Identification of Drivable Road Area from Orthophotos Using a Convolutional Neural Network

This paper addresses the subject of road area determination from photographic images, focusing on the orthoframes that depict gravel or paved roads without road markings. The proposed solution uses the well-known grabCut algorithm to annotate the training images and a U-Net-based convolutional segmentation network to extract the road area. Apart from un-usually shaped roads (e.g. crossings) and unusual circumstances (e.g. hard shadows), the proposed heavily automated solution shows a very good performance (over 96% intersection-over-union on validation images) on the majority of images presented to it.