Trustworthy Vision in Fog: Enhancing Detection Quality and System Reliability for Autonomous Driving

To address critical safety challenges in visionbased autonomous driving under foggy conditions, this paper proposes FOG-DETR, a trustworthy object detection framework that synergizes detection quality enhancement, system reliability assurance, and security-driven design. First, we design an Intensity-Guided Spatial Adaptive Attention Convolution (IGSAAC) mechanism to enhance feature quality by dynamically reorganizing spatial distributions, enabling precise separation of clear and fog-affected regions. Combined with a dual-path histogram self-attention strategy, our method achieves reliable feature representation through globallocal dynamic aggregation. Second, a Small Object-Enhanced Pyramid Module (SOEPM) is proposed to improve detection reliability for safety-critical small targets (e.g., pedestrians and vehicles) by integrating multi-scale features via OmniKernel fusion. The DySample module further ensures operational stability by eliminating upsampling artifacts while maintaining real-time efficiency. Experimental results on the RTTS(Rainy Traffic Testing Set) dataset demonstrate that the optimized FOGDETR achieves 2.8 % and 2.1 % improvements in mAP0.5 and mAP0.5:0.95 metrics respectively compared to the baseline RTDETR, with nearly equivalent parameter count. These advancements establish FOG-DETR as a quality-assured solution for vision systems in security-sensitive scenarios like autonomous driving and surveillance.