Design Of Attitude Estimation Algorithm For Space Non-cooperative Targets Based On Deep Learning

The information acquisition of non-cooperative targets in space has high research value in space confrontation,on orbit maintenance and other fields.To explore the recognition method of non-cooperative targets is also an objective need to maintain national air security.Therefore,it is of great significance to study the attitude recognition scheme of the space non-cooperative target.In this paper,we will use depth learning method and Tensor Flow framework to achieve a gesture recognition algorithm based on visual images.In this paper,a set of relatively complete spatial non-cooperative target recognition algorithm is designed to input a gray image data with satellite images,output satellite attitude angle,or input a series of continuous images of image data,and output satellite angular velocity vector.First,the images with satellite image information are automatically divided,stretched and grayscale through the algorithm of MSER and other image data preprocessing,and the standard gray images of 299×299 pixels are obtained.Then the attitude angle information of the satellite is identified through convolution neural network.By entering a series of continuous time stamp image data,the attitude angle data are obtained by the convolution neural network,and the rotational angular velocity information of the satellite can be obtained by the regression analysis.In addition,the influence of the structure of different optimizer on the network performance and convergence is studied in contrast to the training convolution neural network,and the optimizer and learning rate suitable for the model are found through experiments,and the Loss function suitable for satellite attitude recognition is designed.A Inception V3 convolution neural network with 42 layers of depth is designed and implemented through the Tensor Flow architecture.It first is a common structure alternately between 5 coiling layers and 2 pool layers,connecting 3 Inception modules,each of which contains a number of similar Inception Module structures.When the number of layers is higher,the number of parameters is reduced,the training efficiency is accelerated,and the development cycle is shortened.In the aspect of data collection,the 3DMax software is used to build analog satellite model,to obtain the training set data and to mark up the batch,to generate a data set containing 17000 image samples for the training of the convolution neural network.Data enhancement is performed by setting random brightness and contrast,increasing random noise,and standardization of data,which is equivalent to multiplying the number of data sets used for network training.Experiments on the proposed algorithm are carried out in this paper.It is proved that the algorithm space in this paper is feasible and effective in identifying non-cooperative targets.