A Multi-frame Stereo Vision-Based Road Profiling Technique for Distress Analysis

Abstract

This paper presents a novel technique for road distress identification using an in-vehicle stereo vision system. The novelty of the proposed method is credited by the accumulation of multiple-frame 3D reconstructions which are properly aligned to a road-centred coordinate system. A 3D plane-fitting technique is first carried out to approximate the road manifold at the beginning of accumulation, followed by the construction of a digital elevation model (for the road being analysed) from multiple frames that are integrated using a stereo visual odometry algorithm. Potholes are detected as "valleys" from the built digital elevation model by means of connectedness analysis. Experimental results show significant improvements (> 30%) over a conventional method based on a single-frame construction of plane-based road models.