As autonomous vehicles become increasingly integrated into modern transportation systems, the need for robust, reliable, and interpretable perception systems has become critical. A central component of this perception stack is traffic sign recognition, which enables autonomous vehicles to interpret regulatory and warning cues essential for safe navigation. Although deep learning models such as convolutional neural networks have achieved high accuracy in controlled settings, they can exhibit less robustness under certain real-world conditions, particularly when faced with ambiguous visual contexts or adversarial inputs.

Moreover, the limited interpretability of learned representations presents challenges for safety-critical decision-making, regulatory approval, and public trust, motivating investigation of methods that can complement neural perception with human-readable reasoning.

This thesis proposes a suite of neurosymbolic methods that integrate deep learning with Inductive Logic Programming (ILP) to investigate explainable and robust traffic sign recognition. The thesis investigates the potential of ILP-based symbolic reasoning to complement current deep-learning approaches in addressing four related challenges: interpretability, robustness to adversarial inputs, data efficiency, and context-sensitive disambiguation. Here, \emph{trustworthy traffic-sign perception} is used in a component-level sense to refer to human-interpretable reasoning, robustness to selected perturbations, and logically or semantically consistent prediction under the evaluated conditions. It does not imply comprehensive autonomous-driving safety, formal verification, certification, or deployment readiness.

The research is organised around four core contributions, some of which build on published peer-reviewed papers, while the others are currently under review for publication.

First, an ILP-based traffic sign classifier is developed that uses high-level visual attributes, including shape, colour, and text, for stop sign detection. This approach demonstrates strong robustness against physical adversarial attacks, including Robust Physical Perturbation and Adversarial Camouflage, while requiring only minimal training data.

Second, a neurosymbolic logic-infused deep learning framework is introduced. This method transforms ILP-derived rules into logical consistency constraints and integrates them into the loss function of a deep neural network. The resulting model improves adversarial robustness while preserving predictive performance, providing a novel mechanism for embedding symbolic reasoning into statistical learning.

Third, a graph-based framework is proposed in which symbolic relationships extracted through ILP are encoded as relational graphs and processed by a graph neural network. In the evaluated setting, this ILP-GNN framework enriches relational structure and improves classification robustness under the tested perturbations and environmental conditions.

Finally, DL4LL-Scene is presented as a neurosymbolic pipeline that combines object detection with symbolic reasoning to resolve visual ambiguity. It is particularly effective in distinguishing genuine stop signs from visually similar but functionally irrelevant instances, such as signs appearing on billboards or clothing. This system achieves high accuracy using only eight training examples and substantially outperforms both standard deep learning baselines and vision-language models in low-data settings.

Taken together, these contributions provide evidence that integrating symbolic logic with deep learning can improve traffic sign recognition in several complementary respects. Across the thesis, the proposed methods strengthen robustness under adversarial and ambiguous conditions, improve data efficiency, and produce transparent, human-readable rules that support interpretability and inspection.

More broadly, the thesis argues that autonomous-vehicle perception can benefit from models in which learning is complemented by explicit structure, semantic constraints, and contextual reasoning.

In this way, it provides a foundation for perception systems better suited to the visual complexity, safety requirements, and accountability demands of real-world deployment.