Research On Automatic Target Recognition Of LSS Targets Based On Multi-dimensional Feature Fusion

The detection and identification of Low altitude,Small RCS and Slow speed(LSS)targets such as unmanned aerial vehicles(UAVs)has been paid more and more attention due to their wide application and serious potential threats.This paper studies the real-time automatic target recognition(ATR)method of LSS targets based on multi-dimensional feature fusion and multi-layer classifier with directed acyclic graph.This paper innovatively proposes a complete LSS automatic target recognition method and a complete automatic target recognition chain,and carries out experiment and analysis on the measured LSS target data collected by the holographic staring radar system.The ATR method of LSS targets consists of two key steps: the multidimensional feature selection and extraction of different targets and the design of the multi-layer classification.The complete ATR chain includes the pre-processing of the measured data,the extraction of multi-dimensional features such as the RCS,micro-Doppler and motion,the design of the multi-layer classifier based on directed acyclic graph,and the training and prediction of the whole ATR method for the input target data.The main work is shown as follows:(1)In the paper,the radar echo signals of three typical LSS targets of multi-rotor UAV,helicopter and bird are modeled and simulated in mathematics respectively.The influence of the number,size and rotational frequency of aircraft’s rotors and the size and flapping frequency of bird’s wings on the radar echo is analyzed through mathematical modeling and simulation experiment.Then the simulated echo modulation models of three typical LSS targets are compared with the measured data,and analyze the modulation effects of the motion and micromotion of the typical LSS targets and their moving parts from theoretical analysis and practical measurement,providing theoretical support for subsequent feature extraction and target recognition.(2)The paper introduces in detail multi-dimensional feature selection and extraction.Extract effective features from measured target data from micro-Doppler dimension,RCS dimension and motion dimension.For the micro-Doppler dimension,the theoretical analysis of micro-doppler instantaneous frequency,joint time-frequency analysis,the generation and preprocessing algorithm of doppler modulation spectrum and the feature extraction algorithm are emphatically studied.The preprocessing operation is to optimize the frequency contents of the Doppler modulation spectrum,and to highlight the weak Doppler side-frequency information.It mainly includes regularization processing of Doppler modulation spectrum and main frequency alignment,etc.Then,maximum detection and micro-Doppler feature extraction are carried out for the preprocessed Doppler modulation spectrum.Then feature selection and extraction of RCS dimension and motion dimension based on track are carried out.We can extract six effective features from three dimensions respectively,fed into classifier for LSS target recognition.(3)For the typical LSS targets such as helicopters,rotary-wing UAVs,fixed-wing UAVs,and birds,we design automatic target recognition and classification method based on multi-dimensional feature fusion and multi-layer classifier based on directed acyclic graph.A multi-layer fusion recognition classifier with multi-dimensional features is designed.Then propose two kinds of classification criteria such as multi-dimensional feature fusion classification criteria and multiple equal-weight single-dimension classifiers classification criteria.According to the design of multi-layer classifier,two kinds of classification criteria,three kinds of nonlinear classifiers such as K nearest neighbor(KNN),support vector machine(SVM)and back propagation(BP)neural network and six effective identification features extracted from three dimensions are analyzed on the experiments of target classification and recognition.The experimental results compare the influences of number of input features,the characteristics of different dimensions,three kinds of classifiers and two classification criteria on the recognition accuracy of the classification method.Finally,the experimental results show that the classification method based on multi-dimensional feature fusion based on KNN classifier is the optimal classification and recognition method for typical LSS targets,whether classification accuracy or classification time,compared with other design schemes,and obtaining the final recognition accuracy 97.62% of five kinds of typical LSS targets.Finally,the proposed recognition method and experimental results obtained in this paper are compared with other existing recognition schemes at home and abroad.The results show that the ATR method proposed in this paper has excellent recognition effects for the LSS target recognition,solving the problem of LSS targets recognition in complex natural environment.Finally,the paper summarizes the main work,and points out the research direction near future and the problems needed to further study and solve.