End-To-End Road Graph Extraction Based on Graph Neural Network

Generating road maps from satellite/aerial imagery is a challenging machine learning task due to occlusions and complex traffic environments. Previous works either predict the graph iteratively or employ rule-based algorithms to construct the graph from a pixel segmentation of road areas. In this study, we propose an end-to-end road graph extraction framework to detect road centerline keypoints with convolutional neural networks and construct the road topological structure with graph neural networks. Our method predicts the road graph in a single iteration and does not require complex post-processing rules. We evaluate the method against existing solutions on aerial imagery of urban areas across US cities. The qualitative and quantitative results show that the method extracts highly accurate road graphs of different road conditions.