Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (5): 264-270.DOI: 10.3778/j.issn.1002-8331.2010-0110

• Graphics and Image Processing • Previous Articles     Next Articles

Improved SSD-Based Multi-scale Object Detection Algorithm in Airport Surface

HUANG Guoxin, LI Wei, ZHANG Bihao, LIANG Binbin, HAN Xiaodong, GONG Jianglei, WU Changqing   

  1. 1.National Key Laboratory of Fundamental Science on Synthetic Vision, Sichuan University, Chengdu 610065, China
    2.School of Aeronautics and Astronautics, Sichuan University, Chengdu 610065, China
    3.Wisesoft Co. , Ltd. , Chengdu 610045, China
    4.General Department of Communications and Navigation Satellites, China Academy of Space Technology, Beijing 100094, China
  • Online:2022-03-01 Published:2022-03-01



  1. 1.四川大学 视觉合成图形图像技术国防重点学科实验室,成都 610065
    2.四川大学 空天科学与工程学院,成都 610065
    3.四川川大智胜软件股份有限公司,成都 610045
    4.中国空间技术研究院 通信与导航卫星总体部,北京 100094

Abstract: Aiming at the problem that the existing general object detection method based on deep learning is difficult to detect the airport scene environment object with large scale difference, especially small targets are difficult to detect. This paper proposes a multi-scale object detection algorithm based on SSD algorithm combined with feature pyramid fusion network. The algorithm first uses ResNet-50 as the backbone network, and separately designs six additional feature layers. Secondly, the feature pyramid network is used for feature fusion to obtain more robust semantic information. Finally, Soft-NMS is used to solve the missing detection situation, and the scale ratio of the prior frame is adjusted to better detect small objects. Experiments on the airport surface dataset show that the improved algorithm can achieve 86.31% mAP when the inferred speed is 32?frame/s, which is at the leading level compared with other advanced detectors.

Key words: feature pyramid, object detection, airport surface surveillance, Soft-NMS

摘要: 针对现有基于深度学习的通用目标检测方法对机场场面环境目标尺度差别大,特别是小目标难以检测到的问题,提出了一个基于SSD算法并结合特征金字塔融合网络的多尺度目标检测算法。该算法采用了更深的ResNet-50作为骨干网络,并单独设计了六层额外特征层。使用特征金字塔网络进行特征融合,以获得更鲁棒的语义信息。使用Soft-NMS以解决存在的漏检情况,调整先验框的尺度比以更好地检测小目标。通过在机场场面数据集实验表明,该改进算法能够在推断速度为32?frame/s的情况下,取得86.31%的mAP,对比其他先进的检测器,达到领先水平。

关键词: 特征金字塔, 目标检测, 机场场面监视, Soft-NMS