Research on Robust Multi-Target Detection for Intelligent Driving under Complex Weather Conditions

In intelligent driving systems, the visual perception module often suffers from significant performance degradation under complex weather conditions, which poses a serious challenge for all-weather deployment. Most existing studies concentrate on either image restoration or model optimization tailored to specific weather phenomena such as fog or rain, yet they frequently lack generalization ability when confronted with variable or mixed severe weather. This paper introduces a novel task-driven framework designed for robust perception. The core concept is to treat complex weather not as noise to be removed, but as an inherent environmental attribute that the model should learn to adapt to. Methodologically, we propose a task-oriented weather-invariant feature learning module, integrated with a dynamically weighted multi-modal fusion mechanism. This enables the learning of robust cross-weather domain representations and adaptive information complementarity directly at the feature level. Comprehensive experiments conducted on multiple complex weather datasets—including BDD100K, ACDC, and nuScenes—show that our approach substantially outperforms mainstream baseline methods in terms of mean Average Precision (mAP) and F1 score under diverse conditions like fog, rain, and snow. Thus, it provides a practical technical pathway toward achieving highly reliable all-weather perception for intelligent driving systems.