基于HMM的驾驶员疲劳评估模型研究
发布时间：2019-05-29 04:52
【摘要】：驾驶员疲劳驾驶在道路交通事故发生的原因中占有绝大部分比例。鉴于视觉特征信息直观明显,易于检测,并可实现非接触性测量,因此基于视觉特征检测驾驶员疲劳状态已经成为学者们研究的热点和主流。以往研究大多采用一种或几种驾驶员疲劳时的表现特征运用贝叶斯网络、模糊推理、人工神经网络、机器视觉等开展研究,其局限性在于忽略了驾驶员精神状态的变化是一个随时间变化的过程。本文所建立的隐马尔科夫模型合理地反应了驾驶员精神状态变化与自身特征信息变化过程,它可以描述驾驶员疲劳状态的在时间上的整体非平稳性和局部平稳性,是一种较为理想的驾驶员疲劳评估模型。本文选取了20位实验对象利用驾驶模拟器模拟高速公路工况。实验过程中运用SMI-HED头盔式眼动仪采集实验中驾驶员的眼部特征信息,摄像机记录实验中驾驶员面部视频图像,利用生理参数测试仪采集驾驶员的生理信号。具体工作如下：1.在各种驾驶员疲劳评估模型中,本文重点分析了隐马尔科夫模型(Hidden Markov Models,简称为HMM)的驾驶员疲劳评估模型。选取参数PERCLOS.AECS.PERLVO作为评估驾驶员疲劳状态的参数变量,并建立对应的HMM驾驶员疲劳评估模型。利用实验数据对所建立的模型进行了训练,利用Baum.Welch算法(也称前、后向算法)和基于样本数据得到的HMM模型参数训练得到HMM驾驶员疲劳评估模型的最终参数λ=[A,B,π],使Pr[O/λ]达到最大。2.本文以生理参数仪获得的数据为基础,通过与所建立的HMM驾驶员疲劳评估模型疲劳概率及HMM经典算法Viterbi算法推断驾驶员产生观察值序列时间段内最可能的精神状态对比,验证了所建模型的合理性、准确性。3.基于实验数据,本文对所建立的驾驶员疲劳评估模型进行了详细的对比分析,并获得了相应的分析结论。结果表明,基于单参数和多参数所建立的HMM驾驶员疲劳评估模型能反应驾驶员疲劳是一个随时间变化的过程,这说明所建立的模型可以准确反映驾驶员疲劳形成的时变特征。相应模型的对比分析结果表明,基于多参数PERCLOS、 AECS、PERLVO的HMM的驾驶员疲劳评估模型更符合驾驶员真实的精神状态变化过程,且根据观察值序列得到的最可能隐藏的驾驶员精神状态序列与驾驶员真实的精神状态序列更加吻合。
[Abstract]:Driver fatigue driving accounts for the majority of the causes of road traffic accidents. Because the visual feature information is intuitionistic, easy to detect, and can realize non-contact measurement, the detection of driver fatigue state based on visual features has become the focus and mainstream of scholars. In the past, most of the previous studies used one or more kinds of performance characteristics of driver fatigue by using Bayesian network, fuzzy reasoning, artificial neural network, machine vision and so on. Its limitation is that it ignores that the change of driver's mental state is a process that changes with time. The hidden Markov model established in this paper reasonably reflects the change process of driver's mental state and its own characteristic information, and it can describe the overall nonstationarity and local stationarity of driver's fatigue state in time. It is an ideal driver fatigue evaluation model. In this paper, 20 experimental objects are selected to simulate the working conditions of expressway by driving simulator. In the course of the experiment, the SMI-HED head helmet eye movement instrument was used to collect the driver's eye characteristic information, the camera recorded the driver's facial video image, and the physiological parameter tester was used to collect the driver's physiological signal. The specific work is as follows: 1. Among all kinds of driver fatigue evaluation models, this paper focuses on the driver fatigue evaluation model of hidden Markov model (Hidden Markov Models, (HMM). The parameter PERCLOS.AECS.PERLVO is selected as the parameter variable to evaluate the driver's fatigue state, and the corresponding HMM driver fatigue evaluation model is established. The model is trained by experimental data, and the final parameter 位 = [A, B] of HMM driver fatigue evaluation model is obtained by using Baum.Welch algorithm (also known as forward and backward algorithm) and HMM model parameters training based on sample data. 蟺], so that Pr [O / 位] reaches the maximum. 2. Based on the data obtained by the physiological parameter instrument, this paper infers the most likely mental state within the time period of the observed value series by comparing the fatigue probability of the established HMM driver fatigue evaluation model with the HMM classical algorithm Viterbi algorithm. The rationality and accuracy of the model are verified. Based on the experimental data, the driver fatigue evaluation model is compared and analyzed in detail, and the corresponding analysis conclusions are obtained. The results show that the HMM driver fatigue evaluation model based on single parameter and multi-parameter can reflect that driver fatigue is a time-varying process, which indicates that the established model can accurately reflect the time-varying characteristics of driver fatigue formation. The comparative analysis results of the corresponding model show that the driver fatigue evaluation model based on multi-parameter PERCLOS, AECS,PERLVO HMM is more in line with the real mental state change process of the driver. The most likely hidden driver mental state sequence according to the observed value sequence is more consistent with the driver's real mental state sequence.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U491.254
本文编号：2487669
[Abstract]:Driver fatigue driving accounts for the majority of the causes of road traffic accidents. Because the visual feature information is intuitionistic, easy to detect, and can realize non-contact measurement, the detection of driver fatigue state based on visual features has become the focus and mainstream of scholars. In the past, most of the previous studies used one or more kinds of performance characteristics of driver fatigue by using Bayesian network, fuzzy reasoning, artificial neural network, machine vision and so on. Its limitation is that it ignores that the change of driver's mental state is a process that changes with time. The hidden Markov model established in this paper reasonably reflects the change process of driver's mental state and its own characteristic information, and it can describe the overall nonstationarity and local stationarity of driver's fatigue state in time. It is an ideal driver fatigue evaluation model. In this paper, 20 experimental objects are selected to simulate the working conditions of expressway by driving simulator. In the course of the experiment, the SMI-HED head helmet eye movement instrument was used to collect the driver's eye characteristic information, the camera recorded the driver's facial video image, and the physiological parameter tester was used to collect the driver's physiological signal. The specific work is as follows: 1. Among all kinds of driver fatigue evaluation models, this paper focuses on the driver fatigue evaluation model of hidden Markov model (Hidden Markov Models, (HMM). The parameter PERCLOS.AECS.PERLVO is selected as the parameter variable to evaluate the driver's fatigue state, and the corresponding HMM driver fatigue evaluation model is established. The model is trained by experimental data, and the final parameter 位 = [A, B] of HMM driver fatigue evaluation model is obtained by using Baum.Welch algorithm (also known as forward and backward algorithm) and HMM model parameters training based on sample data. 蟺], so that Pr [O / 位] reaches the maximum. 2. Based on the data obtained by the physiological parameter instrument, this paper infers the most likely mental state within the time period of the observed value series by comparing the fatigue probability of the established HMM driver fatigue evaluation model with the HMM classical algorithm Viterbi algorithm. The rationality and accuracy of the model are verified. Based on the experimental data, the driver fatigue evaluation model is compared and analyzed in detail, and the corresponding analysis conclusions are obtained. The results show that the HMM driver fatigue evaluation model based on single parameter and multi-parameter can reflect that driver fatigue is a time-varying process, which indicates that the established model can accurately reflect the time-varying characteristics of driver fatigue formation. The comparative analysis results of the corresponding model show that the driver fatigue evaluation model based on multi-parameter PERCLOS, AECS,PERLVO HMM is more in line with the real mental state change process of the driver. The most likely hidden driver mental state sequence according to the observed value sequence is more consistent with the driver's real mental state sequence.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U491.254
【参考文献】
相关期刊论文 前2条
1 王荣本;顾柏园;郭烈;余天洪;;基于分形盒子维数的车辆定位和识别方法[J];吉林大学学报(工学版);2006年03期
2 孙伟;张为公;张小瑞;吕成绪;陈刚;;疲劳驾驶检测方法的研究进展[J];汽车技术;2009年02期
，本文编号：2487669
本文链接：https://www.wllwen.com/kejilunwen/daoluqiaoliang/2487669.html
