Multi-Domain Adaptation for Autonomous Driving Perception under Diverse Weather

Abstract

Autonomous driving perception systems are confronted with substantial robustness challenges under diverse weather conditions, where sensor data distortion caused by rain, fog, snow, or intense illumination often leads to degraded performance in critical tasks such as object detection and semantic segmentation. Existing approaches predominantly depend on single-domain models trained under ideal environmental conditions, which suffer from poor generalization across weather domains due to inherent domain shifts. This study explores the application of multi-domain adaptation techniques to enhance perception stability by integrating heterogeneous sensor data, including RGB (Red, Green, Blue) images, LiDAR (Light Detection and Ranging) point clouds, and thermal imaging, while leveraging cross-domain feature alignment mechanisms. The proposed framework employs domain-specific encoders combined with adversarial learning to mitigate weather-induced domain gaps, alongside a multi-task learning objective that simultaneously optimizes perception accuracy and domain invariance. Experimental validation demonstrates that the framework achieves superior performance compared to conventional single-domain and shallow adaptation models, with interpretability analyses revealing key weather-robust features such as thermal edge consistency and LiDAR (Light Detection and Ranging) point density patterns. Its ability to adapt to unseen weather conditions could enable reliable autonomous driving in complex real-world environments and reduce weather-related accidents. By bridging domain adaptation theory with automotive perception requirements, this work advances the translation of robust AI (Artificial Intelligence)-driven systems into practical autonomous driving applications.