Research On Classification Methods Of Ground Targets Based On Micro-doppler Effect And Radar Equipped On The UAV

When many objects in nature are in motion,in addition to the translation of the target subject,they also have micro-motions such as vibration,swing and rotation which will result in additional modulation around the Doppler frequency.This phenomenon is called the micro-Doppler effect.As a unique feature of targets,the micro-Doppler effect can be used as the basis for target recognition.With the progress of modern flight control technology,UAVs have been further developed and are widely used for ground target monitoring and reconnaissance.The premise of these applications is that UAV-based radars can accurately classify ground targets.Therefore,this thesis focuses on the micro-Doppler signal modeling of three typical ground targets including wheeled vehicles,tracked vehicles and pedestrians,extraction of micro-Doppler features with high discrimination,accurate and intelligent classification of ground targets using UAV-based radars.Main research works are as follows:1.Based on micro-motion models of vibration,rotation and swing,the micro-Doppler signal models of UAV-to-ground targets including wheeled vehicles,tracked vehicles and pedestrians are established respectively.Mathematical expressions of related micro-Doppler modulation are derived.Simulation results show that the micro-Doppler modulation differences among the three ground targets can be used as the basis for classification and recognition.Based on this,for the two typical ground vehicles,corresponding Doppler signals are transformed using the bispectrum and singular value spectrum respectively.Specific micro-Doppler features are defined,extracted and sent to the support vector machine for training and classification.Since the third-order cumulant of ground clutter and noise is zero,as the one-dimensional Fourier transform of the third-order cumulant,diagonal slice of the bispectrum can effectively suppress ground clutter and noise.Simulation results indicate that when the SNR(Signal-to-noise Ratio)of Doppler signals is only 15 d B,the recognition rate of the two vehicles of the bispectrum analysis method has exceeded 92%.Compared with bispectral analysis method,the singular value decomposition method further separates specified micro-Doppler components through specific singular values and singular vectors,not only using the amplitude and energy information of micro-Doppler signals,but also considering the micro-motion structural differences.By making the singular values representing ground clutter and noise equal zero,the singular value decomposition and reconstruction method can effectively remove clutter interference.Even if the SNR is only 5 d B,the recognition rate of the singular value decomposition and reconstruction method has been close to100% while the recognition rate of the bispectrum analysis method is only 81%.2.Micro-Doppler modulation of wheeled vehicles and pedestrians is highly similar to each other.Therefore,a compressed sensing method which can separate micro-Doppler signal components in the frequency domain,refine similar frequency spectrums and enhance micro-Doppler features is proposed.First of all,the principal component analysis method is adopted to suppress the ground clutter.Then,the Doppler signal is sparsely represented based on Fourier basis and micro-Doppler signals after the random projection operation are reconstructed using the orthogonal matching pursuit algorithm.Micro-Doppler components are separated in the frequency domain.Finally,micro-Doppler features with high discrimination are extracted from refined spectrums of reconstructed signals and sent to the GA-BP(Genetic Algorithm-Back Propagation)neural network which combines the genetic algorithm and the error-based back propagation algorithm and has more advantages than support vector machines when solving nonlinear multi-classification problems to classify pedestrians and ground vehicles.When the SNR is15 d B,the recognition rates of the bispectrum analysis method and the singular value spectrum method are only 69.13% and 86.22% respectively while the recognition rate of compressed sensing method is 89.67%.3.The compressed sensing method refines the frequency spectrum and reduces the similarity between frequency spectrums of the wheeled vehicle and the pedestrian,but does not make full use of micro-motion structural differences of two vehicles.Therefore,an improved ensemble empirical mode decomposition method that can not only use the micro-motion structural differences,but also accurately divide the frequency spectrum is proposed.It can adaptively determine the standard deviation of added Gaussian white noise and decomposition times of the ensemble empirical mode decomposition algorithm according to the standard deviation of noise and small high-frequency components in the Doppler signal.From high frequency to low frequency,the micro-Dopper signal components are decomposed in turn.A better decomposition effect is achieved with a smaller amount of calculation.The micro-Doppler box dimension,information dimension and grid dimension are extracted from spectrums of intrinsic mode functions to form a three-dimensional fractal feature space with high discrimination.The GA-BP neural network is used for classification and recognition.When the SNR is only 15 d B,the recognition rate still reaches 91.27%,higher than existing classification methods.4.Target classification methods based on machine learning all need to define specific micro-Doppler features according to the signal processing results and classification situations,which are not adaptive,and the classifiers are not uniform.Therefore,a method for intelligent classification of UAV-to-ground targets based on deep convolutional neural networks is proposed.Firstly,the singular value decomposition and reconstruction method is adopted to suppress ground clutter and noise in Doppler signals.Secondly,time-frequency transformation and binarization operation are performed on the preprocessed Doppler signals to improve the contrast of the micro-Doppler envelope in time-frequency images.Finally,based on the Alex Net network,a certain proportion of binary time-frequency images of Doppler signals are randomly selected as training samples to perform migration learning and update the network weight coefficients.The input of the Alex Net network is the time-frequency image and the output is directly the classification result.Not only can complex feature extraction process be avoided,the multi-layer structure of the network can also learn more micro-Doppler features independently.When the SNR is only 15 d B,the recognition rate of this method has exceeded 97%.When the SNR is lower than 5d B,the recognition rate still exceeds 70%.5.A set of UAV-based radar classification and recognition experiment system for ground targets was constructed.Firstly,a pedestrian posture recognition experiment was carried out using a millimeter wave radar,and the Alex Net network was used to learn the preprocessed Doppler signals to classify the three postures including stepping,walking and jogging.The recognition rate can reach 96.67%,indicating that the designed radar prototype can effectively reflect the micro-movement of targets.Then,three ground targets including pedestrians,wheeled vehicles and tracked vehicles are detected when the UAV is hovering or moving linearly.Based on the singular value decomposition and reconstruction method,the measured Doppler signals are preprocessed to suppress clutter and noise.The preprocessed binary time-frequency images are sent to the Alex Net network for target classification.The overall recognition rate is 96%.