Detecting Road Intersections from Satellite Images using Convolutional Neural Networks

Automatic detection of road intersections is an important task in various domains such as navigation, route planning, traffic prediction, and road network extraction. Road intersections range from simple three-way T-junctions to complex large-scale junctions with many branches. The location of intersections is an important consideration for vulnerable road users such as People with Blindness or Visually Impairment (PBVI) or children. Route planning applications, however, do not give information about the location of intersections as this information is not available at scale. As a first step to solving this problem, a mechanism for automatically mapping road intersection locations is required, ideally using a globally available data source. In this paper, we propose a deep learning framework to automatically detect the location of intersections from satellite images using convolutional neural networks. For this purpose, we labelled 7,342 Google maps images from Washington, DC, USA to create a dataset. This dataset covers a region of 58.98 km(2) and has 7,548 intersections. We then applied a recent object detection model (EfficientDet) to detect the location of intersections. Experiments based on the road network in Washington, DC, show that the accuracy of our model is within 5 meters for 88.6% of the predicted intersections. Most of our predicted centre of the intersections (approx 80%) are within 2 metres of the ground truth centre. Using hybrid images, we obtained an average recall and an average precision of 76.5% and 82.8% respectively, computed for values of Intersection Over Union (IOU) from 0.5 to 0.95, step 0.05. We have published an automation script to enable the reproduction of our dataset for other researchers.