目标跟踪中的特征选择与更新方法研究
文献类型：学位论文
| 作者 | 刘荣 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2010-05-20 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 卢汉清 |
| 关键词 | 目标跟踪 特征选择与更新 产生式模型 判别式模型 半监督学习 Object tracking feature selection and update generative model discriminative model semi-supervised learning |
| 其他题名 | Research of Feature Selection and Update in Object Tracking |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 目标跟踪是计算机视觉中的研究热点之一，它在视频监控、人机交互以及多媒体分析等多个领域都有着广泛的应用前景。特征选择与更新是目标跟踪中的关键技术，许多对视觉目标跟踪的研究都集中在它上面。现有的特征选择与更新方法大致可以分为两类：一类是基于产生式模型的，它关注对目标本身的描述；另一类是基于判别式模型的，它关注目标与背景之间的区分性。本文以解决跟踪中的复杂背景、目标表象变化以及遮挡问题为目的对这两类方法都进行了深入的研究，主要贡献包括以下两个方面： 1. 基于判别式模型的特征选择与更新。这类方法运算复杂度较高但算法的鲁棒性很强，在本文的研究中主要关注它的性能。 1) 在半监督的框架下提出了一种特征在线协同提升算法并用于目标跟踪的特征选择与更新。该算法在更新过程中不需要样本标签，避免了由于跟踪结果不准确而引入错误样本导致误差累积，从而保证了整个跟踪过程中特征更新的正确性。本文在理论上证明了提出的算法是一种不仅极小化了训练误差上界而且也极小化了泛化误差上界的优化算法。在这基础上，本文还利用贝叶斯理论框架对算法进行了解释与统一。 2) 在弱监督的框架下提出了一种基于多特征学习的在线特征提升算法用于目标跟踪的特征选择与更新。该算法在更新过程中不需要非常准确的目标位置只需要一个大致的目标范围，从而有效的避免了由于跟踪结果不准确而带来的累积误差。提出的算法把多示例学习的思想应用到每个特征上增加了学习时特征的自由度，从而提高了自由度比较大的目标的跟踪鲁棒性。 2. 基于产生式模型的特征选择与更新。这类方法的运算复杂度相对较低，是实际目标跟踪系统中常用的一类方法。 1) 针对跟踪中的复杂背景问题，本文提出了一种分块显著特征选择的方法；针对跟踪中的目标表象变化问题，本文提出了一种轮廓约束下的特征更新方法；针对跟踪中的遮挡问题，本文提出了一种基于上半身信息的高斯模型更新方法以及一种身份信息辅助的跟踪特征更新方法。 2) 提出了一种基于特征模型反向匹配的误跟（track loss）判断方法，并将其应用到目标特征的更新中。这种更新方法在保证了目标特征的适应性的同时有效地缓解了跟踪中累积误差带来的问题。本文在稳定婚姻匹配的理论框架下证明了提出的误跟判断方法的合理性。为了保证实验评价的公正性与全面性，本文提出了基于ROC曲线的误跟判断评价方法。 |
| 英文摘要 | Object tracking is one of the hottest research topics in computer vision community. It has promising applications in intelligent surveillance, human-computerinterface, multi-media analysis, and so on. Feature selection and update is the core technology in object tracking. It can be roughly categorized into two classes: (1) generative model based approach, which focuses on the description of object; (2) discriminative model based approach, which focuses on the discrimination between object and background. In this dissertation we make an intensive study on the two classes of feature selection and update approaches for the purpose of handling some challenging problems in object tracking, such as background clutter, object dynamics, occlusion. The main contributions of this dissertation can be summarized as follows: 1. Discriminative model based feature selection and update. This kind of methods possesses strong robustness but high computational complexity. This dissertation mainly focuses on its capability of discriminating object from background. 1) A semi-supervised boosting algorithm in a co-training framework is proposed to handle the feature selection and update problem in object tracking. It treats new samples as unlabeled in update, and can avoid accumulative error caused by inaccurate tracking results. The proposed algorithm is proved to not only minimize the training error bound but also the generalization error bound. Moreover, a Bayesian perspective is presented to unify the proposed method in a Bayesian boosting co-training framework. 2) A weakly supervised boosting algorithm based on multiple features learning is proposed to handle the feature selection and update problem in object tracking. Since only a coarse scope of target position is demanded in its feature update, it can avoid accumulative error caused by inaccurate tracking results. The proposed algorithm increases the degree of freedom for each feature through multiple features learning, and so can improve the tracking capability for the object with high degree of freedom. 2. Generative model based feature selection and update. Due to the low computational complexity, it is commonly used in real tracking systems. 1) To handle the problems in real tracking systems, a series of feature selection and update methods are proposed: a blocked dominant feature selection method is proposed to handle background clutter; a contour constrained feature update method is proposed to handle object dy... |
| 语种 | 中文 |
| 其他标识符 | 200618014628046 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6239]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 刘荣. 目标跟踪中的特征选择与更新方法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2010. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。