Research On Target Classification Technology Of Ground Reconnaissance Radar

With the development of modern military electronic technology,modern wars have evolved into high-tech information warfare and electronic warfare.Real-time monitoring and processing of dynamic information on the battlefield is an important factor related to the victory or defeat of war.As a result,the function of battlefield surveillance radar can no longer be satisfied with only detecting the position and velocity of the target.It is of great significance for judging target a friend or foe.Therefore,radar target recognition technology came into being.In view of the high manufacturing cost and large volume of high-resolution broadband radar,most of the active ground surveillance radars are low-resolution radar.It is significant to study the low-resolution ground surveillance radar automatic target recognition technology.The thesis first introduces the preprocessing process,feature extraction methods and theoretical basis of commonly used classifiers for ground surveillance radar echo signals.Then analyze the characteristics of the actual collected human and car echo data,perform Fourier transform on the radar echo data,propose time domain echo structure feature definition and the frequency spectrum structure feature definition,and statistics the distribution of the defined features.Subsequently,the extracted combined features were classified using K-nearest neighbor classifier,SVM classifier,and BP neural network classifier.On this basis,the effectiveness of three hyperparameter optimization methods such as grid search,random search,and Bayesian optimization is verified,and the performance of different classifiers and different hyperparameter optimization methods is compared.Furthermore,the Convolutional Neural Network(CNN)classification and recognition technology is adopted in this paper to simplify the two-dimensional CNN structure applied in the field of image recognition and make it applicable to the one-dimensional data flow of lowresolution radar.On this basis,the echo power spectrum,amplitude spectrum and power transformed amplitude spectrum were combined into three channels of CNN input,and a feature fusion based one dimensional convolutional neural network(1-D CNN)target classification recognition method was proposed,which has better performance than the traditional classifier.In addition,the paper compares the impact of different network hyperparameter configurations on the recognition performance of CNN classifiers.Bayesian optimized hyperparameter optimization methods are used to perform hyperparameter optimization,which overcomes the need to set hyperparameters based on experience and improves the ability of classifiers to build automatically during classifier training and construction.Furthermore,Autoencoder is used to reduce the dimension of features.Finally,this paper uses the built FPGA + DSP as the core architecture of ground surveillance radar signal processing board platform to complete the pre-processing process of down conversion and pulse compression of the echo signal on the FPGA side.The phase coherent accumulation,object detection and object classification recognition modules are realized on DSP.The results of target sample database test and field experiments show that the designed target classification recognition system can achieve the correct classification recognition rate of more than 93%.