Scale-Adaptive Pothole Detection and Tracking from 3-D Road Point Clouds

Abstract

3-D road surface modeling has become an essential part of modern algorithms for road pothole detection when 3-D road point clouds are available. This paper introduces a scale-adaptive road pothole detection and tracking framework. It first fits a quadratic surface to the 3-D road point cloud, generated using GPT-SGM, a state-of-the-art disparity estimation algorithm. The surface modeling process also incorporates the normal vector information, obtained by three-filters-to-normal (3F2N), an ultra-fast and accurate surface normal estimator. By comparing the actual and modeled 3-D road surface point clouds, the pothole point clouds can be extracted. Finally, the discriminative scale space tracking (DSST) algorithm is utilized to track the detected potholes in a sequence of successive video frames. Extensive experimental results demonstrate the robustness of our proposed road pothole detection and tracking framework both qualitatively and quantitatively.