【作者】 钱翔；

【导师】 李新华；

【作者基本信息】 安徽大学 ， 信号与信息处理， 2010， 硕士

【摘要】 目标跟踪被广泛应用于视频监控、安防系统、智能交通系统及机器人技术等领域,是一些需要确定目标位置、运动和身份等智能系统的核心组成部分,可以说是一个非常棘手而富有挑战性的课题。由于实际应用中存在诸如摄像头运动、目标不稳定、背景复杂以及其他相似移动物体等的困难的情况,人们很难找到一个广泛适用的鲁棒性高的跟踪算法。四十多年前,卡尔曼先生提出了卡尔曼滤波算法,它简单而便于实现,是解决线性高斯环境下的问题的最佳方法。近年来由于技术的发展和应用的需要,出现了一个研究非线性非高斯环境下滤波算法的高潮。本文首先介绍了近年来常见的一些目标跟踪滤波算法——卡尔曼滤波器(Kalman filter, KF)、扩展卡尔曼滤波器(Extended Kalman filter, EKF),无敏卡尔曼滤波器(Unscented Kalman filter, UKF),粒子滤波器(Particle filter, PF)。KF简单而优雅,是线性高斯环境下的最佳递归贝叶斯滤波器。EKF利用泰勒级数方法,将非线性问题转化到线性空间,再利用卡尔曼滤波器进行估计滤波,并达到一阶估计精度。UKF通过固定样本集达到对状态概率分布的近似,在精度和计算量上较之EKF优秀,但它是利用高斯分布来逼近系统状态的后验概率密度,在复杂的环境中表现差。PF是一种采用蒙特卡罗采样的贝叶斯滤波方法,它将复杂的目标状态分布表示为一组加权值(称为粒子),通过寻找在粒子滤波分布中最大权重的粒子来确定目标最可能所处的状态分布,已成为复杂环境下进行目标跟踪的最好的方法。本论文通过量测非线性模型(正切)的对比实验,证明了PF在非线性环境下有着最优异表现,UKF表现较之EKF优异,而EKF优于KF,与理论分析的结果一致。在目标跟踪系统中选取描述目标的特征是一个棘手的问题,使用更多的描述目标的特征可以有效提高跟踪的准确性,但会增加计算机的计算量和计算时间,只能取实时性和准确性的折衷。目标颜色直方图特征具有稳定性高、计算量较小的特点,已成为主流的描述目标的特征。描述目标颜色特征的颜色空间有很多种,本文介绍了常见的RGB空间、CMYK空间、HSV空间。其中HSV空间更符合人眼感知色彩的方式,此空间模型具有线性伸缩性良好,色差与颜色分量在相应值上的欧几里德距离成比例等优点。但单一的颜色直方图特征对背景光照变化敏感,而且当有相似颜色干扰信息时,跟踪的准确性大大降低。而目标的结构性特征主要有矩特征,矩特征具有平移、旋转、尺度等不变特性,被广泛应用在图像匹配、姿态识别等领域。本文结合粒子滤波算法提出一种基于融合目标不变矩特征和颜色直方图特征的目标跟踪方法,其中目标颜色直方图特征的计算是在HSV空间进行的。该方法对目标的颜色信息和结构信息进行融合并建立目标模型,粒子权值的融合比例由应用环境决定,在系统更新过程中比较目标与粒子间的欧几里德距离,淘汰劣质的粒子,增加了粒子的可靠度,减少了噪声的影响。实验数据表明此改进方法有效克服了单一颜色特征模型在应用中的不足,并能够在不影响实时性的基础上提高跟踪的有效性。更多还原

【Abstract】 Target tracking is the core components of intelligent systems to determine location, movement and identity targets, which is widely used in the field of video surveillance, security systems and intelligent transportation system. It is really a tough job to realize this system and find a widely used and high robust tracking algorithm due to camera movement, target instability, complexity of background and moving similarity. It is hard to find. Kalman filter algorithm, proposed by Mr. Kalman more than 40 years ago, is the best way to solve the problem in the linear Gaussian environment. However, in order to meet technology and application needs, there has been an emergence of studying the nonlinear non-Gaussian filtering algorithm recent years.First, this thesis introduced the common target tracking filter algorithm proposed recently, such as Kalman Filter (KF), extended Kalman filter (EKF), Unscented Filter (UKF) and particle filter (PF) algorithm. Simple and elegant, KF is the best recursive Bayesian estimator in linear Gaussian environment. By using Taylor series, EKF transforms nonlinear problem into linear space, then using Kalman filter to estimate the results to achieve the first order accuracy. Through the fixed sample set to approximate the probability distribution of the state, UKF is better than the EKF on precision and quantity. Nonetheless, as using Gaussian posterior probability density to approximate the system state, it is poor to perform in complex environment. Particle filter (PF) is a Bayesian filtering adopted by Monte Carlo sampling method. The complex target state distribution is expressed as a set of weights (called particle) in this filter. By finding the largest weight particles in the particle filter to determine the most likely target has been proved as the best way to track target in a complex environment. By measuring the nonlinear model (tangent), this thesis demonstrated that the particle filter has the most outstanding performance dealing with the nonlinear situations, and UKF has superior performance than EKF, EKF is better than KF, which is identical to the theoretical analysis.Second, it is a tough job to select the characteristics of targets in target tracking system. If targets have more features, tracking accuracy could be effectively improved, however, computing quantity and calculation time would also increasing. It is imperative for us to take compromise of real-time and accuracy. As high stability and low computational characteristics, color histogram feature are becoming a main feature to describe targets. This thesis introduced RGB space, CMYK space, HSV space. All these spaces, HSV space is more suitable for human to perceive color. Also, this model has good linear scalability advantages. However, single color histogram is sensitive to the background illumination and what’s more, tracking accuracy could reduced significantly when this system is interfered with similar color objects. By contract, as objects moments feature has structural characteristics, which have the properties of translation, rotation and scale invariant, etc. It is widely used in image matching and gesture recognition.Finally, by combining both properties of target color and moment invariant this thesis presented an improved particle filter based on method, the characteristic of the color histogram is carried out in HSV space. The weights of the particle are determined by application environment. Further, by determined Euclidean distance properties in the process of replacement, poor quality particles were washed out, reliability of particles were increased and the impact of noises were reduced. The experimental results have been demonstrated that this method ameliorated the interference immunity of the single color property for tracking target. In addition, this method also improved the tracking accuracy and robustness while not affecting the real-time characteristics.更多还原