Reliable perception in underground parking remains difficult for autonomous vehicles, as rapid illumination changes degrade detector performance and increase safety risks. To mitigate this issue, we introduce the Urban Underground Parking (UUP) dataset, which contains more than 15,000 annotated images collected under dim, moderate, and bright lighting conditions across multiple regions and parking layouts. In addition, we develop ILLDet, a YOLO-based object detector enhanced with an illumination adapter that refines the input image before feature extraction. The adapter relies on learnable encoder–decoder blocks to form multi-exposure representations that reduce shadows, low-light regions, and glare while preserving spatial detail. Experimental evaluation on UUP indicates that ILLDet reaches 71.2% mAP@50, surpassing recent detection baselines in challenging lighting scenarios while keeping a balanced precision–recall trade-off. The detector shows more robust performance compared to baselines under illumination shifts, noise, and blur, suggesting good generalization to real underground environments. Although the enhancement module increases model complexity, the computational cost remains moderate and suitable for deployment. Overall, the proposed dataset and model establish a reproducible benchmark and provide a practical solution for vision-based perception in GPS-denied, safety-critical underground parking applications.