Design And Implementation Of Foreign Body Detection Algorithms For Airport Runways

As the economy grows,the proportion of air transport and air travel is gradually increasing.Foreign Object Debris(FOD)on airport runways poses a serious threat to aircraft during take-off or landing.At present,domestic and international airports mainly use manual inspection methods to detect foreign objects on airport runways.With the advancement of deep learning technology and the improvement of hardware computing power,the target detection algorithm based on convolutional neural network has been applied to the field of airport runway foreign object detection.However,due to the small size of some foreign objects in airport runways,the accuracy of general-purpose target detection algorithms in detecting these small targets is difficult to meet the requirements of safe airport flight.In addition,there is no public dataset dedicated to airport runway foreign object detection,and the existing generic target detection dataset only contains a small number of FOD species.To this end,this thesis focuses on a convolutional neural network-based algorithm and dataset expansion method for airport runway foreign object detection,with the following work:(1)To address the problem of lacking a common dataset of airport runway foreign objects,this thesis collects images of airport runway foreign objects under various imaging conditions on a highly simulated airport runway surface,including bad weather such as rain,fog and multi-angle images under strong and low light conditions.The FOD3(Foreign Object Debris Detection Dataset,or FODDD),which contains 12 types of airport runway foreign objects such as spanners,screwdrivers and pliers with multi-attribute semantic labels,is constructed.(2)To address the problem of low accuracy of small target detection due to insufficient number of small target samples in the dataset and unbalanced distribution in the training cycle,a multi-scale target detection algorithm YOLOv4-Balance is proposed that combines dynamic loss feedback and data enhancement.first,based on the Mosaic data enhancement method,a data enhancement algorithm U-based on image combination stitching is designed to Mix,in order to balance the distribution of small target samples,eliminate the influence of invalid small target samples on model training,and improve the quality of small target samples in the dataset.Secondly,a multi-scale model training algorithm,DLF(Dynamic Loss Feedback),is designed to balance the distribution of small target samples and improve the contribution of small target sample loss to model learning during the training process by exploiting the loss feedback during model training iterations.(3)The YOLOv4-Sensitive algorithm was designed to alleviate the problem of low accuracy of small target detection due to insufficient information about small target features in the learning process of neural networks.First,a new feature extraction network,U-CSPDarknet53,is designed by reconstructing the residual unit structure in the backbone feature extraction network,CSPDarknet53,to enhance the feature extraction capability of the backbone network and achieve the extraction and retention of small target feature information at a finer granularity.Secondly,parallel expanded convolutional branches are used to extract the contextual information of small targets to alleviate the problem of small target feature loss due to pooling operations in the SPP network.Finally,a CA attention mechanism is introduced into the feature fusion network to integrate location information into the feature aggregation network to enhance the feature description capability and improve the model’s focus on small target feature information.(4)Based on the above algorithm,this study designs and implements a FOD target detection application system based on data enhancement and multi-scale feature fusion.Through the high-speed camera to capture the status of the airport runway surface in real time,the algorithm module outputs the FOD detection information of the runway surface to realize the real-time FOD detection function requirements.The detection results are finally confirmed by the display terminal,realizing real-time FOD detection and removal.Meanwhile,to meet the demand of rapid iteration of the model in practical applications and to improve the training efficiency of the model,this study further improves the feature fusion network in the above algorithm by incorporating the idea of bi-directional cross-scale feature connection,reducing the number of feature fusion nodes that contain less feature information in the feature fusion process,and reducing the number of parameters in the feature fusion network in order to accelerate the convergence of the model.