BEYOND THE PIXEL: ASSESSING REAL-TIME OBJECT DETECTION MODELS IN LOW-VISIBILITY AUTONOMOUS DRIVING ENVIRONMENTS

Abstract

Autonomous driving has made impressive strides in the last decade, but one area where things still get tricky is low-visibility conditions. Fog, heavy rain, snow, dust, and nighttime driving can throw off even the most advanced object detection systems. This paper takes a closer look at how real-time object detection models hold up when the weather and lighting are not on their side. We tested four widely used models, namely YOLOv8, YOLOv9, Faster R-CNN, and EfficientDet, on a curated dataset that combines images from the BDD100K, Foggy Cityscapes, and Nighttime Driving Dataset. The goal was not just to find which model gives the best numbers, but also to understand how each one behaves when pixels get blurry, contrast drops, and edges disappear. We evaluated performance using mean Average Precision (mAP), inference time, and a custom robustness score we designed for this study. Results show that YOLOv8 and YOLOv9 deliver strong real-time performance, but their accuracy drops noticeably in foggy conditions. Faster R-CNN, while slower, holds up better in dense fog and low-light scenarios. EfficientDet sits somewhere in the middle. We also found that simple preprocessing tricks like contrast enhancement and dehazing can recover a fair amount of lost accuracy without slowing things down too much. The paper closes with a discussion on why current models still struggle in these environments and what directions future research could take. Our findings should be useful for anyone working on safety-critical perception systems for autonomous vehicles operating in real-world conditions.