Matching heterogeneous sensing pipelines to digital maps for ego-vehicle localization

In this thesis, we present a probabilistic framework for ego-vehicle localization called Road Layout Estimation framework. The main contribution to the vehicle localization problem is the synergistic exploitation of heterogeneous sensing pipelines, as well as their matching with respect to the OpenStreetMap service. The approach is validated in different ways by exploiting different visual clues. Firstly by using the road graph provided by the OpenStreetMap service, then exploiting high-level features like intersections between roads, buildings façades, and other road features. Regarding the effectiveness of the road-graph exploitation, its is proven by achieving real-time computation with state-of-the-art results on a set of ten not trivial runs from the KITTI dataset, including both urban/residential and highway/road scenarios. Moreover, a probabilistic approach for detecting and classifying urban road intersections from a moving vehicle is presented. The approach is based on images from an on-board stereo rig. It relies on the detection of the road ground plane on one side, and on a pixel-level classification of the observed scene on the other. The two processing pipelines are then integrated and the parameters of the road components, i.e., the intersection geometry, are inferred. As opposed to other state-of-the-art off-line methods, which require processing of the whole video sequence up to when the vehicle is inside the intersection, our approach integrates the image data by means of an on-line procedure. The experiments have been performed on the well-known KITTI datasets as well, allowing the community to perform future comparisons. Besides the pure road interpretation schemes, in this work we also present a technique that takes advantage of detected building façades and OpenStreetMaps building data to improve the localization of an autonomous vehicle driving in an urban scenario. The proposed approach also leverages images from the stereo rig mounted on the vehicle to produce a mathematical representation of the buildings' façades within the field of view. This representation is matched against the outlines of the surrounding buildings as they are available on OpenStreetMaps. All the retrieved features are fed into our probabilistic framework, in order to produce an accurate lane-level localization of the vehicle in urban contexts. Finally, as to achieve a lane-level localization also in highway scenarios, we propose two methods that allow the framework to leverage the lane number and the road width. The proposed approaches have been tested under real traffic conditions, showing satisfactory performances with respect to the map-matching-only settings and compensating the noisy measures of a basic line detector.