【作者】 宋佳声；

【导师】 胡国清；

【作者基本信息】 华南理工大学 ， 机械设计及理论， 2014， 博士

【摘要】 随着计算机技术的不断进步以及计算机运算能力的持续提升，计算机视觉作为模拟人类视觉功能的复杂课题受到了越来越多的关注。视频序列图像中的运动目标检测与跟踪技术是计算机视觉的一个十分重要的研究方向，其目的是在序列图像中根据信息在时间和空间上的相关性确定目标在每一帧的位置、形态和速度等属性，它对后续基于视觉的应用任务具有重要的意义。由于受到了目标自身变化、环境光影改变、成像设备噪声等多方面干扰的影响，使得设计一个兼具鲁棒性、准确性和快速性的检测与跟踪算法依然是一项极具挑战的开放性课题。本文在调研已有研究成果的基础上，探索了背景建模、聚类定位、多特征融合、轮廓演化模型以及滤波估计框架等相关技术领域，创新地提出了针对一些具体问题的解决方法。主要研究工作和结论如下：首先，为提高目标检测系统中高斯背景模型的更新速度，提出了场景运动复杂度的概念和计算方法，并在此基础上提出了一种组合高斯模型的背景建模方法。根据像素的时空采样模型分析场景运动的复杂性并计算出场景的熵值图，按照最大熵阈值将熵值图分割为稳定区域和动态区域，然后在不同的区域采用不同的高斯模型和更新算法。与固定模态个数的高斯模型相比，由于它是建立在对背景区域做出合理分析和解释之后一种组合建模方法，能够避免高斯模态的浪费，又提高了背景参数的更新速度和前景目标的检出速度。其次，提出了一种基于聚类分析的目标定位算法。由于检出的前景常具有野点数据多、区域连通性差等问题，在利用区域生长法检出前景目标属性（大小与位置）时，常会出现错误的检出结果。为此提出了基于近邻分析的前景目标定位算法：首先，经过对背景减法所得到的前景进行适当的下采样和滤波操作后，将目标定位问题转化为像素聚类问题；然后，基于对这些像素间的距离和上下文关系的分析提出了互邻特征矩阵，定义了聚类准则函数；最后，在准则函数最小化标准下设计了完整的聚类算法。通过上述近邻分析过程将前景像素归为几个特定的聚类，从而实现对前景目标的准确定位。然后，提出两种基于区域统计特征的跟踪算法。（1）在无迹卡尔曼滤波框架下，提出了一种基于颜色和边缘统计特征融合的目标跟踪算法。结合了区域的颜色和边缘特征的表观模型比单一区域特征的表观模型更加全面准确地表征了目标，提高了跟踪算法的抗干扰能力和准确性。与此同时，UKF高效的预测更新机制减少了均值漂移算法迭代的次数，提高了目标的搜索效率。（2）在粒子滤波框架下，提出了一种基于颜色与光流特征的目标跟踪算法。为了能够确定目标坐标、速度、大小以及旋转等多方面的信息，并且为了克服单一颜色特征表征造成的可分性差的问题，首次在层次粒子框架下提出了的基于颜色和运动特征的跟踪算法。实验表明，由于采用了层次结构设计和颜色与光流两种区域特征，算法能够自适应地调整跟踪窗口的位置、大小和方向，正确地估计目标状态，得到了比较准确的跟踪结果。在有遮挡时，算法能正确地预测目标位置并在目标重新出现后能够及时捕捉目标继续跟踪，表现了较好的鲁棒性。最后，在UKF框架下，提出了一种基于边缘特征的目标跟踪算法。在经典的几何式主动轮廓分割算法中，为了得到准确的分割结果需要冗长的迭代过程，而且有时并不能如愿。为了提高分割效率和准确性，在UKF框架下设计了全新边缘检测与跟踪算法：首先，采用矢量图像计算图像的梯度值，并设计了能够自适应调整阈值的边缘指示函数；然后，提出了改进后的变分水平集演化模型；最后，在UKF框架下设计了针对运动目标的边缘检测与跟踪算法。实验证明，算法不但显著地提高了轮廓演化模型的灵活性和收敛速度，而且对于阴影、遮挡、目标形变和背景干扰等表现出了较好的鲁棒性。总之，全文针对本课题的有关问题进行了比较系统和深入的研究，相关实验表明，所提出算法在鲁棒性和准确性方面都有较大提高。更多还原

【Abstract】 With the development of computer technology and improvement of computing power, acomplicated research topic, computer vision, attracts more and more attention, which aims tosimulate human visual capabilities. As a very important research interest of computer vision,the detection and tracking of moving objects in sequence images is to recognize the objects’information such as position, shape, velocity, etc. based on their temporal and spatialcorrelation, which has great significance to further vision-based applications. The detectionand tracking technology is facing many challenges such as object defomation, changefulcircumstance, unstable imaging devices, etc., which makes it still a challenging open researchissue to devise a robust, accurate and rapid detection and tracking algorithm. The dissertationexplored the concerned technologies by investigating the existing work such as backgroundmodeling, locating based on clustering, multi-feature fusion, contour evolution, filteringestimation framework, etc. and proposed innovative solutions to some related problems. Itsmain work and conclutions are as follows:Firstly, in order to improve the updating speed of Gaussian models for background, theconcept and computation method of the scene moving complexity were devised, according towhich a combinational Gaussian model for background modeling was proposed. In thismethod, according to the spatio-temporal sampling model of pixels, the scene movingcomplexity was analyzed and the entropy image of the scene was calculated. And this imagewas segmented into the stable region and the dynamic region by means of the maximumentropy threshold. In the two different regions, two different Gaussian models andcorresponding updating algorithms were respectively adopted. Since the modeling method isbased on reasonable analysis and classification for the surveillance scene, the proposed modelcan avoid the waste of Gaussians and be provided with higher updating and detecting speedcomparing to fixed number of Gaussians.Secondly, a object locating algorithm was proposed based on clustering analysis.Because the foreground resulted from background subtract often has many undesirable features such as abundance of outliers, spoiled connectivity, etc., the furter detection byreagion growing almost can’t locate the outline of the object correctlly. In order to solve theseproblems, a locating algorithm of moving objects was proposed based on neighboringanalysis. First, the foreground resulted from background subtract was downsampled andfiltered. And the locating of moving objects was converted to pixels clustering. Second, theneighboring feature matrix and the criterion function were proposed based on analysis of thedistance and context of the pixels. Finally, according to the minimization of the criterionfunction the clustering algorithm was devised. And the pixels were clustered into a certainnumber of clusters corresponding to objects. Thus, the foreground objects were locatedcorrectly.Then, the two tracking algorithms were presented based on in the framework of UKF aobject tracking algorithm was presented based on the region statistical characteristics.(1) Inthe framework of UKF a improved tracking algorithm based on the fusion of color and edgefeatures was proposed. The multi-feature fusion appearance model describes the object morecomprehensively than the single feature one, which has enhanced the accuracy and adaptivityof the tracking algorithm. At the same time, because of the effective “predict-update”mechanism in UKF, the iterations of mean-shift has greatly decreased and the searchefficiency for the tracked object is improved.(2) In the framework of Particle filtering, aobject tracking algorithm were presented based on region statistical characteristics of colorand optical-flow. In order to recognize various information of the object such as coordinates,speed, size and rotation, and to overcome the poor discrimination of color features, a trackingalgorithm in a hierarchical Particle filtering framework was proposed for the first time. Itadopted color and optical-flow features. The tests indicated it could adaptively adjust thetracking window’s position, size and rotation, estimat the target’s state correctly and track theobject accurately because of adopting hierarchical filtering stracture and two-feature fusionalgorithm. And it predicted the object’s position correctly under occlusion and catched itshortly when the occlusion disappeared, which showed it had a robust performance.Finally, in the UKF framework, a object tracking algorithm was proposed based on edge feature. During executing the algorithm of classical geometric active contour segmentation, toobtain accurate segmentation results always involves lengthy iterative process, which doesn’teven work. To improve the segmentation efficiency and accuracy, a novel detection andtracking algorithm was presented. First, the gradient image was calculated based on the vectorimage and an adaptive edge indicator was proposed. Second, the revised evolution modelusing variational level set method was put forward. And then the detection and tracking of theobject’s edge is presented in the framework of UKF. The experiments demonstrate not onlythat it has significantly increased the convergence rate and flexibility of the active contourevolution but also that it is robust to some interference such as shadow, occlusion,deformation of object and background interference.In summary, the dissertation has been presented in a comprehensive way to discuss theserelevant issues. And extensive experiments shows the proposed algorithms have been greatlyimproved in terms of roubustness and accuracy.更多还原