Non-contact Respiratory State Assessment Based On Vision And Machine Learning

The respiratory state can reflect a person's physiological and psychological conditions,which is helpful for screening,diagnosis and prognosis assessment of related diseases.This thesis carries out studies on the non-contact respiratory state assessment.Specifically,this thesis adopts visual sensors to extract respiratory signals,utilizes a variety of signal processing technologies to process the original signals,and uses machine learning to classify respiratory patterns.Based on this framework,two non-contact breathing monitoring systems were developed.In addition,we have creatively developed a method to recognize deep breathing when the human body walks forward.The main contents and innovations can be summarized as follows.(1)A homemade respiratory monitoring system based on color camera and marker tracking,combined with the temporal filtering and the translational cross points algorithm,is developed to realize low-cost and non-contact respiratory state assessment.The experimental results demonstrate that the root mean square error for respiratory rate estimation is 3.29 bpm and F1 for four breathing patterns classification is 89.7%.(2)A depth camera collaborated with human joints tracking is utilized to realize automatic tracking of the region of interest in depth videos,and a respiratory simulation model together with an improved recurrent neural network is proposed to realize the automatic classification of six breathing patterns.The experimental results demonstrate offline classification F1 with 94.8% and real-time classification F1 with 86.5%.(3)To tackle the problem that existing non-contact breath detection methods fail when the human body is walking,we innovatively extract breathing-related signals from multiple body regions,and adopt graph signal processing technology for spatial-temporal filtering.Finally,the established recognition model first realizes deep breath recognition when human moves around.The ablation experimental results prove that the proposed method increases F1 by 16.7%.