University Archives

Autonomous driving and advanced driver-assistance systems rely heavily on computer vision models for perception tasks such as object detection. However, the reliability of these systems can degrade significantly under adverse weather conditions such as heavy snowfall, rain, fog, and reduced visibility. This project investigates the robustness of deep learning-based object detection models under challenging environmental conditions relevant to real-world driving scenarios.

A baseline object detection model based on YOLOv8 is trained on the BDD100K driving dataset after remapping the original labels into three safety-critical classes: vehicle, pedestrian, and cyclist/rider. The trained model is evaluated on the BDD100K validation set to establish in-domain performance. To assess robustness under adverse weather, the model will be further tested on the ACDC dataset, which contains driving scenes captured in snow, rain, fog, and nighttime conditions.

The objective of this work is to analyze performance degradation across weather conditions and identify failure modes in perception systems used for autonomous driving. By comparing in-domain and adverse-weather detection performance, the study aims to highlight limitations of current models and motivate future improvements in weather-robust perception for automotive safety applications.