Design an‎d Implementation of Explainable Artificial Intelligence for Self-Driving Vehicles

10/22/2026 12:00:00 AM

Despite significant advances in deep learning for autonomous driving systems, critical challenges remain in transparency, interpretability, an‎d user acceptance. The inability to explain the decision-making behavior of deep learning models limits their practical implementation in various real-world scenarios, which is particularly critical in safety-critical decision-making networks where incorrect outcomes may lead to dangerous road accidents. This thesis presents an innovative approach to developing interpretable autonomous driving systems that includes three main contributions. First, a model is introduced that combines convolutional block attention modules with deep cross-attention, integrated with an advanced textual decision-making an‎d explanation model. This combination results in a more accurate an‎d comprehensive system for interpretability. The second contribution of this research is the presentation of a multi-task learning approach for simultaneous prediction of various driving behaviors an‎d their underlying reasoning. Additionally, a comprehensive analysis of confidence threshold sensitivity analysis for multi-task classification models is conducted, demonstrating that traditional fixed-threshold approaches are not optimal for multi-task scenarios. Furthermore, the IUST-XAI-AD dataset is introduced, comprising 958 images with detailed human annotations for driving actions an‎d explanatory reasoning. This dataset fills an important gap in domain-specific eva‎luation benchmarks for Persian driving contexts an‎d provides a more challenging testing environment compared to existing datasets. Quantitative eva‎luations are performed using stan‎dard F1 metrics on the BDD-OIA, nu-AR, an‎d the newly introduced IUST-XAI-AD datasets. Experimental results demonstrate that the proposed technique with 54.06% an‎d 36.65% accuracy in overall F1 score an‎d mean F1 score respectively on the BDD-OIA dataset outperforms existing models an‎d shows significant improvement in interpretability. Confidence threshold sensitivity analysis significantly enhances model performance, while the introduction of the IUST-XAI-AD dataset reveals important insights about cross-cultural driving behavior patterns an‎d demonstrates the robustness an‎d generalizability of the proposed model compared to existing models. This research contributes to the development of more trustworthy an‎d interpretable autonomous driving systems by providing methodological advances an‎d practical eva‎luation tools, establishing a foundation for the future of interpretable autonomous vehicles.

يادگيري چندوظيفه , تحليل حساسيت آستانه اطمينان , وسايل نقليه خودران

multi-task learning , confidence threshold sensitivity analysis , autonomous vehicles

Maryam Sadat Hoseini Azad

Dr Shahriar Baradaran Shokohi