UWB-radar-based System For Elderly Indoor Motion Recognition And Behavioral Inference

With the heavily growth of aging,tremendous pressure has been brought to society and individuals around the world,which has given rise to the concept of ambient assisted living.The ambient assisted living system for the care of the elderly helps remote caregivers to get the information about the physical activity capabilities of the elderly and make appropriate strategies,thus to prolong the time for elderly independent living under the premise of reducing time and economic burden.As an emerging sensing technology in the civilian field,ultra-wideband radar has gradually received attentions due to its privacy protection and non-intrusive advantages.In this thesis,ultra-wideband radar is employed to detect the micro-Doppler characteristics of human activities,and different radar signal representation domains and classification technologies are used to recognize motion types.The motion recognition problem based on timeDoppler maps and time-varying range-Doppler maps is studied,and the behavior of the target is inferred by combining motion and positioning information.The main works are listed as follow:1.With the participation of four experimental subjects,an ultra-wideband radar dataset containing eight common indoor motions is established to support the performance evaluation of proposed algorithms;2.This thesis uses the form of time-Doppler maps to characterize the micro-Doppler features generated by human motions.Traditional pre-defined feature classification and advanced deep learning feature classification methods are employed to identify the motion types.The experimental results show that deep learning algorithms have more powerful feature extraction capabilities;3.A new radar signal representation domain named time-varying range-Doppler maps is proposed to overcome the limitation of incomplete information representation of time-Doppler maps.This thesis uses a convolutional auto-encoder to extract time-series features,and finally obtains the motion recognition result through the dynamic modeling of a gated recurrent unit.As a result,a recognition accuracy of 93.88% is achieved with a smaller model size and computational complexity than the time-Doppler-map-based method;4.A target positioning algorithm is proposed based on frequency filtering.Using the internal correlation between indoor areas and common behaviors,the start,progress and end of the behavior are inferred based on the combination of the positioning and motion information of the target.A UI prototype system is used to demonstrate the logical progress of behavior inference,which proves the feasibility of this method.