本文为日本大学（作者：XiaofengLU）的博士论文，共143页。
视频监控已成为近年来图像处理和计算机视觉技术的一个重要研究领域，它尝试从图像序列中检测、识别、跟踪某些物体，并了解、描述目标的行为。视频交通监控系统为智能交通系统（ITS）的交通控制和管理提供最有效的交通信息，为安全驾驶提供帮助。运动目标检测与跟踪方法是智能视频监控领域中最基础、最重要的技术，是实现实时智能视频监控的关键。然而，由于视频监控的发展历史较短，一些重要问题仍未解决，亟需寻找新的技术方法。因此，对运动目标检测与跟踪的研究具有重要理论意义和实用价值。本文主要针对智能交通监控中的关键技术问题进行了研究，主要工作包括：
提出了一种检测运动目标的背景剔除方法。考虑动态条件和干扰因素，提出了基于YCbCr色彩空间的彩色背景统计模型。提出了一种利用Choquet积分对前景和背景进行分类的多特征相似性融合方法，为高精度目标检测提供了一种新的思路。通过分析传统的盲目和选择性背景维护过程，提出了一种自适应的背景维护方法，以适应复杂环境的需要。
针对单目标跟踪，提出了一种基于粒子滤波（PF）的多特征融合算法。本文提出的算法机制不仅融合多个特征来表示跟踪目标，而且能够通过获得自适应特征权重动态平衡目标模型、候选背景和邻近背景之间特征相似性及可区分性的影响。采用Bhattacharyya系数来表示特征相似度，利用对数似然比的方差描述目标模型与邻近背景的区分度。
针对复杂条件下的跟踪问题，如大尺度变化、旋转和目标交互等情况，提出了一种基于粒子滤波框架的加速鲁棒特征（SURF）的鲁棒车辆跟踪方法。此外，我们还提出了目标模板的动态更新机制来捕捉场景变化。具体而言，就是采用新的特征点，丢弃不良的特征点。跟踪窗口的大小也通过平衡三个特征分布的权重来动态修改。此外，利用改进的距离核函数法对每个粒子的权重进行分配。
为了鲁棒地跟踪多个目标，首先分析了传统的数据关联方法，然后提出了一种基于特征测量概率假设密度（PHD）滤波的多目标跟踪方法。在该方法中，使用特征测量来近似后验密度。采用自适应权值融合蒙特卡罗技术提取的颜色特征和LBP特征，实现了基于高斯混合的跟踪方法。
本文给出了每种方法对应的实验结果，并在精度准则下评价、讨论了该方法的有效性。本文的研究将对智能视频监控中的运动目标检测与跟踪技术做出贡献。
Visual surveillance in dynamic scenes has become a very importantresearch area of image processing and computer vision techniques in recentyears, which attempts to detect, recognize, and track certain objects fromimage sequences, and more generally to understand and describe targetbehaviors. Visual traffic surveillance system provides the most efficienttraffic information for traffic control and management, and assistance for safedriving in Intelligent Transport Systems (ITS). Moving target detection and trackingmethods are the most basic and important technologies in the area ofintelligent visual surveillance, and are the key to realizing real timeintelligent visual surveillance. However, due to the short history ofdevelopment, some important problems are still unresolved, and new methods oftechniques are needed. Thus, the research of moving target detection andtracking has great theoretical significance and practical value. This dissertationdoes the researches focused on the key technical problems about intelligent trafficsurveillance, and the major works include as follows: A novel backgroundsubtraction method is proposed to detect the moving target. Due to the dynamicconditions and interference factors, the color statistical background model basedon YCbCr color space is presented. We propose a multiple feature similarityfusion using Choquet integral to class the foreground and background, and providea new idea for high precision target detection. By analyzing of the traditionalblind and selective background maintenance process, an adaptive background maintenancemethod is proposed to adapt the complex condition. For single target tracking,we present a multiple feature fusion algorithm based on the Particlefilter(PF). The proposed mechanism not only fuses multiple features to representthe tracking target, but dynamically balances the effect of feature similarity andfeature discriminability among target model, candidate and adjacent backgroundto obtain the adaptive feature weight. Bhattacharyya coefficient is adopted torepresent the similarity, and the variance of the log-likelihood ratio is usedto describe the discriminability between the target model and the adjacentbackground. For the tracking problems of complex conditions, such as largescale change, rotation and mutual osculation, etc. A robust vehicle trackingbased on the Speeded-Up RobustFeatures(SURF) in aparticle filter framework is presented. What’s more, we propose the dynamicupdate mechanisms of target template to capture the appearance change. Specifically,adopting new feature points and discarding bad feature points. The size of trackingwindow is also modified dynamically by balancing the weights of three feature distributions.Furthermore, the weights of each particle are allocated by an improved distancekernel function method. In order to robustly track multiple target, we firstlyanalyze the traditional data association methods, and then propose a multipletarget tracking method based on feature measurement Probability HypothesisDensity (PHD) filter. In this method, the feature measurement is used toapproximate the posterior density. And, we adopt an adaptive weight to fuse thecolor and LBP features which are extracted by Monte Carlotechnology, and implement the tracking method using Gaussian mixture. Theexperimental results corresponding to each method are presented and the effectivenessof the methods are evaluated and discussed under the criterion of accuracy. Theresearches of this thesis will make a contribution to the technology of movingtarget detection and tracking in intelligence visual surveillance.
1 引言
2 采用背景剔除的运动目标检测
3 采用粒子滤波和多特征融合的单目标跟踪技术
4 采用SURF和PF的单目标跟踪技术
5 采用FM-PHD滤波器的多目标跟踪技术
6 结论与未来研究工作展望
下载英文原文地址：
http://page5.dfpan.com/fs/9lcej222172901665f4/
