This paper proposes an improved RT-DETR (Real-Time Detection Transformer) method to address key challenges in highway target detection, including ambiguous feature recognition, missed multi-scale object detections, and the reliance on manual component design in traditional models. This method introduces the CGResNet (Context-guided Resnet) backbone network that integrates local features, environmental context, and global semantic information to strengthen feature representation for indistinct targets. It further employs an Attention-based Intra-scale Feature Interaction (AIFI) module and CNN-driven Cross-scale Feature Fusion (CCFF) module, optimizing semantic integration efficiency across multi-scale features. Experimental results on the Jiangsu Highway Dataset demonstrate that the improved model achieves 45.4% Average Precision (AP) and 72.6% AP50. This work outperforms the original RT-DETR in critical metrics such as AP75 (+3.0%) and APl (+0.6%), with particularly enhanced robustness in low-light and blurry scenarios. The real-time performance of the model (69 FPS) is slightly lower than YOLOv11-S due to the computational complexity of the CGBlock. However, its superior detection accuracy and adaptability to the scene make it as a robust and high-precision end-to-end solution for intelligent highway surveillance systems.
Fu et al. (Fri,) studied this question.