Roadway feature mapping from point cloud data: A graph-based clustering approach

Abstract

Connected and automated vehicle applications are facilitated by Enhanced Digital Maps (EDMs) of the roadway environment. Due to the high numbers of roadway miles and signalized intersections, there is significant research interest in the automatic extraction of such maps from georectified LiDAR data. Most existing methods convert the LiDAR point cloud to a set of images for feature extraction and mapping. This rasterization step loses information that could be retained if new methods were developed that work directly on the LiDAR point could for feature extraction and mapping, without rasterization. This article presents one such approach that operates on a road surface point cloud, processing small patches at a time using a locally adaptive version of Otsu's method to discard low intensity reflections while retaining reflections from roadway markings. The main new aspect of the approach is a graph-based clustering algorithm implemented directly on the point cloud. A cluster growing method is used to group similar road markings into the same group to enable detection of the stop bars and lane edges. Finally, a SAE-J2735 map message is created from the extracted roadway features.