Authors:
Lekshmi Kalinathan, K. Anusha, K. Devi, K. Divya
Addresses:
School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu, India.
Abstract:
Reliable object detection continues to be the foundation of the safety features for self-driving cars. One of the biggest challenges, though, is that different weather phenomena, such as fog, snow, and darkness, not only affect the driver but also the on-board sensors and, therefore, the performance of perception systems in terms of object detection and recognition. This paper addresses the research gap in combining detection accuracy, robustness to various adverse situations, and explainability for autonomous driving object detection tasks. The Faster R-CNN backbone ResNet-50 has been tested on corrupted datasets: COCO-C, Pascal-C and Cityscapes-C, using the image segmentation techniques K-means, GrabCut, SLIC, Watershed and U-Net. SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) have been employed for model interpretability purposes, and fairness has been addressed through two fairness metrics, namely Demographic Parity and Equalized Odds, to highlight and eventually remove prediction bias. The experiments revealed detection accuracies of 97.2% in clear conditions and 89.0% under adverse conditions, with U-Net performing best at 90.0% even in the most challenging scenarios. The combined use of explainable AI techniques and fairness assessment can improve model transparency, bias detection, and overall system reliability. Overall, this work provides a framework for trustworthy, robust, and equitable computer perception systems for self-driving cars operating in real-world environments.
Keywords: Explainable Ai (Xai); Adverse Weather; Object Detection; Image Segmentation; Bias Mitigation; Autonomous Cars; Fairness Metrics; Shapley Additive Explanations.
Received on: 09/02/2025, Revised on: 08/04/2025, Accepted on: 17/06/2025, Published on: 05/06/2026
DOI: 10.64091/ATICL.2026.000329
AVE Trends in Intelligent Computer Letters, 2026 Vol. 2 No. 2 , Pages: 111-121