Scene Navigability Analysis Based On Deep Learning Model

Scene matching aided navigation is widely used in autonomous navigation of aircrafts because of its high navigation accuracy,strong anti-jamming ability and low support difficulty.Scene navigability analysis is an important factor that affects the performance of scene matching aided navigation system,and the research on this technology has high engineering application value.In this paper,the problem of scene navigability analysis is studied.Firstly,considering the actual application of aided navigation,the problem of scene region adaptation is analyzed.On the one hand,the basic index of scene navigability analysis matching probability is studied.Combined with the situation that there is no corresponding image pair in the task of scene navigability analysis,the calculation method of single image matching probability based on the generation of simulated realtime image is proposed,which lays the foundation for the analysis method of scene region adaptability.On the other hand,the paper studies the factors that affect the adaptability,mainly analyzes the influence of noise interference,distortion and natural scene characteristics and scene matching algorithm on the scene navigability.Then,based on the deep analysis of the problem of scene navigability,this paper puts forward two methods of scene navigability analysis based on artificial neural network,which can automatically acquire the basic features of image through the network,avoiding the problems of large amount of work,strong subjectivity and narrow scope of application for defining the basic features of image.One method is based on PCANet and MLP.Firstly,the basic index of image adaptability is calculated by the single image matching probability calculation method proposed in this paper,and the image adaptability label is obtained after binarization,then the information of image global dimension reduction is extracted by PCANet,and finally the information of image global dimension reduction is input into MLP and the network is trained to get the adaptability evaluation model.PCANet obtains the local information of the image by sliding window block sampling and convolution operation,and then forms the global information of the image by statistical histogram feature information,histogram feature cascade and other operations.MLP can recognize the high dimension output of PCANet,so the image can be divided into two types: adaptive and non-adaptive.Another method is based on the improved block bidirectional 2DPCA and ResNet.Similar to the first method,the basic navigability index of the image is calculated first and then the navigability label of the image is obtained by binarization.In the process of image feature extraction,considering the uniqueness of the basic features of the image,this paper proposes a block bidirectional 2DPCA to obtain the local dimension reduction information of the image.Finally,the local dimension reduction information of the image is input into ResNet training neural network to get the fitness evaluation model.The block bidirectional 2DPCA keeps the spatial information of the image,so the extraction result of the block bidirectional 2DPCA is input to RESNET which has excellent image processing performance for pattern recognition.Finally,the two methods are verified by experiments on the data set which is specially used for scene navigability analysis.The experimental results show that both of the two methods can achieve high accuracy of fitness prediction under the condition of less time-consuming.