Research on Cross-modal and Semantic Collaborative Unsupervised Domain Adaptation for All-weather Autonomous Driving Perception Enhancement

Abstract

Autonomous driving perception systems often encounter severe “domain shift” during real-world deployment. Although deep learning models perform well under ideal weather, their accuracy drops significantly in extreme conditions like rain, snow, fog, or low-light nighttime, as well as during sensor modality transitions. Current Unsupervised Domain Adaptation (UDA) methods mostly focus on class-agnostic global alignment, often neglecting fine-grained semantic correlations and complex visual gaps. This paper systematically explores how to use UDA to enhance the robustness of autonomous driving in complex environments. We propose a multi-dimensional framework that synergizes visual style, feature extraction, and semantic correlation. This study combines theoretical analysis with an algorithmic review to organize several cutting-edge technical paths. First, we introduce Style Prompt Tuning guided by pre-trained Vision-Language Models (VLM). This achieves style transfer while preserving image geometry. Second, we utilize the Transformer-based DTN-DETR model, which cooperates with frequency and spatial domain optimization to improve feature extraction in low-light conditions. Finally, we analyze the Graph Embedding Interclass Relation-Aware Adaptive Network (GelraA-Net) based on Graph Convolutional Networks (GCN) and the Dual Semantic Correlation Alignment (DSCA) mechanism. These methods strengthen cross-scene semantic consistency through topological structures and contextual logic.