DETR for traffic sign detection

Abstract Traffic sign detection refers to accurately perceiving and locating traffic signs in driving scenes, which is crucial for safe autonomous driving. However, existing detection models perform poorly in detecting small objects due to the inherently small size of traffic signs. In addition, the complex and cluttered backgrounds in real-world environments significantly limit detection performance, making traffic sign detection a challenging task in the field of intelligent transportation. To address this challenge, this work proposes a novel traffic object detection model called Traffic-DETR to accurately detect traffic signs in images. For the poor accuracy in detecting small targets, the model introduces a gated residual attention structure GRA, which dynamically controls the contribution of attention features across encoder layers and enhances feature diversity. To address the issue of missed detections in complex backgrounds, we design a Region of Interest-based Decoder module RoID that progressively refines spatial attention regions and helps the model better focus on foreground targets. To resolve the mismatch between classification and localization, we further propose a joint loss function that explicitly correlates classification scores with bounding box regression accuracy, thereby enhancing the quality and consistency of final predictions. Compared with state-of-the-art methods, the proposed Traffic-DETR achieves an AP of 55.7% on the TT100K dataset and 74.3% on the CCTSDB 2021 dataset, demonstrating superior performance and robustness in complex driving scenarios.