Abstract
Traffic sign detection in urban scenes is challenged by small objects, similar category appearances, and background clutter. This paper presents Urban-DETR, a semantics-enhanced framework built on RT-DETR. Frozen image and text encoders from Contrastive Language-Image Pre-training (CLIP), referred to as CityCLIP in the architecture, supply image-level and category-level embeddings. A backbone incorporating LSKA extracts multi-scale visual features. UrbanFusion conditions learnable queries on the image embedding, attends to category text embeddings, and pools the query features into a semantic vector for channel-wise visual feature modulation. On the selected TT100K subset, the reported mean average precision (mAP) over intersection-over-union (IoU) thresholds from 0.50 to 0.95 increases from 60.4% for RT-DETR to 62.6% for Urban-DETR; precision increases from 82.9% to 86.9%. Evaluation on the Changsha University of Science and Technology Chinese Traffic Sign Detection Benchmark 2021 (CCTSDB 2021) gives 61.5% mAP over the same IoU range. These results support improved aggregate detection accuracy for the reported configurations.