Enhancing Vehicle Detection in Adverse Weather with a Weighted YOLO Ensemble Approach and Explainable AI

Adverse weather conditions significantly degrade the performance of vehicle detection models, posing challenges for autonomous driving and traffic surveillance systems. This study enhances vehicle detection in challenging weather environments by proposing an ensemble object detection approach and integrating Explainable AI (XAI) for trustworthiness assessment. This study employs the DAWN dataset—designed to simulate adverse weather scenarios, to assess the performance of three state-of-the-art YOLO models: YOLOvl0, YOLOvll, and YOLOvI2. Our proposed ensemble strategy with weighted box fusion, combining YOLOvl1 and YOLOvI2, improves the mean Average Precision (mAP@0.5) from 0.631 to 0.6479, demonstrating the effectiveness of the proposed ensemble. In addition, EigenCAM-based explainable AI (XAI) methods are utilized to interpret and verify model predictions, ensuring greater transparency and trustworthiness in detection results. The results highlight the effectiveness of ensemble learning for vehicle detection in extreme weather conditions and underscore the importance of XAI in assessing model trustworthiness.