To address the challenges of small-scale traffic sign detection and constrained computational resources, we propose LRP-RTDETR, an edge-friendly efficient high-precision framework based on RT-DETR. We integrate Large Separable Kernel Attention (LSKA) to broaden the receptive field for fine-grained feature capture. The backbone and neck are rearchitected using a GELAN-based framework with RepNCSPELAN4 and ADown modules, reducing structural redundancy while enhancing feature aggregation. Furthermore, an Inner-PIoU v2 loss function, utilizing a non-monotonic attention mechanism, is introduced to mitigate gradient stagnation and improve localization precision. Experimental results on the TT100K dataset show that LRP-RTDETR outperforms the RT-DETR-R18 baseline with a 2.6% increase in mAP@0.5, a 53.4% (30.6 GFLOPs) reduction in computational complexity, and a 12 FPS improvement in inference speed, demonstrating a favorable accuracy-efficiency trade-off for resource-constrained intelligent transportation systems.

autonomous driving, traffic sign, LRP-RTDETR, lightweight model