【作者】 宋策；

【导师】 张葆；

【作者基本信息】 中国科学院研究生院(长春光学精密机械与物理研究所) ， 光学工程， 2014， 博士

【摘要】 目标跟踪技术一直以来都是计算机视觉、图像处理领域的研究热点，其在智能监控、视觉导航、智能交通、人机交互、国防侦察等领域具有重要应用价值，是武器系统的核心技术之一。虽然近二十年来众多学者对目标跟踪技术进行深入研究，但由于跟踪初始阶段目标模板获取不准确、目标在像面内运动规律的复杂性、目标观测特征的实时变化、目标所处背景的复杂干扰、遮挡等因素，导致当前的目标跟踪技术仍不能满足军、民领域的需求，因此仍需对其进行深入研究。目标跟踪问题可以定义为已知目标先验信息，在获取目标新的观测信息后，迭代求取目标状态矢量后验概率密度分布的过程，因此可将目标跟踪过程建模为贝叶斯估计。本论文主要以粒子滤波为跟踪框架，重点对其动态模型及观测模型进行研究；同时针对目标检测、目标分割算法进行研究，试图将目标检测与分割算法与基于粒子滤波的跟踪算法相融合，进而达到减小跟踪误差、提高跟踪精度的目的。本论文主要创新工作及研究成果如下：1.粒子滤波的动态模型是对目标运动方式的描述，若模型描述与目标实际运动方式差异较大，必然导致预测过程后粒子无法准确覆盖目标真实位置，跟踪误差逐渐累积甚至跟踪失败。本文针对机载环境对地面目标跟踪的特点，提出加速度双步动态模型（TSA），其中包含自由模型与保守模型两部分，自由模型将目标速度建模为非零均值的高斯马尔科夫过程，该模型通过参数调整可以较好描述RW模型与NCV模型之间的运动形式；保守模型对目标当前时刻速度进行估计，替代自由模型中高斯马尔科夫过程的目标速度平均值。实验结果表明，该模型对目标在像面内大幅度变速运动有较好的预测能力。2.粒子滤波的观测模型决定粒子权重，直接影响跟踪精度，较为经典的观测特征为目标颜色、轮廓等特征。核密度估计直方图是一种经常被采用的特征，其关键技术为核函数的选取。本文基于Snake轮廓提取算法，构建非对称核函数，融合目标的颜色信息，进而得到目标的核密度估计直方图作为粒子滤波的观测模型。该模型不但可以较好描述目标的观测信息，在目标观测特征变化时可实时更新目标模板。该模型的构建算法可以达到实时，具有实际工程应用价值。3.对于背景复杂的目标跟踪问题，上述非对称核函数的构建算法会出现较大误差。为此，本论文将图像分割算法融入到粒子滤波跟踪算法之中，首先提出多方向GrabCut算法，该算法相比传统分割算法具有更强的分割鲁棒性；进而提出基于该分割算法的MGC-PF粒子滤波算法。实验结果表明，该算法可以较好解决因目标处于复杂背景而引起跟踪精度较差的问题。4.为设计能够更好与跟踪算法相融合的GrabCut分割算法，充分利用跟踪过程目标的时间相关性及空间相关性，本文将随机森林分类器融合到GrabCut分割算法之中，提出适于融入目标跟踪算法中的RF-GC分割算法，最终提出RF-GC-PF粒子滤波跟踪算法。实验结果表明，该算法可以较好解决跟踪过程中由于目标运动规律复杂、目标观测特征实时变化、背景复杂等综合因素引起跟踪精度较差的问题。本论文试图解决目标跟踪过程中由于目标运动规律复杂、目标观测特征实时变化、复杂背景等因素导致的跟踪误差较大、精度较差的问题，对在传统跟踪算法中融入图像分割、目标检测算法的这一发展趋势进行了研究、分析与展望。更多还原

【Abstract】 Target tracking technology has being a research focus in computer vision andimage processing, which has important application value in intelligence monitoring,vision navigation, intelligent transportation, human-computer interaction, defensereconnaissance, and is one of the key technology of weapon systems. Although manyscholars of the past two decades are in depth study of target tracking technology, butbecause of obtaining the inaccurate target template in the initial stage of tracking,complexity movement of the target in the image plane, changing of the targetobservational characteristics, complex background interference, occlusion and otherfactors, the current target tracking technology still can not meet the need of militaryand civilian areas, and therefore target tracking technology still need to be studied indepth.Target tracking problem can be defined as when having the priori information oftarget, after obtaining visual information on new observation, the iterative process ofobtaining the posterior probability of the target state vector, so the target trackingproblem can be modeled as Bayesian estimation. This thesis is mainly based onparticle filter tracking framework, focusing on the dynamic model and observationmodel; simultaneously researching on target detection and target segmentationalgorithm, trying to integrating target detection and segmentation algorithm intracking algorithm based on particle filter to reduce tracking error and improve tracking precision. The main innovation and research results are as follows:1. Dynamic model of particle filter is to describe how the target moving, and thedynamic model describes quite different from the pattern of target actual movement,will inevitably leading the particles cannot accurately coverage of the target trueposition in predicting stage, and leading tracking error gradually accumulate, evenleading tracking failure. For tracking ground target with cameras in airborneenvironment, this thesis proposes a Two-Stage Acceleration dynamic model(TSA),which includes liberal model and conservative model. In the liberal model, targetspeed will be modeled as a zero mean Gaussian Markov process, and can describe themotion pattern between RW model and NCV model through parameter adjustment.Conservative model estimate the current velocity of the target, and alternate theaverage target speed of Gauss-Markov process in liberal model. Experimental resultsshow that the TSA model can predict the target motion accurately when the targetmoving in large scale in the image plane.2. Observation model of particle filter determines the weight of the particles andaffect the tracking precision directly. Classic observational features are color, contourand so on. Kernel density estimation histogram feature is often used and the keytechnology is the selecting of kernel function. In this thesis, asymmetric kernelfunction is constructed based on the Snake contour extraction algorithm, whichintegrated the color information of the target, and then the kernel density estimationhistogram of target is constructed as observation model of particle filter. The modelnot only can describe the target appearance information better, but also can be updatedin real-time when observation features of target changes. Addition, this model isconstructed in real-time algorithm and has practical engineering value.3. For complex background around the target when tracking, the asymmetricalgorithm kernel is constructed with larger error. To deal with such problem, thisthesis will integrate particle filter tracking algorithm with GrabCut imagesegmentation algorithm. First more robust multi-directional GrabCut algorithm isproposed, and then MGC-PF particle filter algorithm is proposed. Experimental results show that this algorithm have better tracking precision when tracking target incomplex background.4. To design a segmentation algorithm based on GrabCut which can integratetracking algorithm better, make full use of the time correlation and spatial correlationin target tracking process, this thesis will integrate random forest classifier to GrabCutsegmentation algorithm, RF-GC segmentation algorithm is proposed, and finallyRF-GC-PF particle filter tracking algorithm is proposed. Experimental results showthat the algorithm can solve the problem of poor tracking precision due to thecombination factors of complex movement of target, observation features changingand the complex background.This thesis attempts to solve the problem of poor tracking precision due tocomplex movement of target, target observation features changing in real-time, thecomplex background and other factors, and the development trend of integratingimage segmentation, target detection algorithm to traditional tracking algorithm isresearched, analysis and outlook.更多还原