Study On 3-D Imaging And Physical Feature-Driven Recognition Of Space Targets

With the development of space technology,space surveillance technology have become a vital part of national security.Because of the all-day,all-weather,and long-range effects,radar plays an important role in space surveillance and national defense.With the development of radar technology,radar imaging and radar target recognition technology emerge as a more refined and intelligent information processing technology.Synthetic aperture radar imaging technology can acquire the high-precision two-dimensional image of targets.By using the target moving information or multi-antenna echoes,the threedimensional(3-D)image can be obtained,which contains more 3-D structure information.The physical features,such as moving and structure features,are the typical features of the space target.In radar automatic target recognition technology,stable and iconic features are extracted and selected from echoes,and the target classification algorithm is used to identify the target category.Therefore,the study on 3-D imaging and physical feature-driven recognition of space target is of great significance to the construction of defense systems and space surveillance.This dissertation is focus on two typical space targets,the space coneshaped target and the space orbit target,and researches on the three-dimensional imaging,micro-range feature extraction,target recognition of the space cone-shaped target,and space orbit target recognition.The main contents are summarized as follows:1.Three-dimensional imaging method for the space cone-shaped target.Aiming at range migration caused by micro-motion and low imaging quality caused by cross-terms and sidelobes,in the third chapter of this dissertation,a three-dimensional imaging method based on smoothed Lv distribution(SLVD)is proposed to obtain the 3-D image of the space coneshaped target.Firstly,the selection criterion for best imaging time based on the timefrequency moment is proposed.When the time-frequency moment is the smallest,the range derivative of the scattering point is the smallest,which can reduce the influence of range migration.Then,range,Doppler frequency,and Doppler chirp rate form the threedimensional imaging space.The Khatri-Rao product operation is operated on the Doppler center frequency-chirp frequency(CFCR)representation and the range-Doppler(RD)image to obtain a three-dimensional image.A short time window process reduces the influence of range migration by reducing the imaging dwell time.SLVD is proposed to obtain the CFCR representation by expressing the Lv distribution(LVD)in a convolution form and introducing a centroid frequency window.The proposed SLVD can improve the energy concentration and parameter estimation accuracy of the CFCR representation,and improve the quality of 3D imaging.After scaling,the three-dimensional Cartesian coordinates of the scattering centers,together with the target size and structure,can be obtained.Numerical analysis experiments verify that the proposed SLVD can suppress cross-terms and sidelobes,and the energy of the obtained CFCR representation is more concentrated.Experiments based on electromagnetic simulation data confirm the effectiveness of the proposed 3-D imaging method.2.Space cone-shaped target recognition.By extracting the micro-range features of scattering centers,typical physical features of space cone-shaped targets,such as target size and micromotion parameters,are extracted.These typical features can train a classifier to achieve target recognition.In the fourth chapter of this dissertation,two important researches on space cone-shaped target recognition,the micro-range estimation and the classifier design,are studied.(1)Aiming at the problems of weak scattering centers’ extraction and wrong association in micro-range extraction,a micro-range feature extraction method based on trajectory association is proposed.Firstly,the orthogonal matching pursuit algorithm is used to reconstruct the micro-range trajectories of weak scattering centers,and the adaptive Kalman filter is used to associate trajectories and to filter out the spurious trajectories caused by noise.A wrong association correction mechanism is proposed to detect the wrong association,and random sample consensus algorithm is used to generate a new track for correcting the wrong association.The windowing technique is also used in the trajectory correction strategy to improve computational speed.Experiments verify that the proposed method can realize the micro-range curve extraction of weak scattering centers.Compared with the existing methods,the micro-range estimation accuracy is higher,and the time consumption is lower.(2)Aiming at the problem that there is no training data and prior information for the false target,and the number of space cone-shaped target training samples is small,a space cone-shaped target recognition method based on the mixed membership function is proposed.In this method,physical features,such as target size and micro-motion parameters,are estimated from micro-ranges,and a light-weight mixed membership function is designed as a classifier.The membership function only needs samples of space coneshaped targets to learn its fuzzy patterns without false targets.The mixed membership function is constructed by multiple membership functions to fit the feature distribution more precisely.Apart from that,the prior information of the space cone-shaped target is used to preset part of classifier parameters in order to construct the light-weight classifier.The test samples are input into the trained mixed membership function to obtain the estimated membership values.After the decision-making fusion,the target recognition results are obtained through judgment.Experiments verify that the proposed method only needs a small number of space cone-shaped targets’ training samples to realize the recognition,and the proposed recognition method has better recognition performance than existing methods.3.Space orbit target recognition.Aiming at the incompleteness of the space orbit target recognition template library,the fifth chapter of this dissertation proposes a space orbit target zero-shot recognition method combined with attribute knowledge library.This method regards classes with training samples as seen classes,and classes without training samples as unseen classes.Since the seen and unseen classes have the same attribution model,by constructing the mapping relationship between seen and unseen classes,the attribute model,which predicts the attributes of unseen samples,can be obtained.Combining the prior information of unseen classes,the identification of unseen samples can be obtained.In the proposed recognition method,the typical and easy-to-learn attributes of the space orbit target(such as: the relative position and size)based on the prior information are firstly designed to generate an attribute space.The attribute vectors of classes are defined and quantified based on the prior attribute information.Attribute vectors of all seen and unseen classes form an attribute knowledge library.Since the number of the training seen classes’ samples are limited,the attribute model is trained insufficiently.By constructing a simplified space orbit target point model,the samples,which have the same attribute space,are simulated to obtain the more sufficient attribute model.The attribute model is jointly trained by seen class training data and point model simulation data,and the mapping from samples to attribute vectors is learned by attribute label embedding method.The estimated attribute of the test samples can be estimated by the trained attribute model,and matched with the prior attribute vector of each unseen class in the attribute knowledge library to obtain the recognition result.Experiments verify the rationality of the proposed attribute space,and the samples generated by simulation can improve the attribute model learning.The proposed zero-shot recognition framework can achieve accurate recognition for unseen space orbit targets.