【作者】 沈志熙；

【导师】 黄席樾；

【作者基本信息】 重庆大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 智能车辆作为智能交通系统的关键技术和重要组成部分,被认为是解决各种交通问题的一个有效途径。智能车辆的研究涉及图像处理、模式识别、人工智能、自动控制、传感器技术等多个学科的理论与技术,集成了信息科学与人工智能技术的最新成果,具有重要的实用价值和学术理论价值。智能车辆是一个集环境感知、规划决策、操作控制等功能于一体的智能体。在智能车辆诸多研究任务中,对环境信息的感知是所有工作的基础和核心,基于视觉的信息感知是车辆理解外部环境的核心技术,其成功与否直接决定未来智能车辆的生存空间。其中,对于障碍物目标的检测和跟踪又是最为重要和不可或缺的功能,是预防危险和安全行驶的前提条件。目前,高速公路等结构化道路环境中的目标检测技术已基本成熟,但复杂的城市交通环境下的目标检测技术尚不能满足实用化要求,成为智能车辆目标检测领域亟待解决的热点和难点问题。本论文围绕基于视觉导航的智能车辆在城区复杂交通环境中的目标检测、识别和跟踪技术进行了深入研究和探讨。主要研究内容概括如下:(1)论文回顾了智能车辆的研究进展,介绍了当前国内外典型智能车辆系统概况,对智能车辆领域的主要研究内容和关键技术进行了探讨。在对各种信息感知手段进行综合比较的基础上,介绍了智能车辆目标检测技术的主要流派及其研究进展,并提出了当前目标检测技术的不足和本文研究方向。(2)论文对机器视觉理论做了综述性介绍,分析了三种具有代表性的视觉理论框架模型,建立了一种针对具体应用改进的计算视觉模型,为算法设计提供指导。从候选目标检测、目标确认识别、目标跟踪定位这三个方面,重点分析了基于视觉的智能车辆目标检测技术的研究进展,提出了本文基于单目视觉的目标检测算法模型,不仅分析了单目视觉方法的可行性,而且阐明了本文算法的总体架构。(3)论文详细论述了基于单目视觉的候选目标检测方法。对智能车辆候选目标检测中的常用视觉特征进行了分析,介绍了小波变换模极大值对奇异信号探测的理论原理。在此基础上,提出了一种基于小波模极大值和多特征融合的候选目标检测方法。实验表明,该算法能在无道路约束条件下直接从整个像平面中提取候选障碍物目标,有效克服复杂场景中目标特征易被背景或其它目标“淹没”的难题,满足城区环境下智能车辆目标检测的鲁棒性要求。(4)论文详细论述了基于单目视觉的目标确认识别方法。对机器学习中的统计学习理论和支持向量机原理进行了介绍,对目前支持向量机的多类分类方法进行了归纳和分析,并介绍了集成学习的理论原理。在此基础上,提出了一种基于混合核函数和集成学习改进的二叉树支持向量机多类分类方法。该方法根据目标在场景中的出现概率和类间差异设计二叉树树型结构,采用混合核函数设计分类器分类函数,并采用AdaBoost与SVM相结合的方法对分类器参数进行自适应选取,有效提高了分类器的分类精度和泛化能力。实验表明,该方法可有效解决小样本条件下,城区场景中所需考虑的“车辆-行人-非机动车辆-背景”等多类分类问题。(5)论文详细论述了基于视觉的目标跟踪定位方法。对Mean Shift算法的理论基础——基于核函数的无参数密度估计及其一般过程进行了介绍,并对算法收敛性条件做了简要分析。在此基础上,提出了一种基于Mean Shift的智能车辆目标跟踪算法。该方法以目标颜色空间为特征空间,采用颜色特征的统计直方图对目标模型进行描述,利用Bhattacharyya系数定义目标模型在相邻帧间的相似性度量,并基于Mean Shift迭代实现被跟踪目标的准确定位。该方法着力改进并解决了经典Mean Shift算法中的以下关键问题:基于Kalman滤波器改进机动目标跟踪初始点选取问题;提出一种分块匹配策略改进目标遮挡情况下的处理问题;提出一种基于匹配贡献度的选择性子模型更新机制,克服了实际跟踪过程中的模型偏移问题,保证算法在实际应用中对目标的长时间稳定跟踪。最后,对全文的研究工作进行总结,并指出今后工作中进一步研究的方向。更多还原

【Abstract】 Intelligent vehicle as the key technology and important part of Intelligent Traffic System (ITS) has been regarded as the effective way to resolve these traffic problems. The research of intelligent vehicle involve in image processing, pattern recognition, artificial intelligence, automatic control, sensing technology and so on. It has integrated the latest achievements of the information science and artificial intelligence. Thus, the research of intelligent vehicle has important values of practical and academic theory.Intelligent vehicle is an intelligent agent which integrates the functions of environmental perception, planning decision and operation control. In the lots of research tasks of intelligent vehicle, the information perception of driving environment is the foundation and key of all works. Machine vision is the key technology of information perception, and its success or failure will directly decide the survive space of intelligent vehicle in the future. Furthermore, the object detection and tracking is the most important and indispensable function in information perception, and is also the precondition of preventing risk and safe driving for intelligent vehicle. At present, the object detection technical in highway traffic scenes has been maturely almost, but it can not meet the practical requirement in complex urban traffic scenes. Thus, it is urgent and hot problem to develope the research of object detection for intelligent vehicle in complex urban traffic scenes. This paper further discusses the vision-based object detection, recognition and tracking in complex urban traffic scenes for intelligent vehicle. Main study content has been summarized as follows:Firstly, the research progress and typical systems of intelligent vehicle in domestic and international are reviewed, the main research contents and the key techniques of intelligent vehicle are also discussed. Based on the comparison of common information perception methods, the main technical schools of object detection in intelligent vehicle are discussed, and these research progresses are reviewed. Furthermore, the shortages of current research and the direction of future research are proposed in the paper.Secondly, the machine vision theory and three representative vision theory models are discussed in the paper. A new application-based improved computing vision theory model is established and provides the guiding for the practical algorithm design. From the three views of candidate object detection, object recognition and object tracking, the research progress of vision-based object detection in intelligent vehicle is reviewed. A practical monocular vision-based object detection algorithm model is presented, in which the feasibility of monocular vision and the general algorithm architecture of the paper are discussed.Thirdly, the monocular vision-based candidate object detection method is discussed detailedly in the paper. The common visual characters for candidate object detection in intelligent vehicle are discussed, and the theory principle of the singular signal detection based on Wavelet Transform Modulus Maxima (WTMM) is introduced. Furthermore, a novel WTMM-based candidate object detection method is presented in the paper.Experiment shows that the method can directly detect the candidate objects from the whole image plane without road constraint, overcome the problem of the object character submerged by the background and other objects, and meet the robustness acquirements of intelligent vehicle in complex urban traffic scenes. Fourthly, the vision-based object recognition method is discussed detailedly in the paper. The statistical learning theory and the principle of Support Vector Machine (SVM) are introduced, the main SVM-based multi-class classification methods are analysed, and the ensemble learning theory is discussed. Furthermore, a multi-class classification method based on binary tree SVM (BT-SVM) improved by mixtures of kernels and ensemble learning is presented in the paper. In the proposed method, a decision-tree structure is designed based on the distributing probability and pattern diversity of common objects in urban traffic scenes. The classifier function is designed based on the mixtures of kernels. Moreover, the parameters of classifier function are adaptively selected based on AdaBoost. The proposed method effectively improves classification accuracy and generalization ability of the classifer. Experiment shows that the method can effectively resolve the multi-class classification problem including vehicle, pedestrian, non-motor vehicle, background, etc in urban traffic scenes, especially the small samples condition.Fifthly, the vision-based object tracking method is discussed detailedly in the paper. The theory of no-parametric kernel density estimation based on Mean Shift algorithm is introduced, and the algorithm convergence condition is discussed briefly. Furthermore, a Mean Shift-based intelligent vehicle object tracking method is presented in the paper. In the proposed method, the object color space is used as feature space, the color feature statistical histogram is used to describe the target model, the Bhattacharyya coefficient between target model and target candidate model of adjacent frames is defined as the similarity metric, and the Mean Shift iterative algorithm is applied to locate the tracked object. The proposed method emphasizes to improve the following key problems of classical Mean Shift algorithm: 1) the Kalman-based motion prediction is introdced to improve the tracking initial point selection of maneuvering object; 2) a block matching strategy is presented to improve the processing for object occlusion in complex traffic scenes; 3) a selective submodel updating mechanism based on matching contribution degree is presented to overcome the model migration problem during tracking and guarantee the stability of long-time tracking in the practical application.Finally, the conclusion of whole research work in the paper is given. Furthermore, the further work and research prospects are introduced.更多还原