Research On Key Technologies For Airport Runway Foreign Object Debris Detection And Recognition Based On Video

Foreign Object Debris(FOD)presents on the runway poses a significant threat to the safety of air travel.FOD has the potential to damage aircraft during critical phases of the flight,which can cause catastrophic loss of life and airframe,and lead to increased maintenance and operating costs.FOD detection system can efficiently detect and classify the FOD,analyze the hazard level of the FOD and make alarm.At present,the detection of FOD in China is based on manual searching,which is poor reliability and takes up the valuable runway.Imported FOD detection system is too expensive,over several million dollars,and the technology is blockaded.Therefore,developing a FOD detection system is very important for the Chinese aviation industry.Compared with other detection systems,a video-based detection system has wider monitoring range and complete information,this paper proposes a video-based FOD detection system.Considering the feasibility and flexibility of vehicular detection system,cameras and information processing apparatus are mounted on a vehicle,to form a removable FOD detection and recognition system.This paper studies the key issues of vehicular FOD detection and recognition system.This system can detect and classify FOD,analyze the hazard level of the FOD and make alarm.The main research of this paper can be summarized as follows:1.FOD detection based on feature fusion.Background subtraction is widely used in the filed of object detection with stationary cameras,but this method is not suitable to a mobile system because it is hard to acquire the same region to form the background by using the vehicular cameras.A feature-fusion method is proposed according to the characteristics of the mobile FOD system.In order to avoid the disadvantage of single feature detection method,features such as grayscale,Fourier transformation,edges,textures and shapes are analyzed for runway surface and foreign objects.Moreover,features such as grayscale,edge and brightness are fused to detect the FOD using D-S evidence theory.Multi-frame splicing is used to increase the signal to noise ratio and improve target signal strength,which has good robustness and adaptability.Experimental results present the proposed approach can increase the accuracy up to 95.85% by using multi-features fusion while the offline detection precision increases by 1.37%.2.Pattern features extraction of FOD.The foregoing analysis of features is to detect FOD fast and accurate.While in reality,workers should make a different treatment for the FOD according to its risk level,in order to reduce the occupancy of the runway.But the preliminary detection could not make a judgment of the risk level,this paper further studies the classification of FOD,by which may help the workers make a rapid and correct disposal.Pattern features extraction is the foundation of classification and identification.In this study,we construct Gabor filter to extract pattern features of FOD and test the availability and reliability.Experimental results present the brightness response of 2D Gabor filter can overcome the effect of changes in lighting conditions.In addition,2D Gabor response of position can tolerate slight geometric distortion,and reduce the influence of image noise.3.Dimension reduction of pattern features.Gabor transformation will lead to a very high feature space,which makes it difficult for computers to calculate so much data.Therefore,mapping or transformation schemes are used to convert the vectors to a low-dimensional subspace,which can significantly reduce the dimension of features,and also improve effectiveness of these lower-dimensional space features.Traditional methods of dimension reduction,such as principal component analysis,linear discriminant analysis,local linear embedding are analyzed.This paper proposes a novel dimensionality reduction method following weighted kernel locally linear embedding and comparison is made with classical methods to verify the proposed approach.Experimental results present the proposed approach outperform other methods,when the dimension is 30 and the nearest neighbor is 30.4.Classification of foreign object debris.Classifiers usually need a long time to finish the data training,and have a poor adaptability to unbalance samples.This paper proposes a novel classification algorithm named Directed Acyclic Graph and Twin Support Vector Machine(DAG-TWSVM for short).The main idea of DAG-TWSVM is to use the decision method of DAG-SVM,decomposing a k class problem into(7)(8)kk-21 binary classes,and TWSVM is used as each binary classifier,which solves two Quadratic Programming Problems.Experimental results present the proposed DAG-TWSVM performs better than some widely used multi-class SVMs,it uses less time and outperforms other multi-class approaches on unbalanced dataset.This paper studies the key issue of the video-based FOD detection and recognition system,including FOD detection,pattern feature extraction,feature dimension reduction and classification.The work in this paper is a powerful supplement to the research of FOD detection and recognition system in China,which proposes a low-cost mobile FOD detection and recognition system.The innovations of this paper are summarized as follows:1.This paper proposes a novel FOD detection method based on feature fusion and optimizes the FOD detection system.In the actual experiments,several performance indicators are calculated,such as precision,recall rate,false alarm rate and missing rate.Experimental results present the proposed approach can achieve a better performance.2.This paper creates image bank of FOD,which may provide basic information and statistical analysis for the development and optimization of FOD detection systems.Gabor filter is used to extract pattern features of FOD,and the availability and reliability are validated.3.This paper proposes a novel algorithm for dimension reduction named as weighted kernel local linear embedding,this approach reduces the impact of noise and the points out of samples.Comparison is made with traditional methods PCA,LDA,LE,LLE using image bank of FOD.Experimental results present the proposed approach is superior to other methods in dimension reduction and classification.4.This paper presents a novel classification algorithm termed as Directed Acyclic Graph and Twin Support Vector Machine,which combining the advantages of DAG-SVM and TWSVM.The proposed algorithm uses less time in training than other multi-class SVMs,and performs better with unbalanced samples.