QQ客服 官方微博 反馈留言

题名	目标跟踪关键技术研究及应用
作者	吴毅
学位类别	工学博士
答辩日期	2009-05-20
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 粒子滤波 协方差区域描述子 多视角说话人脸 球员跟踪 Object tracking, particle filter covariance region descriptor multi-view speaking face player tracking
其他题名	Research and Application of 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领域中研究热点之一。在智能监控、人机交互、视频检索等领域中有着广泛的应用前景和商业价值。本文在理论上对目标跟踪的关键技术进行了研究，基于一种鲁棒的区域描述子提出了一些有效的算法；另一方面，针对特定的应用，本文从如何充分利用先验知识（目标的先验模型）的角度来研究如何提高实际应用中跟踪的鲁棒性。论文的主要贡献归纳如下： 1. 提出一种基于协方差区域描述子的贝叶斯跟踪算法。协方差描述子可以融合多种底层视觉特征，该算法将协方差区域描述子融入到贝叶斯框架下，处理复杂背景下的目标跟踪问题。另一方面，协方差快速计算方法被扩展到贝叶斯跟踪框架下，提高了跟踪的效率。 2. 提出了一种基于分块目标表示的黎曼流形上的贝叶斯跟踪算法。将目标的分块表示扩展到贝叶斯跟踪框架下，提高了跟踪的鲁棒性，特别是在处理目标部分遮挡情况下的鲁棒性。 3. 提出了一种增量协方差张量学习算法，并用于目标跟踪中的模型更新。该算法采用增量的方式学习一个低维紧致协方差张量目标表示，可以快速有效的在线适应目标表观的变化。模型更新中采用了一种加权的策略，通过降低以前观测对模型的贡献、增加当前观测贡献来提高跟踪的鲁棒性。 4. 提出了一种基于Boosted协作分布式粒子滤波的自动多目标跟踪算法，较好的解决了足球视频中球员间的遮挡问题。基于前景观测提出的势能函数，能有效对目标间的相互遮挡建模。在粒子滤波实现上，采用一种综合了动态模型和Boosting检测信息的混合模型作为建议分布，这不仅能够快速检测新球员而且也提高了跟踪性能。 5. 提出了一种基于视觉的多视角说话人脸跟踪框架，并给出了一种用于影视剧分析的多视角说话人脸跟踪方法。该方法在人脸跟踪上后，通过人脸匹配实现嘴部配准，并利用一种归一化描述子刻画嘴部的变化，判别目标是否在说话。
英文摘要	Object tracking is one of the hottest research topics in computer vision. It has broad application and commercial value in intelligent surveillance, human-computer interface, video retrieval, and so on. In this dissertation we focus on the key techniques of object tracking and based on a robust region descriptor some effective algorithms are proposed. For the tracking of specific object, we use the prior knowledge (the prior model of the target) to improve the tracking performance. Following is the main contribution of this dissertation: 1. A Bayesian tracking algorithm based on the covariance region descriptor is proposed. The covariance descriptor can fuse different low level visual features and the proposed algorithm encodes it in the Bayesian tracking framework to handle difficult background. Moreover, the fast covariance computation is extended to Bayesian tracking framework, which makes the tracking process more efficient. 2. We present a robust Bayesian tracking approach on Riemannian manifolds via fragments-based representation. The fragments-based representation is extended to the Bayesian tracking framework to improve the tracking performance, which is especially robust to partial occlusion. 3. An incremental covariance tensor learning algorithm is proposed, which is used to update the object model during tracking process. It incrementally learns a low-dimensional covariance tensor representation, efficiently adapting online to appearance changes for the target. Moreover, a weighting scheme is adopted to ensure less modeling power is expended fitting older observations. 4. We present a multi-object automatic tracking approach based on boosted interactively distributed particle filter, which can handle the occlusion problems among targets. The foreground is used to develop a data-driven potential function to model the occlusion among targets. In the particle filter implementation, the proposal distribution using a mixture model that incorporates information from the dynamic model and the Boosting detection is adopted. The Boosting proposal distribution allows us to quickly detect targets and improves the tracking performance. 5. A framework for vision-based speaking face tracking is proposed. Based on this framework a multi-view speaking face tracking approach for video analysis is presented. After the face is tracked mouth is aligned through face matching. Then a novel descriptor is introduced to describe the change of the mouth and identify...
语种	中文
其他标识符	200618014628063
内容类型	学位论文
源URL	[http://ir.ia.ac.cn/handle/173211/6148]
专题	毕业生_博士学位论文
推荐引用方式 GB/T 7714	吴毅. 目标跟踪关键技术研究及应用[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2009.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

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。