Perception multi-capteurs avec des cartes vectorielles pour la localisation des véhicules autonomes

For autonomous vehicles, it is crucial to find localization solutions that meet the requirements of navigation tasks. For safety reasons, the localization system must provide accurate and reliable pose estimates, with a high availability and a low latency. To this end, multisensor data fusion techniques are employed. They generally combine GNSS receivers with proprioceptive sensors that provide vehicle kinematics and dynamics. In this PhD thesis, we particularly focus on the use of additional exteroceptive sensors such as lidars and cameras which can provide measurements on features georeferenced in maps. These sensors help overcome GNSS limitations in complex environments like urban areas, enabling lane-level positioning. When using perception sensors, various methods can provide localization information. A common approach in robotics is implementing Simultaneous Localization and Mapping (SLAM) with GNSS constraints. This builds a local map online from raw sensor data, which can then be used for re-localization with the same sensors. Using accurate prior maps offers an interesting alternative enabling immediate localization upon entering a new area without the need to create a map anew. In this PhD thesis, we consider high-definition (HD) vector maps containing georeferenced road features represented as points or polylines. They encompass a wide range of physical elements essential for navigation such as traffic signs or road markings. The main goal of this thesis is to leverage all the potential offered by HD vector maps to improve localization, through a perception system whose performance has been optimized for this purpose. As a case study, our research focuses on detecting pole-like features, which are commonly found throughout road environments and georeferenced in HD maps. More specifically, we present camera and lidar perception approaches enabled by a map-based automatic annotation method. This method can annotate any kind of mapped poles. To enhance annotation accuracy and completeness, we integrate this primary method with additional automatic annotation sources. We train detectors to identify pole bases in camera images and from clusters of lidar points. This integration ensures that detected poles conform to the definitions used in the HD map. Finally, these detection approaches are integrated in a multi-sensor fusion system to assess their benefits for a localization system. Given the approaches explored, the thesis heavily relies on experimental data collected from vehicles equipped with lidar sensors and cameras. This work was carried out in synergy with the European project ERASMO, which aimed to develop a highly accurate and reliable localization system for autonomous vehicles.