Journal of Applied Optics, Volume. 46, Issue 2, 300(2025)

Lightweight traffic sign detection based on RT-DETR

Haolong XIE1... Xiaolong ZHANG1,2,*, Peixu WEI1 and Chuanjin CUI1 |Show fewer author(s)
Author Affiliations
  • 1College of Electrical Engineering, North China University of Science and Technology, Tangshan 063200, China
  • 2Zhejiang Yaoheng Optoelectronic Technology Co.,Ltd., Hangzhou 311600, China
  • show less

    Traffic sign detection is the key in the driving process of intelligent vehicles, which is of great significance to vehicle analysis of road conditions. Aiming at the problems of too many parameters and low accuracy of existing traffic sign detection algorithms, a lightweight traffic sign detection algorithm based on improved real-time detection transformer (RT-DETR) was proposed. Firstly, the ResNet network in the backbone network of the model was replaced by VanillaNet network to reduce the number of layers and parameters. Secondly, the bi-directional feature pyramid network (BiFPN) was used to replace the path aggregation network (PAN) structure in the feature fusion module of RT-DETR to increase the fusion ability of the model and extract more abundant feature information. Finally, the global attention mechanism (GAM) was added to the feature fusion module to enhance the model perception of global information and improve the detection performance of multiple targets and obscured targets. The proposed algorithm was tested on traffic sign datasets. The mean average precision (mAP) value of the improved RT-DETR algorithm reaches 87.7%, 3.6% higher than that of the original algorithm, and the parameters are reduced by 21%, meeting the deployment needs of intelligent vehicle equipment, which proves the effectiveness of the improved algorithm.

    Keywords
    Tools

    Get Citation

    Copy Citation Text

    Haolong XIE, Xiaolong ZHANG, Peixu WEI, Chuanjin CUI. Lightweight traffic sign detection based on RT-DETR[J]. Journal of Applied Optics, 2025, 46(2): 300

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Dec. 21, 2023

    Accepted: --

    Published Online: May. 13, 2025

    The Author Email: ZHANG Xiaolong (张孝龙)

    DOI:10.5768/JAO202546.0202002

    Topics