基于视频序列的人眼检测与跟踪的研究及应用

发布时间：2018-01-13 11:13

本文关键词：基于视频序列的人眼检测与跟踪的研究及应用　出处：《太原理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着互联网的快速发展和成像设备的迅速普及,图像和视频成为信息的主要传播方式并爆炸式的充斥在我们的日常生活中。为了让计算机主动地分析和处理获取到的信息,计算机视觉(Computer Vision)受到越来越多人的关注。由于人眼涵盖了包括性别、身份、表情在内的多种信息,人眼检测和跟踪是人脸识别、心理测试和疲劳检测等应用的核心技术,吸引了众多研究人员致力于人眼定位的探索与研究,使其成为计算机视觉领域的重要课题。本文基于视频序列研究了人眼检测和跟踪,并将其应用于疲劳检测中。具体研究工作如下:(1)基于Adaboost算法研究了人眼检测。当采用Adaboost算法训练分类器时,随着训练迭代次数的不断增加,一些处于正负样本交界处的临界样本可能由于多次错分而导致权重越来越大,使算法重心逐渐转移到其身上而引发权重失衡,结合漏检正样本的价值要远远大于误检负样本,本文改进了Adaboost算法,并针对人眼检测中误检率较高的缺陷提出了基于改进Adaboost算法的三层结构人眼检测方法。(2)基于Kalman滤波器和CamShift算法研究了人眼跟踪。CamShift算法在目标跟踪过程中缺乏预测模块,而且当目标颜色与背景颜色相近或目标出现大面积遮挡时容易丢失目标,且一旦丢失便无法恢复对目标的正确跟踪。本文将Kalman滤波器引入到CamShift算法并对其进行了改进,并针对人眼跟踪受局部搜索窗口大小的影响提出一种基于Kalman滤波器和改进CamShift算法的双眼跟踪方法。(3)当人处于疲劳状态时人眼变化最为明显,本文在人眼检测和跟踪的基础上,基于人眼状态进行了疲劳检测。具体分为两部分:首先基于图像处理方法分析研究了人眼状态,采用Ostu方法对定位到的人眼区域进行阈值化处理;接着采用Freeman链码方法提取二值图像中人眼外围轮廓,再采用最小二乘椭圆拟合方法对人眼外围轮廓进行拟合,最后依据椭圆的长短轴比值分析判断人眼状态。其次引入了PERCLOS和眨眼频率两个疲劳判定标准,通过统计一段连续时间内人眼的PERCLOS参数和眨眼频率进行疲劳检测。
[Abstract]:With the rapid development of the Internet and the rapid spread of imaging equipment. Images and videos become the main means of information dissemination and explosion in our daily life. In order to enable the computer to actively analyze and process the information obtained. Computer vision (computer vision) has attracted more and more attention. Because the human eye covers a variety of information, including gender, identity, expression and so on. Human eye detection and tracking is the core technology of face recognition, psychological testing and fatigue detection, which attracts many researchers to explore and study human eye location. It has become an important subject in the field of computer vision. This paper studies human eye detection and tracking based on video sequences. The specific research work is as follows: 1) Human eye detection is studied based on Adaboost algorithm. When using Adaboost algorithm to train classifier. With the increasing number of training iterations, some critical samples at the junction of positive and negative samples may have more and more weight due to multiple errors, which makes the weight unbalance. Considering that the value of missed positive samples is much greater than that of false negative samples, this paper improves the Adaboost algorithm. Aiming at the defect of high false detection rate in human eye detection, a three-layer structure human eye detection method based on improved Adaboost algorithm is proposed. Based on Kalman filter and CamShift algorithm, the human eye tracking. CamShift algorithm lacks prediction module in the process of target tracking. And it is easy to lose the target when the color of the target is close to the background color or there is a large area occlusion of the target. Once lost, the correct tracking of the target can not be recovered. In this paper, the Kalman filter is introduced into the CamShift algorithm and improved. A binocular tracking method based on Kalman filter and improved CamShift algorithm is proposed for human eye tracking, which is influenced by the local search window size. The change of human eyes is most obvious when people are in fatigue state. On the basis of human eye detection and tracking, this paper carries out fatigue detection based on human eye state. It is divided into two parts: firstly, the human eye state is analyzed and studied based on image processing method. The Ostu method was used to thresholding the human eye region. Then Freeman chain code method is used to extract the peripheral contour of human eyes in binary images, and the least square ellipse fitting method is used to fit the contour of human eyes. Finally, according to the long and short axis ratio analysis of the ellipse to judge the human eye state. Secondly, the PERCLOS and blink frequency fatigue criteria are introduced. Fatigue was detected by counting the PERCLOS parameters and blink frequency of human eyes for a continuous period of time.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】