Casado Herráez, Daniel: Localization and Mapping for Autonomous Vehicles Using Radar. - Bonn, 2026. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-90279
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-90279
@phdthesis{handle:20.500.11811/14181,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-90279,
author = {{Daniel Casado Herráez}},
title = {Localization and Mapping for Autonomous Vehicles Using Radar},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2026,
month = jun,
note = {Human error remains the leading cause of road fatalities. By minimizing human intervention behind the wheel, autonomous vehicles aim to reduce traffic accidents and provide a safer means of transportation. These vehicles must know their precise location to navigate reliably within an environment. Without accurate localization, trajectory planning and obstacle avoidance systems will have an incorrect estimate of the vehicle's state, potentially leading to a collision.
Although localization is commonly achieved using global navigation satellite systems, satellite visibility is limited in areas with tall skyscrapers and indoor settings such as tunnels and parking garages. Therefore, autonomous cars also need the ability to estimate their own position and orientation solely leveraging local sensing and previously recorded maps of the environment, without relying on satellite availability. Onboard localization and mapping are typically achieved using cameras or LiDAR sensors. Cameras, however, are affected by low light and adverse weather conditions, and the performance of LiDARs degrades in challenging weather scenarios such as heavy rain, fog, and snow. On the contrary, radar sensors are resilient to environmental conditions and are already integrated into consumer vehicles today. Additionally, automotive radars also provide velocity information and the radar cross section of the measured targets, which can be leveraged to enhance localization performance. As a downside, these sensors produce a sparse and noisy point cloud compared to LiDAR data, resulting in challenges for sensor-based pose estimation. Although some works on radar localization and mapping already exist, these methods face limitations at high noise levels and with the limited number of points present in radar scans. Moreover, they do not exploit the specific properties of radar sensing.
In this thesis, we push the boundaries of radar localization and mapping for autonomous vehicles by introducing novel and impactful techniques specifically designed for automotive radar sensors. We begin by estimating the pose of the vehicle and creating a map of the environment over time. To achieve this, we propose novel algorithms that exploit the sparsity of automotive radar scans and their associated Doppler velocities, yielding accurate pose estimates. Due to the low vertical resolution of automotive radars, we observe that localization performance degrades when there are changes in elevation. To address this limitation, we propose an approach that exploits the radar properties to extract road features, enhancing accuracy during slope variations. We further improve pose estimation performance by introducing an additional inertial measurement unit into the system. While these strategies achieve high short-term accuracy, errors can accumulate over long trajectories, leading to inconsistencies in large-scale maps. Therefore, we develop a radar-specific module that performs place recognition to identify previously visited locations. We integrate this information within a novel radar-inertial simultaneous localization and mapping system, achieving accurate online pose estimation and producing consistent maps. To record and combine different regions of the drivable area at various points in time, we propose a multi-session mapping system that merges multiple maps. Our proposed approach also accounts for temporal changes in the environment, such as parked cars and new constructions. This is crucial for maintaining a long-term, accurate representation of the environment. Moreover, our technique accurately localizes within previously recorded radar maps with minimal accumulated error.
Our proposed estimation approaches achieve state-of-the-art results on automotive radar data. Additionally, we present one of the first place recognition methods explicitly tailored for automotive radar sensors, capable of generalizing to different datasets. We also present one of the first long-term map construction and localization techniques for automotive radars. All of our approaches have been evaluated on public datasets and have been published in peer-reviewed conferences and journals. Some of our methods have also been open-sourced, collectively enhancing the capabilities of safe localization and mapping for autonomous vehicles using radar.},
url = {https://hdl.handle.net/20.500.11811/14181}
}
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-90279,
author = {{Daniel Casado Herráez}},
title = {Localization and Mapping for Autonomous Vehicles Using Radar},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2026,
month = jun,
note = {Human error remains the leading cause of road fatalities. By minimizing human intervention behind the wheel, autonomous vehicles aim to reduce traffic accidents and provide a safer means of transportation. These vehicles must know their precise location to navigate reliably within an environment. Without accurate localization, trajectory planning and obstacle avoidance systems will have an incorrect estimate of the vehicle's state, potentially leading to a collision.
Although localization is commonly achieved using global navigation satellite systems, satellite visibility is limited in areas with tall skyscrapers and indoor settings such as tunnels and parking garages. Therefore, autonomous cars also need the ability to estimate their own position and orientation solely leveraging local sensing and previously recorded maps of the environment, without relying on satellite availability. Onboard localization and mapping are typically achieved using cameras or LiDAR sensors. Cameras, however, are affected by low light and adverse weather conditions, and the performance of LiDARs degrades in challenging weather scenarios such as heavy rain, fog, and snow. On the contrary, radar sensors are resilient to environmental conditions and are already integrated into consumer vehicles today. Additionally, automotive radars also provide velocity information and the radar cross section of the measured targets, which can be leveraged to enhance localization performance. As a downside, these sensors produce a sparse and noisy point cloud compared to LiDAR data, resulting in challenges for sensor-based pose estimation. Although some works on radar localization and mapping already exist, these methods face limitations at high noise levels and with the limited number of points present in radar scans. Moreover, they do not exploit the specific properties of radar sensing.
In this thesis, we push the boundaries of radar localization and mapping for autonomous vehicles by introducing novel and impactful techniques specifically designed for automotive radar sensors. We begin by estimating the pose of the vehicle and creating a map of the environment over time. To achieve this, we propose novel algorithms that exploit the sparsity of automotive radar scans and their associated Doppler velocities, yielding accurate pose estimates. Due to the low vertical resolution of automotive radars, we observe that localization performance degrades when there are changes in elevation. To address this limitation, we propose an approach that exploits the radar properties to extract road features, enhancing accuracy during slope variations. We further improve pose estimation performance by introducing an additional inertial measurement unit into the system. While these strategies achieve high short-term accuracy, errors can accumulate over long trajectories, leading to inconsistencies in large-scale maps. Therefore, we develop a radar-specific module that performs place recognition to identify previously visited locations. We integrate this information within a novel radar-inertial simultaneous localization and mapping system, achieving accurate online pose estimation and producing consistent maps. To record and combine different regions of the drivable area at various points in time, we propose a multi-session mapping system that merges multiple maps. Our proposed approach also accounts for temporal changes in the environment, such as parked cars and new constructions. This is crucial for maintaining a long-term, accurate representation of the environment. Moreover, our technique accurately localizes within previously recorded radar maps with minimal accumulated error.
Our proposed estimation approaches achieve state-of-the-art results on automotive radar data. Additionally, we present one of the first place recognition methods explicitly tailored for automotive radar sensors, capable of generalizing to different datasets. We also present one of the first long-term map construction and localization techniques for automotive radars. All of our approaches have been evaluated on public datasets and have been published in peer-reviewed conferences and journals. Some of our methods have also been open-sourced, collectively enhancing the capabilities of safe localization and mapping for autonomous vehicles using radar.},
url = {https://hdl.handle.net/20.500.11811/14181}
}





