Inferring Road Boundaries Through and Despite Traffic

This paper is about the detection and inference of road boundaries from mono-images. Our goal is to trace out, in an image, the projection of road boundaries irrespective of whether or not the boundary is actually visible. Large scale occlusion by vehicles prohibits direct approaches - many scenes present 100% occlusion and so we must infer the boundary location using scene context. Such a problem is well suited to CNN derived approaches but the sinuous structure of a hidden narrow continuous curve running through the image presents challenges for conventional NN-architectures. We approach this as a coupled, two class detection problem-solving for occluded and non-occluded curve partitions with a continuity constraint. Our network output is in a hybrid discrete-continuous form which we interpret as measurements of segments of the true road boundary. These measurements are passed to a model selection stage which associates measurements to minimal number of a-priori unknown set of geometric primitives (cubic curves) representing road boundaries. We present a semi-supervised method which leverages a visual localisation to generate 25 thousand labelled images for training and testing - the results of which are presented in the conclusion of the paper.