机场跑道异物FOD(Foreign Object Debris）会对飞行安全造成重大影响。尤其在飞机起降阶段如果机体卷入异物后果十分严重，现阶段跑道异物的清理主要靠人工机器在每日开飞前进行巡检清除，在应对突发异物入侵跑道时只能靠人工视频监控发现，漏检率高，不利于飞行安全。因此本文基于深度学习，提出一种改进YOLOv5算法的机场跑道异物检测方法，通过建立数据集训练学习，实现利用机场摄像头自动检测跑道异物的功能，提升跑道巡检的效率与质量。

解决了目前缺乏相关机场跑道异物检测数据集的问题，通过在某民航机场实际拍摄并进行数据标注，设计并构建了机场跑道异物数据集“FODD数据集”。针对FOD样本量小不易进行大量训练的问题，提出一种改进DCGAN算法，生成机场跑道异物补充样本集。另外通过加深网络深度并引入残差模块解决网络退化问题，改进损失函数提升图片生成质量，用于提高算法的训练效果。最终使用YOLOv5目标检测网络验证数据集有效性。

针对机场跑道与普通地面分割的问题，提出了一种改进的Deeplabv3+跑道分割网络，采用轻量级卷积网络特征提取骨干网络Efficient Netv2，提高参数利用率；在ASPP模块的基础上引入混合条带池化和深度可分离膨胀卷积，构建N-ASPP模块，减少参数量并提高学习多尺度信息的能力。针对机场跑道异物形态各异、小目标居多等特点，提出一种改进的YOLOv5算法，添加小目标检测头，设计多尺度检测头，提高网络对多尺度目标检测的召回率；添加CBAM注意力机制，提高小目标检测精度；使用特征融合网络BIFPN，提升算法的特征融合能力。将两个网络进行融合后并在自制FODD数据集上进行了一系列的实验，实验证明，该算法显著的提高机场跑道异物目标检测网络的性能。

最后设计机场跑道异物检测系统，该系统具备跑道异物的图片检测、视频检测与实时检测功能。通过测试证明本系统能够快速高效完成跑道异物检测任务。

FOD ( Foreign Object Debris ) on airport runway will have a significant impact on flight safety. Especially in the take-off and landing stage of the aircraft, if the body is involved in the foreign body, the consequences are very serious. At this stage, the cleaning of the foreign body on the runway mainly depends on the manual machine to carry out inspection and removal before the daily flight. In response to the sudden invasion of foreign bodies into the runway, it can only be found by manual video monitoring. The missed detection rate is high, which is not conducive to flight safety. Therefore, based on deep learning, this paper proposes an airport runway foreign object detection method based on improved YOLOv5 algorithm. By establishing data set training and learning, the function of using airport camera to automatically detect runway foreign objects is realized, and the efficiency and quality of runway inspection are improved.

The problem of the lack of relevant airport runway foreign object detection data sets is solved. Through the actual shooting and data labeling in a civil aviation airport, the airport runway foreign object data set ' FODD data set ' is designed and constructed. Aiming at the problem that the FOD sample size is small and it is not easy to train a lot, an improved DCGAN algorithm is proposed to generate the airport runway foreign body supplement sample set. In addition, by deepening the network depth and introducing the residual module to solve the network degradation problem, the loss function is improved to improve the quality of image generation, which is used to improve the training effect of the algorithm. Finally, the YOLOv5 target detection network is used to verify the validity of the data set.

Aiming at the problem of airport runway and ordinary ground segmentation, an improved Deeplabv3 + runway segmentation network is proposed. The lightweight convolutional network feature extraction backbone network Efficient Netv2 is used to improve the parameter utilization. On the basis of ASPP module, hybrid strip pooling and deep separable dilated convolution are introduced to construct N-ASPP module, which reduces the number of parameters and improves the ability to learn multi-scale information. Aiming at the characteristics of different foreign bodies and small targets in airport runways, an improved YOLOv5 algorithm is proposed, which adds small target detection head and designs multi-scale detection head to improve the recall rate of multi-scale target detection. Add CBAM attention mechanism to improve the accuracy of small target detection ; the feature fusion network BIFPN is used to improve the feature fusion ability of the algorithm. After the fusion of the two networks, a series of experiments were carried out on the self-made FODD dataset. Experiments show that the algorithm significantly improves the performance of the airport runway foreign object detection network.

Finally, the airport runway foreign body detection system is designed, which has the functions of image detection, video detection and real-time detection of runway foreign body. The test proves that the system can quickly and efficiently complete the runway foreign object detection task.