康轶譞, 刘宇, 王亚伟, 周立君, 郭城, 王怡恬. 基于DETR的道路环境下双目测量系统[J]. 应用光学, 2023, 44(4): 786-791. DOI: 10.5768/JAO202344.0402003
引用本文: 康轶譞, 刘宇, 王亚伟, 周立君, 郭城, 王怡恬. 基于DETR的道路环境下双目测量系统[J]. 应用光学, 2023, 44(4): 786-791. DOI: 10.5768/JAO202344.0402003
KANG Yixuan, LIU Yu, WANG Yawei, ZHOU Lijun, GUO Cheng, WANG Yitian. DETR-based binocular measurement system in road environment[J]. Journal of Applied Optics, 2023, 44(4): 786-791. DOI: 10.5768/JAO202344.0402003
Citation: KANG Yixuan, LIU Yu, WANG Yawei, ZHOU Lijun, GUO Cheng, WANG Yitian. DETR-based binocular measurement system in road environment[J]. Journal of Applied Optics, 2023, 44(4): 786-791. DOI: 10.5768/JAO202344.0402003

基于DETR的道路环境下双目测量系统

DETR-based binocular measurement system in road environment

  • 摘要: DETR(detection transformer)算法是一个基于Transformer的目标检测算法,具有检测速度快、检测效果好的优势。介绍了一种利用DETR算法及双目视觉原理对道路环境下的人、车、自行车、信号灯等目标进行构建的测量系统。分析了双目测距、相机标定、目标检测以及目标匹配的原理,并以此为基础构建了测量系统。采用目标检测算法检测视野中的目标,利用双目视觉原理对检测到的目标进行测距,同时分析了测量系统中测量误差的来源,并计算其对结果的影响。该算法在KITTI数据集及现实环境中进行测试,测量系统基线为45 cm,对15 m~80 m的指定目标检出率高于90.6%,测距误差小于5.8%,在RTX 2080Ti平台上能够实时运行。

     

    Abstract: Detection transformer (DETR) is a target detection algorithm based on Transformer, which has the advantages of fast detection speed and good detection effect. A measurement system based on DETR and binocular vision principle for people, vehicles, bicycles, signal lights and other targets in road environment was introduced. The principles of binocular ranging, camera calibration, target detection and target matching were analyzed, and the measurement system was constructed on this basis. The target detection algorithm was used to detect the targets in the field of vision, and the principle of binocular vision was used to measure the distance of the detected targets. The source of measurement error in the measurement system was analyzed and the influence on the results was calculated. The algorithm was tested in KITTI data set and real environment. The system baseline is 45 cm, the detection rate of 15 m~80 m specified targets is higher than 90.6%, and the ranging error is less than 5.8%. The proposed algorithm can run in real time on RTX 2080Ti platform.

     

/

返回文章
返回