【作者】 李雪松；

【导师】 周一帆； 张志胜；

【作者基本信息】 东南大学 ， 机械工程， 2021， 硕士

【摘要】 呼吸检测技术分为接触式和非接触式,对于一些特殊的人群来说,接触式检测操作不方便,而非接触式呼吸检测,尤其是红外热像仪呼吸检测,由于它是非接触的且操作简单,在临床上对某些不方便行动的病人进行呼吸检测是一种非常有价值的检测手段,因而目前对于它的研究最多,但是利用分辨率较低的红外热像仪进行检测,尚未有成熟的实现方案。本文利用红外热像仪采集热图像,利用鼻孔中气流变化的特点作为研究对象来测量鼻子的呼吸,主要的研究内容及成果如下:在鼻子识别方面,由于目标图像是红外热图像,图像中的轮廓、细节信息比可见光图像弱,并且鼻子部位的结构信息在热图像中是无法清楚地被观察的,因此本文对图像作了预处理,例如去噪、直方图均衡化、图像锐化,这些预处理操作有效地增加了图像的轮廓、细节信息。并在后续的处理中利用人脸部位温度特征和鼻子边缘特征,结合多级大津法和Canny边缘检测算法,对面部图像进行相应的处理,多级大津法能有效地将人脸部分和背景分离,通过Canny边缘检测找到了鼻子的轮廓边缘,完成了对鼻子部位的自动识别。在目标跟踪方面,本文的跟踪目标是红外热图像中鼻子区域,该区域的特点是特征不明显且含有噪声,本文采用了基于稀疏解的粒子滤波跟踪算法。针对红外热图像特征不明显的特点,将候选目标用目标模板和补丁模板进行稀疏表示,补丁模板有效地解决了跟踪目标中出现的噪声以及低分辨率的问题,稀疏性解解决了跟踪目标特征不明显的问题。并用实验对象进行了验证,后续的跟踪目标与第一帧目标的结果相似性均值在0.998以上,跟踪效果良好。在信号提取方面,从跟踪目标中提取的呼吸信号含有较大的噪声信号,为了提高信噪比,结合两种方案,首先利用改进的帧差法识别出鼻孔区域,并计算该区域的平均灰度值。得到的波形与鼻子区域平均灰度值相比波形噪声有明显的降低。对鼻孔区域平均灰度值波形进行滤波处理,对比了几种滤波器的区别,通过对比均方误差和信噪比,选用了椭圆滤波器进行滤波,并利用实验集进行验证,得出滤波的信噪比平均值为5.58d B、均方误差平均值为0.245。更多还原

【Abstract】 Respiration detection technology is divided into contact type and contactless type.For some special patients,the contact type detection operation is inconvenient.Therefore,the respiration detection of contactless type,especially the thermal imaging camera respiration detection,which is a very valuable detection method to perform respiration detection on some patients who are inconvenient to move in clinical practice because of it’s untouchability and simplicity of operator.Consequently,there is currently the most research on it.However,there is no mature implementation plan to detect by infrared thermal imaging cameras with lower resolution.This article uses infrared thermal imaging camera to collect thermal images,and uses the characteristics of airflow changes in the nostrils as the research object to measure the breathing of the nose.The main research contents and results are as follows:In terms of nose recognition,since the target image is an infrared thermal image,the contour and detail information in the image is weaker than the visible light image,also the structural information of the nose part cannot be clearly observed in the thermal image.Therefore,this article has made a preprocessing for the image.Such as denoising,histogram equalization,image sharpening,these preprocessing operations effectively increase the outline and detail information of the image.And in the subsequent processing,using the temperature characteristics of the face part and the nose edge feature,combined with the multi-Otsu method and the Canny edge detection algorithm,the facial image is processed accordingly.The multi-Otsu method can effectively separate the face part from the background part,the contour edge of the nose is found through Canny edge detection,and the automatic recognition of the nose part is completed.In terms of target tracking,the tracking target in this article is the nose area in the infrared thermal image.The characteristics of this area are that the features are poor and contain noise.This article uses a particle filter tracking algorithm based on sparse solution.Aiming at the inconspicuous characteristics of infrared thermal image,the candidate targets are sparsely represented by target template and patch template.The patch template effectively solves the problems of noise and low resolution in the tracking target,and the sparse solution solves the problem that the tracking target features are not obvious.It was verified with experimental subjects,and the average similarity between the following tracking target and the first frame target was above0.998,and the tracking effect was perfect.In terms of signal extraction,the breathing signal extracted from the tracking target contains a larger noise signal.In order to improve the signal-noise ratio,the two schemes are combined.First,the improved frame difference method was used to identify the nostril region,and the average gray value of the region was calculated.Compared with the average gray value of the nose area,the obtained waveform has a significant reduction in waveform noise.The average gray value waveform of the nostril area was filtered,and the differences between several filters were compared.By comparing the mean square error and the signal-noise ratio,the elliptic filter was selected for filtering,and the experimental set was used to verify the filter.The average signal-noise ratio is 5.58 d B,and the average mean square error is 0.245.更多还原