Research On FOD Detection Algorithm Based On Key Point Estimation And Lightweight

Foreign Objects Debris(FOD)on the airport runway will bring a huge threat to aviation safety,which will not only bring a large amount of direct and indirect economic losses,but also endanger the lives of passengers.In order to automatically detect and identify FOD,researchers at home and abroad have conducted a lot of research on FOD detection systems.There are already mature systems that have been widely adopted in foreign countries,but China has no access to the key technologies.And they are very expensive,while domestic research on the detection of FOD is inadequate.Most of the FOD detection systems developed are in the testing stage,and haven’t been invested in the airport runway for practical application.Therefore,in order to realize the localization of the FOD detection system as soon as possible,further research is needed.With the rapid development of computer vision technology,object detection algorithms based on deep learning have become a powerful tool to solve this problem.This paper mainly studies the key technologies of airport FOD detection system based on optical images.The main research of this paper can be summarized as follows: 1.A FOD detection algorithm based on keypoint estimation and semantic segmentation is proposed.In the actual environment,we have a big problem that most of our targets are very small in the images because of the wide airport runway and the small FOD.In order to improve the detection accuracy of small targets,this paper first constructs the implementation based on key point estimation,which can enhance the ability to recognize small targets by performing target detection at a higher resolution.At the same time,a more flexible method to set training sample labels is proposed to make the model more suitable for FOD target detection;a multi-scale prediction strategy based on atrous convolution is added to improve the detection accuracy of targets of various sizes;Local prediction results are fused in the post-processing stage to enhance the model’s positioning capabilities.In addition,due to the special nature of the FOD target detection task,all foreign objects on the runway should be detected,regardless of the category,which is different from traditional target detection.In order to detect all foreign objects appearing on the runway without being limited by the target category of the training set,a method that uses semantic segmentation to separate backgrounds to guide the detection of FOD targets with unknown categories is proposed in this paper.In order to meet the needs of the segmentation task,The bilinear interpolation layer used for upsampling in Hourglass was replaced by the deconvolution layer;Facol loss is used to alleviate the imbalance of the positive and negative samples of the segmentation branch;In order to reduce the workload of sample labeling,the training of segmentation branches is performed using a weakly supervised method.Experiments show that compared with the current representative methods YOLO,Faster R-CNN,and Center Net,the proposed method can detect smaller FOD targets more effectively,and the model has a stronger ability to detect FOD targets with unknown categories.2.A lightweight FOD detection model is designed.Considering the large amount of data in actual applications and the possible shortage of hardware equipment,we need to increase the detection speed of the model as much as possible.In order to improve the training and prediction efficiency of the model and reduce the storage cost and calculation cost of the model,the design concept of lightweight models was introduced.The ideas of Mobile Net and Squeeze Net are used to lighten and improve the model in Chapter 3.A model improvement method that uses both deep separable convolution and channel compression methods to achieve compression of computational and model parameter quantities.Experiments show that the training and prediction efficiency of the proposed network has been improved significantly.