【作者】 李雪梅；

【导师】 秦华标；

【作者基本信息】 华南理工大学 ， 电路与系统， 2013， 硕士

【摘要】 近年来越来越多人关注自动人眼定位和人眼状态的判别，主要由于鲁棒性的人眼状态提取在很多领域都发挥着重要的作用，人眼状态为人脸表情识别、人机交互系统、视线跟踪和驾驶员疲劳检测系统提供重要的辅助信息。在驾驶员疲劳检测中，PERCLOS法是最有效的驾驶疲劳测评方法，而眼状态的准确判别是PERCLOS方法的关键问题。但是目前的系统还无法克服复杂光照的影响、头部转动等问题，因此需要对眼状态判别算法进行深入研究。为了使得系统能够全天候工作，本文使用红外主动光源进行图像照明。在此条件下，本文研究了复杂光照条件下的红外眼睛图像预处理算法，同时针对红外眼睛图像的特点，研究并提取能够对睁闭眼状态进行有效区分的判别特征，之后根据特征的特点建立起最优模型，从而实现对眼睛状态的鲁棒性判别算法。本文的主要工作如下：（1）针对眼状态中复杂光照影响的问题，本文提出将光照模型的光照不变量提取方法应用到眼状态判别的预处理上。首先提取不同的光照不变量并定性分析其效果，之后建立基于光照不变量的人眼状态判别模型，定量分析其是否能够克服光照变化，有效提高眼睛判别的准确率。实验证明，基于光照不变量的眼状态判别方法能够在一定程度上克服光照的影响，提高睁闭眼判别的准确率。（2）针对红外光照下的人眼图像的特征提取问题，首先对常用的红外眼睛特征进行分析比较，找到具有尺度不变性和旋转不变性的伪Zernike矩特征、简单而有效的复杂度特征和对光照不敏感的HOG特征，之后对三种特征进行提取和分析，最后使用SVM分类模型进行联合训练，提出了一种基于多特征决策融合的眼状态识别算法。（3）自建红外人眼样本库，对所提出的识别算法进行检测，同时将结果与目前常用的眼状态识别算法进行比较。结果表明，基于多特征决策融合的眼状态识别算法能够综合伪Zernike矩特征、复杂度特征和HOG特征的优点，将算法的总体准确率提高到91%以上，实现对红外人眼状态的鲁棒判别。更多还原

【Abstract】 In recent years, more and more people focus on automatic eye location and eye statedetection. The reason is that robust eye state recognition plays an important role in manyfields. Eye state detection provides auxiliary information for facial expression recognition,HCI system, eye gaze tracking and driver fatigue detection system. PERCLOS is the mosteffective method for driver fatigue detection, and the accuracy of eye state is the key problemto PERCLOS method. But the current systems still can’t overcome the challenge of complexillumination and head rotation. Therefore it’s necessary to go into the eye state recognitionalgorithm.This paper use active infrared sources for lighting, in order to enable the system to workall day long. In this condition, the paper study the infrared eye image postprocessor methodsunder complex illumination and then extract the distinguished features in the light of thecharacteristic of infrared image to realize the robust algorithm for eye state recognition.The main work of this paper is as follows:(1) In order to weaken the affect of complex illumination, this paper proposes the idea toapply the illumination invariant extraction methods in the postprocessor of eye state detection.First extract different illumination invariants and analysis their effect qualitatively. and thenuse PCA to reduce the dimensionality of illumination invariant, finally use LDA to build a eyestate detection model. The experimental result shows that the method can weaken the affect ofillumination and raise up the accuracy of eye state detection.(2) In order to solve the problem of eye image feature extraction under infrared illumination,the paper firstly analyzes the common used features and chooses three discriminate features:the pseudo Zernike for its scale invariant and rotate invariant, the complexity for itssignificant difference, and the HOG for its insensitive to light variation. Three features arethen joint training in Support Vector Machine model, A new method for eye state detectionbased on multi features is proposed.(3) A infrared eye sample library is built to test the proposed algorithm, and compared to othercommon used algorithm. the result shows that pseudo Zernike feature can overcome thechange of eye lead by head rotation, and after combined with complexity and HOG features, the average accuracy of proposed algorithm reaches91%, and have good robustness.更多还原