Global pose estimation is a fundamental step in the advancement of autonomous driving technology. While Global Navigation Satellite Systems (GNSS) are conventionally viewed as the instinctive solution, their reliability is frequently compromised in urban environments due to signal blockage and multipath interference. Consequently, this project addresses the critical need for a robust real-time positioning system capable of achieving high-precision accuracy (<10 cm) within densely populated urban settings. A defining characteristic of our methodology is the strategic departure from High Definition (HD) maps, which are costly and computationally intensive. Instead, we leverage Level of Detail 2 (LoD2) based abstract maps, which offer a significantly more cost-effective alternative. By integrating LoD2 models with Digital Terrain Models (DTM) derived from open data, we construct a comprehensive abstract reference map. Uniquely, we turn the sparsity of these abstract maps into a computational advantage; by employing fast and efficient machine learning algorithms that exploit this reduced data density to maximize processing speed. To facilitate this positioning strategy, we have developed a state-of-the-art data acquisition system. The sensor suite incorporates LiDARs, cameras, and Inertial Measurement Units (IMU) to capture high-fidelity environmental information. These multimodal inputs are utilized to identify unique landmarks for geo-referencing the vehicle against the abstract reference map. To ensure the precision and consistency of the geo-referencing process, we partnered with Quality Match GmbH. Together, we developed a machine learning algorithm capable of accurately detecting landmarks and associating them with their map counterparts. To assess our approach, and given the complexity of this task, we constructed unique databases based on real-life measurements and utilized a human-in-the-loop data annotation pipeline, to produce the “ground-truth” data needed for tuning and benchmarking our algorithms. Finally, specialized filters are employed to smooth and integrate the multimodal measurements, ensuring continuous and stable pose estimation.

This mFUND project is funded by the BMDV (Bundesministerium für Digitales und Verkehr).