【作者】 洪晓鹏；

【导师】 高文；

【作者基本信息】 哈尔滨工业大学 ， 计算机应用技术， 2010， 博士

【摘要】 随着在视频监控、自然人机交互系统和智能交通系统中越来越多地应用,物体检测与跟踪技术已经成为计算机视觉领域的关键技术之一。本文针对物体检测与跟踪任务中普遍存在的类内散度大和类间可区分度低的问题,选择了具有较好判别力和鲁棒性的基于图像二阶统计特性的区域描述子作为技术路线。针对传统二阶统计描述子计算效率较低的缺点,本文提出了一种新型的二阶统计图像区域描述子,定义了该描述子之间的距离度量以及多个描述子的均值;在此基础上,本文探讨了如何将所提区域描述子用于图像整体表达,并结合常用的分类学习算法构造物体检测器;此外,本文研究了如何利用所提区域描述子对跟踪目标进行有效建模;最后,本文探讨了在集成学习框架下如何设计与描述子之间距离定义相对应的高效分类器。具体地,本论文的主要研究内容如下:第一、二阶统计区域描述子包含了给定图像区域的多种特征(信息)的统计信息,具有判别力较强、鲁棒性较好和维数较低的优点。但是传统的二阶统计描述子,即协方差矩阵存在计算效率较低和对应的分类器设计过于复杂的问题。为克服这些问题,本文提出了一种名为Sigma set的新型图像区域描述子。Sigma set是由少量特征向量构成的集合,其和给定图像区域具有相同的二阶统计特性。它可以通过矩阵分解唯一高效地得到。为了使其成为可胜任多种计算机视觉任务的有效描述子,本文进一步为其定义了距离度量和描述子的均值。所提描述子具有和传统协方差矩阵类似的强判别力和鲁棒性。集合形式使得描述子间的距离度量最终转化为向量空间上的集合元素间距离,因此计算效率要远高于传统二阶统计描述子相应的黎曼流形距离的计算。结合近邻分类器在纹理分类任务中验证了所提描述子的有效性。第二、本文致力于将所提区域描述子应用到物体检测任务。为此,本文将Sigma set区域描述子扩展到对包含物体的整个图像进行描述,称为Sigma set图像表示。和协方差矩阵图像表示相似,Sigma set图像表示同样具有对图像区域信息较强的概括和描述能力。Sigma set描述子的特定距离定义使得Sigma set物体表示可以向量表示并直接与常用学习算法相结合,与协方差矩阵物体表示的黎曼流形分类器相比更为高效。物体检测的实验表明,Sigma set物体表示结合常用分类器能够以更高的效率获得与传统协方差矩阵物体表示及相应黎曼流形分类器类似的性能。第三、本文致力于将所提出的Sigma set描述子应用于实时跟踪任务。由于跟踪的目标物体往往要经历表观的变化,如何能有效对跟踪目标进行建模并适应其表观变化一直是跟踪研究的重要课题。为此,在物体分块Sigma set表示的基础上,本文提出了贝叶斯推理框架下的核函数隐含在线学习实时物体跟踪方法。所提方法具有以下优点:首先,分块Sigma set表示可以有效地对跟踪目标进行描述;其次,目标的局部分块由核函数隐含在线学习算法进行建模,可以从历史观测中有效得到目标估计,快速而鲁棒地适应物体的表观变化;再次,多个分块核函数隐含在线模型的输出由仲裁器进行融合,有效地克服部分遮挡所带来的影响;最后,贝叶斯推理框架确保了跟踪的实时性。多个视频序列上的实验结果证明了所提方法可以有效克服由物体表观变化、运动和部分遮挡带来的影响。第四、区域描述子只有结合相应的距离度量才能取得较好的性能。传统的非线性分类器可以使用和描述子相对应的距离度量,但其计算效率较低。为了解决这个问题,本文致力于在集成学习框架下设计一种既能与特定区域描述子的距离定义相适应,又可以高效计算的弱分类器学习算法。为此提出了样本预映射弱分类器学习算法。样本预映射是在弱学习或者弱分类“之前”进行的一个增强步骤,它来源于与特定区域描述子的距离度量相对应的非线性核函数,它是与输入向量等维数的非线性变换。在预映射之后的空间进行线性弱学习可以有效逼近与核函数所衍生的隐空间上的最优线性分类面。人体与车辆检测的实验证明了所提方法可以以线性分类算法的效率达到非线性分类算法的性能。更多还原

【Abstract】 Along with increasing real-world applications of visual surveillance, human machine interaction systems, and intelligent transport systems, object detection and object tracking have been one of the most active research areas in computer vision during the last decades. The main challenges for object detection and tracking lie in the following two aspects: large intra-class divergence and low inter-class separability. In this dissertation, the technical road map to address these two challenges is to develop the second order statistics based region descriptors, which shows high discriminative power and good robustness in many computer vision tasks. However, operations of the traditional second order statistics based region descriptors, i.e., the covariance matrices, on Riemannian manifold are usually computationally demanding. To solve this problem, this dissertation proposes a novel second order statistics based region descriptor, named Sigma set, and defines how to calculate the distance between two Sigma sets and the mean of multiple Sigma sets. In addition, this dissertation extends the proposed region descriptor to represent a whole image and studies how to combine the proposed descriptor with the commonly used machine learning algorithms. Furthermore, a robust modeling method to represent the target in real-time tracking environment is proposed. Finally, this dissertation discusses on how to design the efficient boosted weak classifiers where the distance metric of commonly used region descriptors is embedded.The detailed descriptions for these techniques are introduced in the following manner.Firstly, second order statistics based region descriptors, which captures the correlations among extracted features inside an object region, is of low dimensionality, discriminative and robust against the variations in illumination, view, and pose, etc. Despite such advantages, because the traditional second order statistics based region descriptors, i.e., the covariance matrices (COVs), do not lie in Euclidean space, computationally demanding operations based on Riemannian manifold are required to measure the distance between two COVs and calculate the mean of COVs accurately. To solve this problem, this dissertation proposes a novel efficient region descriptor, named Sigma set, which encodes second order statistics of the given image region in the form of a small set of vectors. Sigma set can be uniquely constructed through Cholesky decomposition of the covariance matrix. As the covariance matrix descriptors, Sigma sets are of low dimension, powerful and robust. Moreover, compared with COVs, Sigma sets are not only more efficient in the calculations of the distance between two Sigma sets and the mean of multiple ones, but also easier to be enriched with first order statistics. Experimental results in texture classification verify the effectiveness and efficiency of the proposed region descriptor.Secondly, this dissertation extends Sigma sets to holistic representations of images. Sigma set image representation is capable of encoding information of the given image and achieves accuracy comparable to COV representation. In addition, compared with the traditional covariance matrix image representation, Sigma set image representation is more efficient, since that it lies in the vector space. Experiments show that Sigma set image representation works well in object detection as well as object classification.Thirdly, in the real-time object tracking task, how to adapt to the appearance changes of moving objects promptly is crucial. Based on the multiple-patch Sigma set representation, this dissertation proposes to utilize the online one-class support vector machine algorithm, named Implicit online Learning with Kernels Model (ILKM) to model each of the patches. ILKM is simple, efficient, and capable of learning a robust online target predictor in the presence of appearance changes. Responses of ILKMs corresponding to multiple target patches are fused by an arbitrator, which makes the decision and triggers the model update. The benefits of the proposed appearance model are fourfold: firstly, patch based representation increases the robustness against partial occlusions; secondly, ILKM is capable of learning a robust predictor based on the observed samples efficiently; thirdly, the arbitrator including an analysis of possible partial occlusions further improves the robustness; finally, Bayesian inference framework ensures the efficiency of the proposed tracking approach. Experimental results demonstrate that the proposed tracking approach is effective and efficient in ever-changing and cluttered scenes.Fourthly, there are different distance metrics used for measuring different region descriptors. In most cases, the usage of not only the region descriptor but also its distance metric ensures the high performance. It is possible for traditional non linear classifiers to use the kernel functions to reflect the distance metrics of a specific region descriptor as the kernel functions. However, the computational costs of the traditional non linear classifiers are extremely high. To solve this problem, we propose sample pre-mapping as a novel pre-processing step before samples are fed into weak classifiers in Boosting. Sample pre-mapping, from the original input space to another space with the same dimensionality, is a point dependent mapping derived from non linear kernel functions. Linear weak classifiers in the pre-mapped space provide effective and efficient approximations of the optimal separating hyperplanes in the kernel-induced high-dimensional space. Hence the resulting classifiers achieve the accuracy comparable to that of kernel methods, with the computational cost of linear classifiers in both training and detection. Experimental results in both pedestrian detection and car detection on public datasets show that weak learning in the pre-mapped space is capable of achieving accuracy comparable to traditional non-linear classifiers using the compuatational costs of linear classifiers.更多还原