汽车工程 ›› 2025, Vol. 47 ›› Issue (6): 1188-1197.doi: 10.19562/j.chinasae.qcgc.2025.06.017

• • 上一篇    

基于轻量级RT-DETR-tiny的车辆目标检测算法

隆艾岐1,冯治国1,2(),张振博1,田兴强3,向巍3,4   

  1. 1.贵州大学机械工程学院,贵阳 550025
    2.贵州省特色装备及制造技术重点实验室,贵阳 550025
    3.贵州交通职业大学,贵阳 551400
    4.贵州省山区公路安全与车路协同研究实验室,贵阳 551400
  • 收稿日期:2024-12-03 修回日期:2025-01-07 出版日期:2025-06-25 发布日期:2025-06-20
  • 通讯作者: 冯治国 E-mail:zgfeng@gzu.edu.cn
  • 基金资助:
    第二十七届中国科协年会学术论文。贵州省科技重大专项(黔科合重大专项字 ZNWLQC[2019]3012)和贵州省交通运输厅科技项目(2021-321-020)

Vehicle Target Detection Algorithm Based on Lightweight RT-DETR-tiny

Aiqi Long1,Zhiguo Feng1,2(),Zhenbo Zhang1,Xingqiang Tian3,Wei Xiang3,4   

  1. 1.School of Mechanical Engineering,Guizhou University,Guiyang 550025
    2.Guizhou Key Laboratory of Special Equipment and Manufacturing Technology,Guiyang 550025
    3.Guizhou Communications Polytechnic University,Guiyang 551400
    4.Guizhou Laboratory for Mountain-Area Highway Safety and Vehicle-Road Collaboration Research,Guiyang 551400
  • Received:2024-12-03 Revised:2025-01-07 Online:2025-06-25 Published:2025-06-20
  • Contact: Zhiguo Feng E-mail:zgfeng@gzu.edu.cn

摘要:

针对自动驾驶场景的硬件限制以及轻量级算法对小目标车辆检测困难的问题,提出了一种新的轻量级车辆目标检测算法RT-DETR-tiny。首先,提出了一种新的冗余图快速生成模块(ReduFast block),利用级联式特征提取结构,避免冗余信息导致的小目标特征信息丢失,并降低计算冗余。基于此模块设计的轻量级网络ReduFastNet作为特征提取网络,相比其他轻量级网络可实现更快的推理速度。其次,在特征融合阶段引入DGSTM模块,使得模型进一步轻量化;同时设计EAAIFI模块,保证了特征融合阶段的实时性。最后,针对小目标车辆检测中边界框易受噪声影响的问题,引入DIOU来优化原损失函数,提高目标中心位置准确性,减少预测框宽高比波动对模型的过度惩罚。实验结果表明,在BDD100K-Urban nighttime数据集上相较于基线算法,所提算法检测精度达到75.3%,仅损失0.1%,而参数量和计算量分别下降37.1%、33.5%,每秒检测帧数达到45.1,检测速度提升了5个百分点。在UA-DETRAC-Small Car数据集上与其他主流轻量级目标检测模型相比,RT-DETR-tiny兼顾了较高检测精度和较小参数量、计算量,优于同类目标检测算法,更有利于自动驾驶场景对车辆目标实时检测的准确率及边缘部署。

关键词: 自动驾驶, 轻量化, 车辆目标检测, RT-DETR算法

Abstract:

For the hardware limitation in autonomous driving scenarios and the challenges faced by lightweight algorithms in detecting small target vehicles, a novel lightweight vehicle detection algorithm, RT-DETR-tiny, is proposed. Firstly, a new Redundant Graph Rapid Generation Module (ReduFast block) is proposed, which uses a cascade feature extraction structure to avoid the loss of small target feature information caused by redundant information, and reduce the computational redundancy. The lightweight network ReduFastNet, designed based on this module, serves as the feature extraction network, achieving faster inference speed compared to other lightweight networks. Then, during the feature fusion stage the DGSTM module is incorporated to further streamline the model, while the EAAIFI module is designed to ensure real-time performance during feature fusion. Finally, for the problem of boundary boxes being susceptible to noise in small target vehicle detection, the DIOU is introduced to optimize the original lossoriginal loss function, enhancing the accuracy of the target center positions and mitigating excessive penalties of the model caused by on aspect ratio fluctuations of the prediction box. Experimental results demonstrate that, on the BDD100K-Urban nighttime dataset, the proposed algorithm achieves a detection accuracy of 75.3%, with only a 0.1% loss, while parameters and computational load decreases by 37.1% and 33.5%, respectively, achieving a detection frame rate of 45.1 frames per second and enhancing detection speed by 5 percentage points. In comparison to other mainstream lightweight object detection models on the UA-DETRAC-Small Car dataset, RT-DETR-tiny balances high detection accuracy and minimal parameter count and computational load, outperforming similar object detection algorithm, which facilitates accuracy and edge deployment for real-time vehicle detection in autonomous driving contexts.

Key words: automatic driving, lightweight, vehicle target detection, RT-DETR