Semantic mapping of road scenes

The problem of understanding road scenes has been on the fore-front in the computer vision community for the last couple of years. This enables autonomous systems to navigate and understand the surroundings in which it operates. It involves reconstructing the scene and estimating the objects present in it, such as ‘vehicles’, ‘road’, ‘pavements’ and ‘buildings’. This thesis focusses on these aspects and proposes solutions to address them. First, we propose a solution to generate a dense semantic map from multiple street-level images. This map can be imagined as the bird’s eye view of the region with associated semantic labels for ten’s of kilometres of street level data. We generate the overhead semantic view from street level images. This is in contrast to existing approaches using satellite/overhead imagery for classification of urban region, allowing us to produce a detailed semantic map for a large scale urban area. Then we describe a method to perform large scale dense 3D reconstruction of road scenes with associated semantic labels. Our method fuses the depth-maps in an online fashion, generated from the stereo pairs across time into a global 3D volume, in order to accommodate arbitrarily long image sequences. The object class labels estimated from the street level stereo image sequence are used to annotate the reconstructed volume. Then we exploit the scene structure in object class labelling by performing inference over the meshed representation of the scene. By performing labelling over the mesh we solve two issues: Firstly, images often have redundant information with multiple images describing the same scene. Solving these images separately is slow, where our method is approximately a magnitude faster in the inference stage compared to normal inference in the image domain. Secondly, often multiple images, even though they describe the same scene result in inconsistent labelling. By solving a single mesh, we remove the inconsistency of labelling across the images. Also our mesh based labelling takes into account of the object layout in the scene, which is often ambiguous in the image domain, thereby increasing the accuracy of object labelling. Finally, we perform labelling and structure computation through a hierarchical robust P Markov Random Field defined on voxels and super-voxels given by an octree. This allows us to infer the 3D structure and the object-class labels in a principled manner, through bounded approximate minimisation of a well defined and studied energy functional. In this thesis, we also introduce two object labelled datasets created from real world data. The 15 kilometre Yotta Labelled dataset consists of 8,000 images per camera view of the roadways of the United Kingdom with a subset of them annotated with object class labels and the second dataset is comprised of ground truth object labels for the publicly available KITTI dataset. Both the datasets are available publicly and we hope will be helpful to the vision research community.