Enhanced Pothole Detection in Complex Environments Using ARCH‐RTDETR : A Lightweight and Efficient Approach

ABSTRACT Detecting potholes in complex environments poses challenges such as varying illumination, shadows, and occlusions. Traditional methods often suffer from insufficient detection accuracy and poor real‐time performance. To enhance detection robustness without sacrificing inference speed, this paper adopts the RT‐DETR (Real‐Time Detection Transformer) framework—which requires no NMS (Non‐Maximum Suppression) post‐processing and features an efficient hybrid encoder—as its foundation. We propose the lightweight and efficient ARCH‐RTDETR detection model. The model introduces targeted enhancements to the backbone, feature‐fusion module, and multi‐scale architecture. Specifically, an AFGCA (Adaptive Fusion Global Context Attention) mechanism strengthens sensitivity to subtle cues; RepBN (Reparameterized Batch Normalization) is deeply integrated into the AIFI (Adaptive Instance Feature Integration) module to optimize feature distributions and increase multi‐scale representational capacity; and the proposed CA‐HSFPN (Coordinate Attention‐guided Hierarchical Scale Feature Pyramid Network) improves the effectiveness of cross‐scale feature fusion. Experiments on diverse datasets show that ARCH‐RTDETR achieves an average detection accuracy of 85%, outperforming the RT‐DETR baseline by 2.9%, while also improving detection precision and inference efficiency. These results indicate strong potential for deployment in intelligent transportation systems. This research provides a technical reference for small object detection, addressing the low efficiency of traditional manual inspections and the high detection latency of existing equipment in intelligent transportation systems, thereby offering a reliable technical solution for road safety assurance.