Reconstructing Road Network Graphs from both Aerial Lidar and Images

Abstract

We address the problem of reconstructing road networks as undirected graphs over large geographic regions in cold start scenarios where neither the preliminary graph nor any on-road trajectory information is available. The goal of this paper is to transform bimodal aerial data in the form of 3-dimensional Lidar scans and high resolution images into road network graphs. We use a fully convolutional architecture that fuses the two datasets by reducing the disparity in their modalities to segment out roads. We then apply a simple, disk-packing based algorithm that covers the segmented regions with a minimal set of variably sized disks, connect the intersecting disks and use a provable curve reconstruction algorithm to obtain the road network graph. We show that our method is better at removing outliers and gives improved connectivity and topological accuracy than the existing state of the art thinning based method.