Object detection under adverse environmental conditions remains a critical challenge for autonomous driving systems, with performance degrading by over 40% under dense fog, heavy precipitation, or nighttime scenarios. Existing approaches either specialize for single degradation types or remain fundamentally condition-agnostic, failing to address the compound nature of real-world adverse conditions. We present a Condition-Aware Cross-Scale Transformer (CACST), a unified framework integrating explicit environmental condition estimation, adaptive image preprocessing, and cross-scale feature interaction. Our approach introduces: (1) a condition estimation module quantifying degradation parameters via supervised and contrastive learning; (2) a gated image adaptation mechanism with learnable ISP operations dynamically modulated by estimated conditions; and (3) an Enhanced Density-aware Cross-Scale Transformer (EDST) performing global attention across feature pyramid scales with spatial reliability weighting. Comprehensive experiments across five benchmarks demonstrate state-of-the-art performance: CACST achieves +4.1% mAP on ACDC (47.8% vs. 43.7%), +3.2% on Foggy Cityscapes (49.7% vs. 46.5%), and +6.7% on ExDark (74.9% vs. 68.2%) compared to the strongest specialized baseline, maintaining balanced accuracy across all conditions
Copyrights © 2026