End-to-End Road Centerline Extraction via Learning a Confidence Map

Abstract

Road extraction from aerial and satellite image is one of complex and challenging tasks in remote sensing field. The task is required for a wide range of application, such as autonomous driving, urban planning and automatic mapping for GIS data collection. Most approaches cast the road extraction as image segmentation and use thinning algorithm to get road centerline. However, these methods can easily produce spurs around the true centerline which affects the accuracy of road centerline extraction and lacks the topology of road network. In this paper, we propose a novel method to directly extract accurate road centerline from aerial images and construct the topology of the road network. First, an end-to-end regression network based on convolutional neural network is designed to learn and predict a road centerline confidence map which is a 2D representation of the probability of each pixel to be on the road centerline. Our network combines multi-scale and multi-level feature information to produce refined confidence map. Then a canny-like non-maximum suppression is followed to attain accurate road centerline. Finally, we use spoke wheel to find the road direction of the initialized road center point and take advantage of road tracking to construct the topology of road network. The results on the Massachusetts Road dataset shows an significant improvement on the accuracy of location of extracted road centerline.