Options
April 16, 2024
Master Thesis
Title
Self-localization of roadside radar unit based on vehicle trajectory
Abstract
Roadside sensor unit plays an important role in the Intelligent transportation system (ITS). Millimeter wave (mmWave) radar has become one of the most popular sensors in roadside units, due to its unique advantage of long-range detection and stability in all weather, etc. A key step before further traffic analysis or data communication in ITS is the localization of the radar sensor. The localization task for the mobile sensor units is more important, as it is necessary every time it is deployed. In this thesis, a map-based data processing pipeline is presented and implemented to estimate both the position and orientation of roadside radar in a world coordinate. It solves the localization task by matching two representations of roads. A street map is one of them. It could provide topological information on lanes. Turning behavior is the attribute that this thesis focuses on. By utilizing a street map, it is possible to extract road points from aerial laser scans and assign labels to these points. Besides, the pipeline uses the internal measurements obtained from the sensor. Moving objects detected can be considered as the existence of roads. With a sufficient duration of sampling, these moving objects can reach any accessible point in the field of view of the sensor. The second road representation is then extracted. Tracking moving objects helps to learn the turning behavior of each trajectory, and to assign labels to radar measurement points. Point cloud registration is applied to find the transformation matrix between the above two point sets. Semantic information is introduced in this step to calculate the weights of point pairs, enhancing the robustness.
The performance of the pipeline is evaluated in both simulator and real-world experiments. The results indicate that labeling point clouds based on the results of object tracking is reliable. The proposed method can achieve positional accuracy below 1 meter and angular accuracy below 0.5 degrees. Compared to methods that do not use semantic information, there is an improvement in the result.
The performance of the pipeline is evaluated in both simulator and real-world experiments. The results indicate that labeling point clouds based on the results of object tracking is reliable. The proposed method can achieve positional accuracy below 1 meter and angular accuracy below 0.5 degrees. Compared to methods that do not use semantic information, there is an improvement in the result.
Thesis Note
Stuttgart, Univ., Master Thesis, 2024
Author(s)
Advisor(s)