A hybrid framework for heterogeneous object detection amidst diverse and adverse weather conditions employing Enhanced-DARTS

Abstract

Autonomous vehicles face significant challenges in accurately identifying vehicles, objects, and traffic signals under adverse weather conditions and poor lighting. To address these issues, we introduce a novel detection system utilizing automatic white balance techniques, specifically the Adaptive Retinex algorithm, to restore visibility and enhance color. This system is integrated with a Faster R-CNN framework enhanced by non-maximum suppression to improve the accuracy of object detection. Employing a combination of three datasets—Dawn, MCWRD, and Indian Roads Dataset (IRD)—our method includes over 6000 augmented images representing diverse environmental conditions. We also implement an optimized version of Differentiable ARchiTecture Search (DARTS) to dynamically fine-tune the architectural parameters of our detection model. This approach has successfully achieved a detection accuracy of 97.43% with a minimal loss rate, demonstrating significant potential for enhancing navigation safety in autonomous vehicles across various challenging environments.