion of outliers [26]. The other is that the larger the proportion of outliers, the higher the possibility that the empirical binary weighting function selects the wrong 3-D points [12]. This situation frequently occurs when a road surface is severely occluded by obstacles close to the ego-vehicle or the proportion of outliers increases on the distant road surface due to the perspective projection.…”

“…Among the popular vertical road profile models discussed above, a cubic B-spline curve has been known as the most accurate and flexible model because of its high DOF [8]- [12]. Wedel et al [8], [9] estimated a cubic B-spline curve-based vertical road profile using a Kalman filter.…”

“…Keller et al [10], [11] utilized a region of interest (ROI)-based outlier removal approach. Schauwecker and Klette [12] presented an M-estimator-based region growing method. Despite all these efforts, the previous methods are unable to work properly in cases of non-Gaussian noises, increasing triangulation errors at the distant road surface, and a larger portion of outliers.…”

“…A B-spline curve has been presented as a generalization of polynomial and piecewise linear functions [8]- [12]. Among these popular vertical road profile models, a cubic B-spline curve has been known as the most accurate and flexible model.…”