Random-Walker Monocular Road Detection in Adverse Conditions Using Automated Spatiotemporal Seed Selection

A key module of modern advanced driver-assistance systems (ADASs) is the road detector, which has to be robust, even under adverse conditions. The ultimate goal of such a system, which uses only visual information acquired from a color video camera, is to classify each frame pixel as belonging to the road or not. In this direction, this paper proposes a new fully automatic algorithm that combines both time and spatial information using the efficient random-walker algorithm (RWA) as a segmentation tool. A novel technique for automatic seed selection is proposed, utilizing features derived from a shadow-resistant optical flow estimator using the c₁ channel of the c₁c₂c₃ color space, along with a priori information and previous frame segmentation results. The proposed system is qualitatively assessed using video sequences in both typical and adverse conditions, including heavy traffic, shadows, tunnels, rain, night, etc. It is also quantitatively compared with previous efforts on a publicly available manually annotated onboard video database, providing superior results.

I. Introduction

The research on advanced driver-assistance systems (ADASs) has gained increasing interest over the past decade. The incentive for this evolvement was the worldwide effort to reduce road accidents. The triggering effects for the blossoming of ADAS research were the technology advances that allowed cost-effective implementation of real-time image processing algorithms used in such systems. Nowadays, an increasing number of commercially available vehicles are equipped with lane detectors activating lane departure warnings, parking radars, collision warning sensors, etc. The goal of a completely autonomous commercially available vehicle has yet to be reached due to the vast diversity of driving environments and the unpredictability of incidents that could occur. When the available data used are limited to the visual information taken from onboard cameras, typical tasks such as road detection and collision avoidance become even more difficult. So far, only assistive systems have been implemented, a variety of which having been presented in [1] and [2].