Research On Micro-Doppler Features Based Human Behavior Classification Methods

The Doppler effect caused by the relative motion(i.e.,micro-motion)of the target different structural components is called the micro-Doppler effect.Researchers can study normal human behavior by micro-Doppler features to effectively detect abnormal behav-ior or threat.Therefore,the research on human behavior classification and related issues on micro-Doppler features has broad application prospects in the fields of medical rehabil-itation,life rescue,public safety,and medical monitoring,and has been widely concerned by researchers.Because of the existing shortcomings of the micro-Doppler signal pro-cessing methods,the research of this thesis mainly focuses on three aspects,including noise suppression,feature extraction,and target classification.Based on the existing the-ories and methods,this thesis intends to further expand the theoretical basis of the human behavior classification based on the micro-Doppler features and designs more universal,robust,and efficient methods for the applications of the micro-Doppler features to provide the theoretical and technical support.The algorithms and models proposed in this thesis are all verified to have excellent performances by the theoretical analysis,the simulation,and the real experiments.The specific research contents are summarized as follows:1?The relationship between target and micro-Doppler effect is studied.Aiming at the problem of insufficient samples of micro-Doppler radar measurement,a method based on the motion capture data for producing radar echo of human behaviors is proposed.The method uses the motion capture data to model human behaviors,and through the model and the position relationship between human and radar transmitter and receiver,the radar echo is simulated.The consistency of the feature parameters of the simulation samples and radar measurement samples verify the effectiveness of the method.2?The problem of signal denoising using empirical mode decomposition and its derivative algorithms is studied.To solve the problem of the empirical selection of the decomposed component in signal denoising,a direct extraction rule of the decomposed component is proposed by the mathematical expectation comparison of the frequency be-tween decomposed components and the target to avoid the mistakes caused by selecting the decomposed component by experience.Simulated and measured results show that the directly extracted decomposition component is the efficient signal component.Aiming at the problem of the features broken in the sifting process of signal denoising,an algorithm based on the learning mechanism of the wavelet dictionary is proposed to reconstruct the missing or interrupted signal components.Simulated and measured results show that the algorithm can remove unwanted information while reconstructing the signal component destroyed by the sifting process.Comparing to the sum method of efficient components,our algorithm has a stronger denoi sing ability.3?The problem of multi-component micro-Doppler frequency estimation is stud-ied.Aiming at the defect that the existing multi-component frequency estimation meth-ods rely heavily on the reliable prior model,this thesis proposes a modified sparse multi-component frequency estimation algorithm,which no need the prior knowledge of the signal model.The algorithm searches coarse signal components by the discrete Fourier transform dictionary and refines signal components by the mapping on the best spatial regression kernel.In the implementation process,the algorithm utilizes the alternating di-rection multiplier method to improve the learning efficiency of the discrete Fourier trans-form dictionary.If the signal is unbiased,the best spatial regression kernel can be directly calculated by the coarse signal component.If the signal is biased,the components can be partitioned via the m-fold cross-validation in time series,and the sub-kernel of spa-tial regression corresponding to each fold is calculated,the spatial regression kernel is calculated via these sub-kernels for the coarse signal component mapping.The fold num-ber of cross-validation is addressed by the granularity selection rule,which can balance precision and processing time.Simulated and measured results demonstrate that the al-gorithm can estimate linear or sinusoidal frequency modulation components without any prior knowledge of signal mode,and obtain a better resolution than time-frequency image.4?The classification problem of human behavioral samples based on the micro-Doppler radar is studied.Aiming at the issue of poor classification performance caused by time-frequency analysis,based on the reason that signal autocorrelation function in the time domain can be mapped to the time-frequency image in the time-frequency domain,the feature extraction of the autocorrelation function of radar echo in the time domain is transformed into a convex optimization problem by introducing l2 norm.An itera-tive convolution neural networks framework is designed to solve the convex optimization problem,which can automatically extract the micro-Doppler features.Furthermore,a classification algorithm is proposed.In this algorithm,the automatically extracted micro-Doppler features are used as the input of classifiers to realize the classification to solve the problem of poor classification performance caused by the error of the time-frequency analysis.Simulation and experimental results show that the classification performance of this algorithm is more robust than that of the methods using the features of the time-frequency image for classification.One activity produces different micro-Doppler shifts at various aspect angles in three-dimensional space.Aiming at the problem of the aspect angle effect on the classification performance of micro-Doppler features,different clas-sification algorithms are used to classify data samples,which are collected at different aspect angles,and their classification performance is studied.Simulated and measured results show that different features have a significant effect on classification performance,and different classifiers have a small effect on classification performance.The closer the aspect angle of training samples and that of testing samples,the better the classification performance.