基于机器视觉的人体状态监测关键技术研究

发布时间：2018-07-13 10:06

【摘要】：我国青少年近视发病率逐年递增,其中不良坐姿是引起青少年近视的原因之一。在道路交通事故中,很多是由驾驶员注意力分散和疲劳驾驶引起的。针对以上两个问题,本文基于机器视觉对坐姿状态和疲劳状态的监测进行了研究。在人脸检测的基础上针对两种不同场景分别提出了两种不同的坐姿行为监测方法和疲劳监测方法,包括基于人脸肤色统计的坐姿判别、基于区域关键特征点匹配的头部状态判别、基于嘴巴活动区域融合边缘统计的打哈欠判别、基于人眼与瞳孔检测的闭眼判别四种有效的监测方法。并设计了"不良坐姿行为监测系统"和"辅助驾驶中头部状态与疲劳监测模拟系统"。首先,在不良坐姿行为监测系统中,改进了基于RGB彩色视频的人脸检测方法:采用基于肤色的人脸检测,能够有效地减少人脸误检率;采用最大单目标的人脸检测方法,能检测出主要目标的人脸;提出了检测窗口尺度自适应的单目标人脸检测方法,利用前一帧中所检测的单个主体目标的尺寸,自适应地调节检测窗口的变化范围,使得检测速度得到了大大的提升。其次,提出了一种基于人脸肤色统计的坐姿监测方法,该方法先根据检测出的人脸框规划出左、中、右三个肤色判别区域,然后通过对比这三个区域的肤色与正确姿态下的情况,来判别坐姿靠左/靠右;通过统计对比当前与正确姿态下人脸框内的肤色面积来判断靠前/靠后。实验表明,在避免背景为肤色的前提下该方法对左右的检测正确率为100%,对前后的检测正确率为97.3%。针对驾驶员不良头部状态和疲劳状态的监测,提出了基于主动红外视频的三种判别方法:(1)基于区域关键特征点匹配的头部状态判别方法,分析正确姿态下的模板与实时监测区域中的三对最佳匹配的SURF特征点的位置,从而判断出当前头部状态正确与否;(2)基于嘴巴活动区域融合边缘统计的打哈欠判别方法,统计表明,嘴巴几乎活动在人脸检测框中下端的区域内,所以打哈欠的判别主要是在人脸框上规划出嘴巴活动区域,然后统计在该区域内Prewitt与Canny算子融合边缘纵向投影比判断嘴巴开合程度,并通过开合程度来判断打哈欠的状态;(3)基于人眼与瞳孔检测的闭眼判别方法,根据人脸框规划出眼睛的大概区域,在此基础上能够更好地定位眼睛,大大减少了全局检测带来的误检,同时提高了检测效率和正确率,然后将检测出来的眼睛适当放大并做霍夫圆检测,能够通过霍夫圆的存在与否判断眼睛开合状态。实验结果表明,设计完成的"辅助驾驶中头部状态与疲劳监测的模拟系统"中头部状态判别、打哈欠判别和眼睛疲劳判别模块的正确率分别为:98.9%、100%、97.8%。
[Abstract]:The incidence of myopia is increasing year by year in China, among which poor sitting posture is one of the causes of juvenile myopia. In road traffic accidents, many of them are caused by driver's distraction and fatigue driving. In view of the above two problems, this paper studies the monitoring of sitting posture and fatigue based on machine vision. On the basis of face detection, two different sitting behavior monitoring methods and fatigue monitoring methods are proposed for two different scenes, including sitting-posture discrimination based on face skin color statistics. There are four effective monitoring methods: head state identification based on matching key feature points of region, yawning discrimination based on fusion edge statistics of active areas of mouth, and closed eyes discrimination based on human eyes and pupil detection. The monitoring system of bad sitting behavior and the simulation system of head state and fatigue monitoring in auxiliary driving are designed. First of all, the method of face detection based on RGB color video is improved in the bad sitting behavior monitoring system: the face detection based on skin color can effectively reduce the false detection rate of face, and the maximum single target face detection method is adopted. A single target face detection method based on window scale adaptive detection is proposed, which adaptively adjusts the range of detection window by using the size of a single object detected in the previous frame. The detection speed is greatly improved. Secondly, a sit-down monitoring method based on face skin color statistics is proposed, in which the left, middle and right skin color discriminant regions are planned according to the detected face frame. Then, by comparing the skin color of the three regions with the correct posture, the left / right side of the sitting position is judged, and the forward / backward position is judged by statistical comparison of the skin color area in the face frame under the current and the correct posture. The experimental results show that the accuracy rate of this method is 100 for left and right and 97.3 for front and rear without background skin color. Aiming at the monitoring of bad head state and fatigue state of driver, three discrimination methods based on active infrared video are proposed: (1) head state discrimination method based on regional key feature point matching; The position of three pairs of best matched SURF feature points in the real time monitoring region and the template under the correct posture are analyzed to determine whether the current head state is correct or not. (2) the yawning judgment method based on the edge statistics of the mouth active region fusion. Statistics show that the mouth almost moves in the lower end of the face detection box, so the yawning is mainly based on the planning of the mouth activity area on the face frame. Then, the longitudinal projection ratio of Prewitt and Canny operator is counted to judge the degree of mouth opening and closing, and the state of yawning is judged by the degree of opening and closing. (3) the method of judging the closed eyes based on the detection of human eyes and pupils. According to the face frame, the eye area can be mapped out, and the eyes can be located better, which greatly reduces the error detection brought by global detection, and improves the detection efficiency and accuracy. Then the detected eyes are enlarged properly and the Hoff circle is detected, which can judge the opening and closing state of the eyes by the existence or not of the Hoff circle. The experimental results show that the correct rates of head state discrimination, yawning discrimination and eye fatigue identification module in the "Simulation system for head condition and fatigue Monitoring in Auxiliary driving" are: 98.9% and 100% and 97.8%, respectively.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】