Eintrag in der Universitätsbibliographie der TU Chemnitz
Volltext zugänglich unter
URN: urn:nbn:de:bsz:ch1-qucosa2-1003063
Tilly, Julius Frederik
Wanielik, Gerd (Prof.) ; Mößner, Klaus (Prof.) ; Thomanek, Jan (Prof.) (Gutachter)
Road User Detection with Automotive Polarimetric Radars
Kurzfassung in englisch
With the availability of the first automotive polarimetric 77Ghz radars that are small enough to be integrated in cars, a new information dimension given by measuring the polarization of received electromagnetic waves and knowing the polarization of transmitted waves can be exploited for automotive applications. In this thesis, the benefits of polarimetric information for road user classification and detection in the automotive context are investigated. For this purpose, a variation of real-world traffic scenarios has been recorded and annotated to form a large dataset that allows the evaluation of the potential of polarimetric radars. A general impression of how different road users appear to polarimetric radars is obtained by measuring and analyzing their polarimetric signatures. The signatures show that, especially for cars and bicycles, the scattering behavior changes depending on the aspect angle under which they are observed, potentially allowing for better classification and orientation estimation of these objects. Using sparse radar point clouds in which each point has additional polarimetric information, a classical random forest classifier is applied with hand-crafted features computed on clusters of radar points. The classification performance with additional polarimetric features is significantly improved compared to the performance without the availability of polarimetric features. In the next step, a machine learning model is used to learn features directly on the polarimetric radar point clouds. Again, it is observed that the scattering information measured with polarimetric radars improves the class and instance segmentation performance compared to when only conventional radar information is available. An additional importance analysis shows that the polarimetric information is more valuable than the RCS. Sparse radar point clouds are extracted from denser underlying data that contains more information, so potentially valuable data can be lost in the radar point cloud generation step. Therefore, a machine learning model using this layer of data, also referred to as a 'radar data cube', is presented and evaluated for improving road user detection. The model predicts oriented bounding boxes with class labels. Experiments with the model show that the raw data layer enriched with polarimetric scattering data provides significantly better detection and orientation results than without this data.
| Universität: | Technische Universität Chemnitz | |
| Institut: | Professur Nachrichtentechnik | |
| Fakultät: | Fakultät für Elektrotechnik und Informationstechnik | |
| Dokumentart: | Dissertation | |
| Betreuer: | Wanielik, Gerd (Prof.) | |
| DOI: | doi:10.60687/2025-0204 | |
| SWD-Schlagwörter: | Radar , Polarimetrie , Objekterkennung | |
| Freie Schlagwörter (Englisch): | Radar , Polarimetric , Road User Detection , Machine Learning | |
| DDC-Sachgruppe: | 621.3848 | |
| Sprache: | englisch | |
| Tag der mündlichen Prüfung | 16.10.2025 | |
| OA-Lizenz | CC BY-SA 4.0 |
