Abstract: Accurate positioning and real-time detection of traffic signs are the core links of intelligent inspection of transportation facilities. However, in real road scenarios, long-distance traffic signs are easily partially occluded by vehicles, pedestrians, trees and other objects, which seriously restricts the performance of existing object detection algorithms. To solve the above problem, an occlusion-aware feature-interactive robust traffic sign detection model based on the YOLOv8s framework is proposed, with detection capability improved through two key architectural innovations. Firstly, drawing on the attention mechanism of Real-Time Detection Transformer (RT-DETR), an Attention-Guided Intra-Scale Feature Interaction (AIFI) module is designed and embedded into the neck network of YOLOv8s. By adaptively learning the internal contextual information of feature maps, it generates more discriminative multi-scale semantic features and enhances the representation of small target features. Secondly, to address the occlusion issue, the Separated Enhanced Attention Module (SEAM) from YOLO-Face V2 is introduced to construct an occlusion-aware SEAM-Head detection head. It adaptively amplifies the local feature weights of occluded regions, improving the model’s robustness to partially occluded targets. Training and evaluation are conducted on a traffic sign dataset composed of real-world road images and online open-source images and containing partially occluded samples. Experimental results (including ablation experiments) show that all performance metrics of the proposed model are superior to the baseline YOLOv8s, with the mAP@0.5:0.95 metric improved by 11.03 percentage points. It performs particularly well in the detection tasks of small-sized and partially occluded traffic signs, fully verifying the effectiveness of the AIFI module and SEAM-Head design. This provides reliable technical support for the intelligent inspection of traffic signs in complex road scenarios.