h1

TY  - THES
AU  - Wenning, Marius Julian
TI  - Kontextabhängig Anomaliedetektion zur visuellen Hinderniserkennung für automatisierte PKW im End-of-Line-Bereich; 1. Auflage
VL  - 37/2022
PB  - RWTH Aachen University
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2022-11143
SN  - 978-3-98555-121-7
T2  - Ergebnisse aus der Produktionstechnik
SP  - XIII, 154 Seiten : Illustrationen, Diagramme
PY  - 2022
N1  - Druckausgabe: 2022. - Auch veröffentlicht auf dem Publikationsserver der RWTH Aachen University 2023
N1  - Dissertation, RWTH Aachen University, 2022
AB  - For automated driving on company premises of car manufacturers, the cars’ protective device needs to be implemented using standard sensor equipment. Built-in mono cameras cannot be used for obstacle detection yet, as existing computer vision algorithms do not provide reliable object detection. The analysis of the computer vision algorithms leads to a specification of requirements for an innovative obstacle detection algorithm. The method of anomaly detection shows inherent advantages compared to existing obstacle detection algorithms. The state of the art includes promising methods. However, anomaly detection has not been tested in the use case of vehicle automation so far. Therefore, the objective of this thesis is to design a suitable anomaly detection algorithm and to test it in the use case of factory-automated cars. To this end, two data sets are developed. Data from a simulated factory environment is used to develop a performant anomaly detection and to test it against state-of-the-art benchmark algorithms in use case-specific test cases. Real data enables validation of the simulation results and proves the algorithm’s practicality. A detailed analysis reveals implicit design rules that should be considered in the development of an anomaly detection for obstacle detection. The classification quality can be improved by taking into account the spacial and temporal context of the processed images. In comparison with state-of-the-art algorithms, the anomaly detection shows a competitive classification quality at only a fraction of required training data. In scenes with high complexity due to illumination, dirt or smoke, the anomaly detection is more robust than the semantic segmentation and the depth estimation. Real world experiments confirm the simulation results and prove practical applicability.
LB  - PUB:(DE-HGF)11 ; PUB:(DE-HGF)3
DO  - DOI:10.18154/RWTH-2022-11143
UR  - https://publications.rwth-aachen.de/record/856819
ER  -