【作者】 李鹏；

【导师】 王博亮；

【作者基本信息】 国防科学技术大学 ， 电子科学与技术， 2012， 博士

【摘要】 随着科学技术的进步，数字图像序列已成为一种重要的信息载体并广泛应用于国防、工业生产、文化传媒、以及医疗诊断等领域。目标提取发现和提取目标在图像序列中的时间和空间分布，是图像序列分析中的核心科学问题之一，在科学研究和工程应用上都有着十分重要的意义。近年来，低帧频（low-frame-rate）图像序列（帧频≤5帧/秒）在移动成像、无线数据传输及存储容量受限等复杂场景中得到了广泛应用。在低帧频图像序列中，相邻帧时间间隔长、目标时空连贯性差、目标外观与尺度变化剧烈，给目标提取带来了新的挑战。本文围绕低帧频图像序列，研究了目标提取的三个关键技术：目标检测、目标跟踪以及图像序列目标分割。目标检测是目标提取的起点和支撑；目标跟踪获得目标的时空连贯性；图像序列目标分割得到目标在序列中的时空区域。本文的主要研究成果如下：一、提出了一种预筛选Hough森林目标检测算法。针对现有的Hough森林算法存在的随机抽取图像块样本利用效率低、无效样本干扰大的问题，提出了基于图像块表述质量的随机样本预筛选思路；利用量化灰度级空间相关直方图描述图像块样本的灰度与局部空间结构的统计特性，并结合二维熵建立图像块表述质量的度量，进而建立了预筛选Hough森林目标检测算法。多组数据集上的实验结果证明了所提出的预筛选机制能够降低随机森林的不确定性并提高Hough森林的目标检测性能。二、在低帧频图像序列目标跟踪方面，针对目标大尺度变化的难题，提出了一种基于核的变尺度目标跟踪算法（SIKBOT）。首先提出了一种基于集合分析的目标相似性度量，进而结合均值漂移过程求解尺度维加权核密度函数的模值搜索问题。在目标跟踪迭代中，并行地使用尺度维与空间维上的两个均值漂移过程估计目标的尺度与位置，实现变尺度目标跟踪。与现有主流方法相比，该方法提高了对变尺度目标跟踪的能力。三、在低帧频图像序列目标跟踪方面，针对目标形状不规则的难题，提出了一种基于目标形状的Epanechnikov核函数（形状核，shaped kernel）并用于本文提出的SIKBOT方法中，形成基于形状核的变尺度目标跟踪方法（SK+SIKBOT）。所提出的形状核可以避免背景噪声对目标建模的影响，而且严格满足均值漂移算法收敛的充分条件。多组目标跟踪实验证明SK+SIKBOT方法不但提高了目标跟踪的精度与鲁棒性，也提高了目标跟踪的效率。四、针对低帧频图像序列目标分割中分割误差累积的问题，提出了一种时空Grab Cut算法。将Grab Cut图像分割算法的不完整标记与迭代式估计的思想推广到图像序列目标分割中，建立了目标/背景统计分布的帧间传递机制，有效克服了图像序列目标分割错误的累积问题。实验结果证明了所提出的时空Grab Cut算法的性能超过了目前主流的基于图割的图像序列目标分割算法。更多还原

【Abstract】 With the development of scientific technology, image sequences become anextremely important data source and have been extensively applied in many fields, e.g.national defense, industry production, entertainment, media broadcast, and medicaldiagnosis, etc. Object extraction is to find and extract the space-time distribution objectsin image sequences. As a fundamental problem in image sequence analysis, objectextraction receives attentions from both scientists and engineers. Low-frame-rate imagesequences (frame rate≤5fps) have, for the past few years, been put into use in manycomplicated scenes, such as moving imaging platforms, wireless surveillance or limitedstorages. Low-frame-rate image sequences bring object extraction new challenges, e.g.,the long temporal interval between adjacent frames, weak spatiotemporal coherence,and intensive change in object appearance and scale.In this thesis, we focus on three key techniques for object extraction inlow-frame-rate image sequence object detection, object tracking and image sequenceobject segmentation. Object detection is the start point for object extraction; objecttracking explores the spatiotemporal coherence of objects; object segmentation obtainsspatiotemporal regions occupied by the objects. The main contributions of thedissertation are summarized as follows:1. To solve the problem of underutilization of the randomly extracted patches,we proposed a prescreening mechanism for the Hough forest based on therepresentation quality. The gray level spatial correlation histogram (GLSCH) wasintroduced and improved to characterize the randomly extracted patches. Then weemployed2D image entropy to measure the representation quality of the patches andconstructed the prescreening-based Hough forest. Extensive experiments on standarddatabase demonstrated the proposed pre-screening mechanism decreased the uncertaintyHough forest and improved the detection performance.2. We developed a novel scale invariant kernel-based object trackingalgorithm (SIKBOT) for tracking fast scaling objects in low-frame-rate imagesequences. We first proposed a novel set analysis based object similarity measure andthen employed the mean shift procedure to estimate the object scale. During eachiteration in tracking, object scale and object position were simultaneously estimated bytwo mean shift procedures in parallel. Compared with state-of-the-art methods, theproposed SIKBOT method improved the performance for tracking fast scaling objects.3. To accurately describe irregular-shaped objects during tracking, weproposed a new object-shape-based Epanechnikov kernel (shaped kernel, SK), whichwas then combined with the proposed SIKBOT algorithm to construct the shaped-kernelSIKBOT algorithm (SK+SIKBOT). The proposed shaped kernel can alleviate the influence of the background noise during object modeling. Moreover, the Epanechnikovprofile guarantees the strict convergence of the mean shift procedures. Extensiveexperiements demonstrated the proposed shaped kernel achieved improvements in bothaccuracy and efficiency.4. To overcome the error accumulation problem in low-frame-rate imagesequence object segmentation, we proposed a novel spatiotemporal Grab Cut algorithm.The object/background distribution propagation mechanism was established by tracking.Then by introducing the concepts of incomplete labeling and iterative estimation of theGrab Cut, we effectively alleviated the problem of error accumulation. Experimentalresults demonstrated the proposed spatiotemporal outperformed the state-of-the-artGraph Cuts-based image sequence object segmentation algorithms.更多还原