Vectorized Road Centerline Extraction from Aerial Imagery Using Reinforcement Learning

Abstract

Road boundary extraction is an important task having applications in city planning and autonomous driving. Extracting them from aerial images is a cost-friendly solution and also scalable. This is usually performed in two steps, semantic segmentation and vectorization. In this thesis, we aim at direct vectorized road boundary extraction using reinforcement learning i.e imitation learning by omitting the segmentation step. However, doing this would require a refinement method to refine the generated road network, for doing this we develop a novel Graph neural network based approach to refine any generated road network. We use this analyze the road networks generate in different areas of urban settings such as square grid cities, bridges over water bodies and also in scenarios where there is significant green cover creating occlusions. The results are compared quantitatively to the previous baselines and we achieve comparable performance to the state of art. Our method can be used with any of generated road networks to finally refine it to be more accurate and remove incorrect topologies.