Year: 2026
Type: dissertation
Abstract: Camera-based pedestrian detection utilizing deep neural networks (DNNs) represents a challenging milestone for automated driving. Changes in the perceived environment of an automated vehicle can expose functional insufficiencies of embedded DNNs, potentially resulting in harmful behavior. Accomplishing safe pedestrian detection means providing a structured safety argumentation based on verifiable evidence for the correct functionality of DNNs. Hence, this work proposes a novel methodology for machine learning engineering to support a safety argumentation that addresses functional insufficiencies besides the systematic evaluation of DNNs. The methodology presents verifiable evidence with detection metrics designed for automated driving and assurance metrics to quantify the safety impact of inherent interpretability of DNNs. This work demonstrates that leveraging inherent interpretability enables a safety argumentation. Modifying the DNN architecture and loss formulation to establish interpretable reasoning enforces a structured latent space that inherently provides evidence. The proposed DNN for concept-based pedestrian detection uses semantic concepts such as body parts to structure the latent space by enforcing a high intra-class concentration and inter-class separation. Ultimately, predictions become interpretable since detected body parts have significant similarities to semantic concept vectors. Moreover, structuring the latent space improves the generalization capabilities of DNNs. Despite training with unlabeled real-world data and labeled synthetic data, the DNN for concept-based pedestrian detection adapts successfully to the real world. The development of baseline methods and progressive advancement to novel interpretable methods with state-of-the-art performance conforms to a generic DNN for pedestrian detection. This counteracts the prevalent lack of transparency in pedestrian detection and ensures valid benchmarking. Meanwhile, the systematic evaluation of methods focuses on safety-critical detection errors from the perspective of automated driving systems. The presented experimental results demonstrate the substantial contribution of this work, highlighting the significance of inherent interpretability for safe pedestrian detection with DNNs.
Author: Feifel, Patrick
Institution: Carl von Ossietzky Universität Oldenburg
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7172240106
