Research On Human Motion Recognition Theory And Technologies Based On Radar System

Human motion recognition has attracted great interest for different purposes,such as surveillance,search and rescue operations,smart home,and senior people care in assisted living facilities.Various methods for human motion recognition have been proposed.Compared to traditional contacted sensors and video-based sensors,radar-based human motion recognition may complement them because of its potential for high accuracy,robustness,adaptability,penetrability and privacy preservation.It shows significant practical implication in both military and civil area.However,there are still some issues and problems to be improved and solved.First,most studies focused on the classification of human motions with strong regularity and good stability while others only paid attention to fall detection leading to a limited application.In addition,among the existing feature extraction algorithms of human motions,the performance of envelope detection algorithm based on spectrogram depends on exmpirical threshold and is susceptible to noise interference.And the feature vectors extracted by the ones based on data dimensionality reduction present less relationship with physical characteristics of human motions.Further,current work focused on motion recognition in a laboratory environment,whereby the different activities are recorded as separate and individual snapshots.Practical applications would need to deal with continuous human motion recognition in real-living conditions,where the human subject monitored can perform activities one after another with unknown durations and trasitions in between.In this thesis,ultra wideband-impulse radar(UWB-IR)radar and frequency-modulated continuous wave(FMCW)radar system are used to consider a fusion of multi-domain features including time,range and radar-cross-section(RCS).Furthermore,weighted range-time-frequency transform(WRTFT),sparse representation and dynamic range-Doppler trajectory(DRDT)algorithms are proposed for non-rhythmic motion recognition,multi-domain feature extraction and continuous human motion recognition,respectively.The contributions of this thesis can be summarized as below:1.In daily life,there exists various kinds of human motions with widely different characteristics and meanwhile they also exhibit some clustering features,which make it difficult for recognition.In this thesis,a weighted range-time-frequency transform method is introduced for comprehensive human motion recognition,including the largest number of motion types ever studied with an UWB-IR radar system.First,in the pre-screening layer,information in the time-range domain is used to distinguish in-situ motions and non-in situ motions with binarization algorithm.According to different kinds of human motions,the weighted rangetime-frequency transform method is proposed to obtain corresponding high-resolution spectrograms.Then,physical empirical features are extracted from in-situ motions,while principal component analysis-based features are extracted from non-in situ motions.Furthermore,the recognition performance of different classic machine learning algorithms is discussed.Finally,the recognition of twelve non-rhythmic human motions is achieved.2.Most researches utilize envelope-based feature extraction method based on timefrequency spectrogram in radar-based human motion recognition.However,it has been proven easy to be affected by empirical threshold and noises.Therefore,a sparse-representation-based feature extraction method in both time-frequency and range-Doppler spectrograms is proposed implemented on a FMCW radar system considering a fusion of multi-domain features.First,spare representations in both time-frequency and range-Doppler domains of received signals are achieved through sparse dictionary,respectively.Then,time-Doppler and range-Doppler trajectories can be obtained by orthogonal matching pursuit(OMP)and 2-D OMP algorithms.Further,the comparision between spectrograms of sparse recovered signals and original ones demonstrated that it allows to use sparse vector or matrix with few elements to reserve most effective information.Experiments are conducted and proved its feasibility and robustness in fall and fall-similar motion recognition.3.Recent studies focus on recognition of snapshot human motions.In this thesis,a novel dynamic range-Doppler trajectory method based on FMCW radar system is proposed to recognize continuous human motions with various conditions emulating real-living environment.This method can separate continuous motions and process them as single events.First,range-Doppler frames consisting of a series of range-Doppler maps are obtained by sliding-window fast Fourier transform(FFT)from the backscattered signals.Next,the DRDT is extracted from these frames to monitor human motions in time,range,and Doppler domains in real time.Then,a peak search method along with a voting method is applied to locate and separate each human motion from the DRDT map.Next,range,Doppler,RCS and dispersion features are extracted and combined in a multi-domain fusion approach as inputs to a machine learning classifier.Finally,continuous human motion recognition can be achieved when arranging a sequence of individual snapshot recognitions in chronological order.Extensive experiments have been conducted to show its feasibility and superiority in two or even more continuous human motion recognition.Furthermore,its performance even in various conditions of distance,view angle,direction,and individual diversity is discussed.