Mori, Ken Thaddäus (2024)
Defining object detection requirements for safe automated driving.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00026622
Dissertation, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
Automated Driving (AD) requires reliable environment perception to ensure safety. One common perception task is 3D object detection, which aims at perceiving location and attributes of dynamic objects. The evaluation of object detection typically relies on datasets. While these datasets provide evaluation metrics, they fail to connect detection results to safety outcomes such as accidents. Therefore, there is a lack of clear requirements for object detection which consider the safety of the driving task.
Accordingly, the objective of this work is to identify requirements for 3D object detection which consider safety. Furthermore, it is desirable to obtain requirements which are interpretable. Therefore, the overall research question regarding requirements for safe object detection is decomposed. Requirements for the three interpretable aspects of classification, relevance and attributes of objects are treated separately. Finally, the last objective is to provide a method to evaluate and validate the different requirements for each of these aspects.
The methodology of this work first identifies common principles which further specify the overall objective of requirements for safety. The four principles are interpretability, legal requirements, safety requirements and the human baseline. Applying these principles allows developing different methods for the aspects classification, relevance and attributes, respectively. In this work, the required object categories, criteria for relevance and attribute requirements are successfully identified. In addition, a novel validation method based on a motion prediction leveraging a deep neural network is presented. Applying this validation method to the requirements proposed in this work is successful, thus supporting the results. Remaining limitations of the proposed methodology including the available data and algorithms are identified and discussed. Furthermore, the implications of the novel requirements on datasets, algorithms and sensor setups for 3D object detection are considered.
The overall methodology presents a two-pronged approach to requirement definition. Firstly, simple and interpretable requirements are developed based on a safety argumentation. These requirements are then additionally substantiated by the validation method, which relies upon a deep neural network. The results thus provide the requirements, which are required to test and validate object detectors in the context of safety for the task of AD. The author hopes that the requirements encourage the explicit evaluation and improvement of safety for future object detection.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2024 | ||||
Autor(en): | Mori, Ken Thaddäus | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Defining object detection requirements for safe automated driving | ||||
Sprache: | Englisch | ||||
Referenten: | Peters, Prof. Dr. Steven ; Kuijper, Prof. Dr. Arjan | ||||
Publikationsjahr: | 7 März 2024 | ||||
Ort: | Darmstadt | ||||
Kollation: | XIII, 170 Seiten | ||||
Datum der mündlichen Prüfung: | 30 Januar 2024 | ||||
DOI: | 10.26083/tuprints-00026622 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/26622 | ||||
Kurzbeschreibung (Abstract): | Automated Driving (AD) requires reliable environment perception to ensure safety. One common perception task is 3D object detection, which aims at perceiving location and attributes of dynamic objects. The evaluation of object detection typically relies on datasets. While these datasets provide evaluation metrics, they fail to connect detection results to safety outcomes such as accidents. Therefore, there is a lack of clear requirements for object detection which consider the safety of the driving task. Accordingly, the objective of this work is to identify requirements for 3D object detection which consider safety. Furthermore, it is desirable to obtain requirements which are interpretable. Therefore, the overall research question regarding requirements for safe object detection is decomposed. Requirements for the three interpretable aspects of classification, relevance and attributes of objects are treated separately. Finally, the last objective is to provide a method to evaluate and validate the different requirements for each of these aspects. The methodology of this work first identifies common principles which further specify the overall objective of requirements for safety. The four principles are interpretability, legal requirements, safety requirements and the human baseline. Applying these principles allows developing different methods for the aspects classification, relevance and attributes, respectively. In this work, the required object categories, criteria for relevance and attribute requirements are successfully identified. In addition, a novel validation method based on a motion prediction leveraging a deep neural network is presented. Applying this validation method to the requirements proposed in this work is successful, thus supporting the results. Remaining limitations of the proposed methodology including the available data and algorithms are identified and discussed. Furthermore, the implications of the novel requirements on datasets, algorithms and sensor setups for 3D object detection are considered. The overall methodology presents a two-pronged approach to requirement definition. Firstly, simple and interpretable requirements are developed based on a safety argumentation. These requirements are then additionally substantiated by the validation method, which relies upon a deep neural network. The results thus provide the requirements, which are required to test and validate object detectors in the context of safety for the task of AD. The author hopes that the requirements encourage the explicit evaluation and improvement of safety for future object detection. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
Freie Schlagworte: | requirements, testing, object detection, automated driving | ||||
Status: | Verlagsversion | ||||
URN: | urn:nbn:de:tuda-tuprints-266228 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften und Maschinenbau | ||||
Fachbereich(e)/-gebiet(e): | 16 Fachbereich Maschinenbau 16 Fachbereich Maschinenbau > Fachgebiet Fahrzeugtechnik (FZD) 16 Fachbereich Maschinenbau > Fachgebiet Fahrzeugtechnik (FZD) > Sicherheit |
||||
Hinterlegungsdatum: | 07 Mär 2024 12:43 | ||||
Letzte Änderung: | 08 Mär 2024 11:29 | ||||
PPN: | |||||
Referenten: | Peters, Prof. Dr. Steven ; Kuijper, Prof. Dr. Arjan | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 30 Januar 2024 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |