【作者】 刘倩；

【导师】 侯建华；

【作者基本信息】 中南民族大学 ， 通信与信息系统， 2013， 硕士

【摘要】 智能视频监控是机器视觉领域一个颇具应用价值的研究方向，被广泛地应用于航空、交通、公共安全等方面。作为视频监控中最基础的环节-目标的检测与跟踪技术，一直以来更是人们关注的热点问题。目标是否能够被成功跟踪主要取决于能否成功将目标从它周围环境中分割出来，因此可以把目标的检测与跟踪问题转换成二分类问题，只要找到合适的特征和判别分类器就可以实现高效准确的跟踪。然而，视频序列中目标的准确检测与跟踪是一个颇具挑战性的任务，其中制约跟踪效果的最大因素莫过于目标的外观变化，包括姿势和形状等内因变化，以及光照、移位、相机视角及遮挡等外因变化。针对以上问题，本文对目标检测与跟踪展开研究，从机器学习的角度，将目标的检测与跟踪视为一个二分类问题。本文的工作主要包括两部分。第一，目标模型的表述。为了全面有效地表述目标，深入分析了尺度不变特征变换（ScaleInvariable Feature Transform，SIFT）、PCA-SIFT、SURF（SpeedUp Robus tFeature）、DAISY等特征算子，对比各种特征算法的优缺点和应用场合，最终选择利用多尺度空间的高斯差分（DoG）检测图像中的特征点，然后利用DAISY特征描述子对目标图像进行描述。使目标模型对旋转、视角、尺度具有不变性，对光照和局部遮挡有一定的稳定性。第二，分类器的设计。重点研究了近年来热门的随机森林分类器，在此基础上，引入霍夫变换的随机森林，即霍夫森林（HoughForest）作为区分目标和背景的判别分类器。将图像块和其对应的类别标记作为霍夫森林分类器的训练数据，通过结构性的训练和学习，生成霍夫森林相应的结构参数，把目标图像块和背景图像块分开，相应的霍夫森林的各叶节点就形成了判别式的码本集。为了提高目标模型的对目标和背景变化的自适应性，文中提出了一种新的模型更新方法，即在后续视频序列中，利用新标记的图像块与模型码本中的图像块的相似性测量对模型进行更新。实验证明，将快速有效的DAISY描述子和识别率高、鲁棒性强的霍夫森林分类器相结合，在处理局部遮挡、消失目标重现检测和低分辨率图像等场合具有显著的效果，同时新的模型自适应更新的方法使算法对复杂场景下的目标检测与跟踪有很大的优势。更多还原

【Abstract】 Intelligent video surveillance is a valuable research in machine vision field, and is widely applied in the field of aviation, traffic and public security, etc. As the most basic component in video surveillance, object detection and tracking has attracted more and more attention. Whether a tracking is successful or not mainly depends on the degree in which an object is segmented from its surroundings. In this dissertation, the problem of detection and tracking has been treated as a binary classification, and efficient and accurate tracking can be realized by suitable features and the design of discriminative classifier.However, object detection and tracking is a very challenge task due to target appearance variability, which is the primary restricting element, including pose variation and shape deformation, as well as illumination change, camera motion, viewpoint change, and occlusions. The problems above mentioned have been studied in this paper in a machine learning point of view, in which detection and tracking is regarded as a binary classification problem.The thesis consists of two parts as followed. The first part is the description of target model. In order to represent the target model comprehensively and effectively, several features are analyzed in-depth such as Scale Invariable Feature Transform (SIFT), PCA-SIFT, SURF (Speed Up Robust Feature), and DAISY etc. The merits and drawbacks of each feature are expounded respectively, and their application scenarios are indicated. Difference of Gaussian in muti-scale space is selected for feature points detection. Considering invariant to image rotation, scale, change in viewpoint, change in illumination, and partial occlusion,DAISY feature descriptor is chosen to describe the object image.The second part is the design of classifier. Following the discussion of random forest which is popular in recent years, the thesis has elaborated Hough Forest (HF). Hough Forest combines random forest and hough transform, and serves as classifier to separate the target from background. Hough forest classifier is trained by image partial patches and their labels, and the structure parameters of Hough forest are generated to classify the object patches and background patches. The set of leaf nodes of each tree in the Hough forest can be regarded as a discriminative codebook. In order to improve the adaptability of the target model, the paper proposes to update the discriminative codebook by computing the similarity measurement between the detection results of the video coming in sequence and the codebook.Experiment results show that the proposed algorithm, combining efficient DAISY feature descriptor and robust discriminative Hough Forest which, has satisfactory tracking precision and good real time performance, and works well under the condition of partial occlusions, different image resolutions, and complex scenes.更多还原