Research On Foreign Object Debris Detection Algorithm Based On Deep Learning

In the past decade,the volume of civil aviation passenger and cargo traffic is at a steady high level,and the runway free time is getting less and less due to the frequent flight takeoffs and landings at various civil aviation airports,which brings increasing pressure to the airport operation and safety management,especially the difficulty of runway foreign body detection and treatment.At present,most of the runway foreign body management work is done manually,so the automatic runway foreign body detection technology,which can reduce the runway occupation time,has become a promising research direction.The runway foreign body detection technology is mainly divided into three technical lines: radar,infrared and image,among which the image-based detection system has become a hot spot for current research due to its low hardware deployment cost,flexible setup and ability to identify foreign body types.Especially,the field of computer vision has made significant progress in recent years,and a number of target detection algorithms based on deep learning theory have been born.Based on the research results in the field of deep learning and target detection,this paper improves the general target detection algorithm YOLOv5 based on convolutional neural network,and researches a more effective runway foreign object detection algorithm based on the characteristics of runway foreign objects and the needs of small and medium-sized airports,in order to address the problem that existing deep learning detection algorithms have low detection accuracy for small targets in complex environments.First,the basic theories and common algorithms of deep learning and target detection are sorted out in detail,the performance indexes of target detection tasks are determined,the structure and advantages of the generic target detection algorithm YOLOv5 are analyzed,and the YOLOv5 network model is selected as the base algorithm.By analyzing and comparing the main runway foreign object datasets,the dataset used in this study is determined and preprocessed.Secondly,the effective channel attention(ECA)module is added to the backbone network of YOLOv5 to strengthen important features and suppress non-essential features using a weighting approach,which brings significant performance gains by increasing a small number of parameters,allowing the network to pay attention over a larger area and thus capture targets efficiently.In the neck network,the original feature pyramid module(PA-Net)is replaced with a weighted bi-directional feature pyramid(Bi FPN)network structure to achieve efficient bidirectional cross-scale connectivity and weighted feature fusion.The EIo U Loss is used to replace CIo U Loss as the loss function at the prediction side,which speeds up the convergence of the function.Based on these improvement points,the improved YOLOv5 algorithm for runway foreign object detection is proposed.Finally,a deep learning training environment is built based on Python and Pytorch,and the improved algorithm proposed in this thesis is trained,tested and analyzed on the processed FOD-A dataset.The test results show that the mean average accuracy(m AP@0.5)index of the improved YOLOv5 model reaches 97.4%,which is 1.6 percentage points better than the original model,and achieves the requirements for airport runway foreign body detection in complex environments.