复杂机场道面外来异物高精度实时检测算法

PDF下载HTML

Title:

High-precision real-time detection algorithm for foreign object debris on complex airport pavements

作者:

李海丰¹, 李纪霖¹, 王怀超¹, 桂仲成²

1. 中国民航大学 计算机科学与技术学院, 天津 300300;
2. 成都圭目机器人有限公司, 四川 成都 610310

Author(s):

LI Haifeng¹, LI Jilin¹, WANG Huaichao¹, GUI Zhongcheng²

1. College of Computer Science and Technology, Civil Aviation University of China, Tianjin 300300, China;
2. Chengdu Guimu Robot Co., Ltd., Chengdu 610310, China

关键词:

机场道面外来异物; 小目标检测; 多尺度融合; 纹理信息提取; 超分辨率; 亚像素卷积; 特征提取; 完全交并比

Keywords:

foreign objects debris on airport pavement; small target detection; multi-scale fusion; texture information extraction; super-resolution; subpixel convolution; feature extraction; complete intersection ratio

分类号:

TP391

DOI:

10.11992/tis.202110014

摘要:

机场道面外来物(foreign object debris, FOD)具有类型多样、形状各异、背景复杂、目标弱小等特点，并且严重影响飞行器安全，故其高精度实时检测具有重要意义。针对以上问题，提出基于超分辨率特征金字塔并带有纹理信息提取模块的FOD实时检测网络(FOD real-time detection network, FOD-RDN)。该网络采用Darknet-53作为主干网提取特征，通过超分辨率特征金字塔对形状各异的小目标进行检测，并设计了纹理信息提取模块降低复杂背景的干扰。同时采用双通道YOLO检测器和基于CIoU的损失函数，进一步提升网络对FOD的检测精度和速度。实验结果表明，本文算法在满足实时性要求的情况下，在FOD数据集上整体检测精度达到了91.8%，相比于主流目标检测网络在FOD目标检测方面具有更好的检测效果。

Abstract:

Foreign object debris (FOD) on the airport pavement has the characteristics of diverse types, different shapes, complex backgrounds, and weak targets, which seriously affect safety of aircrafts. Therefore, high-precision real-time detection of FOD on the airport pavement is of great significance. To solve the above problems, we propose a FOD real-time detection network (FOD-RDN) based on super-resolution feature pyramid with texture information extraction module. The network uses Darknet-53 as the backbone network to extract features, and detects small targets with different shapes through super-resolution feature pyramid. Then the texture information extraction module is designed to reduce the interference of complex background. At the same time, a dual-channel YOLO detector and a CIoU-based loss function are used to further improve the accuracy and speed of the network detecting FOD. The experimental results show that the algorithm in this paper can achieve an overall detection accuracy of 91.8% on the FOD dataset under the condition of meeting the real-time requirements, which is better than the mainstream object detection network in terms of FOD detection.