Multimodal road perception with illumination adaptation in autonomous vehicles

The development of autonomous driving technology is reshaping transportation methods. However, the significant decline in perception capabilities under low-light conditions such as rain, fog, and nighttime remains a key bottleneck restricting its widespread application. This paper proposes SafeDrive-Fusion, a novel illumination-adaptive multimodal perception framework that introduces dynamic cross-modal attention fusion and safety-critical region prioritization for robust road passability assessment in autonomous driving. Unlike existing static fusion approaches, this framework addresses the problem of significantly reduced perception capabilities in dim environments through an innovative illumination-adaptive weighting mechanism that dynamically integrates visual and radar information while prioritizing safety-critical road regions. Experimental results show that SafeDrive-Fusion maintains 87.5% accuracy even under extremely low-light conditions, with a performance degradation of only 6.1%, far superior to existing methods' degradation range of 16.6%-34.8%. Particularly in safety-critical "hazardous situation" recognition, the system achieves an F1 score of 0.82, which is 0.40 higher than baseline methods. Through its illumination-adaptive modal weight dynamic adjustment mechanism, the system can automatically adjust its dependence on different modalities according to environmental conditions. As illumination decreases from normal to extremely low, the radar modality weight increases from 25% to 62%. Meanwhile, SafeDrive-Fusion features a low computational latency of 45ms, meeting the real-time requirements of autonomous driving.