Pantnagar Journal of Research

Abstract

Object detection under fog, rain, and dust remains one of the hardest open problems in autonomous driving. Most existing detectors either collapse in poor visibility or treat near and far objects with equal priority — neither of which is acceptable for real-world deployment. This paper introduces a framework that tackles both problems at once. We propose a Weather Restoration GAN that cleans up degraded images in real time, and a Depth Fusion Gate that steers the detector^’s attention toward the 0–20 m zone where collision risk is highest. Training follows a three-stage curriculum on the KITTI dataset. In dense fog the system reaches 78.2% mAP, a 25% gain over baseline, and 87.5% mAP in the near range, running at 30 FPS with 33 ms latency and a 6.5 GB memory footprint. Compared to LM-CNN-SVM, YOLOv4, and Bayesian Neural Networks, our approach is the only one that jointly handles weather degradation and range-specific prioritization.