CASIA OpenIR  > 毕业生  > 硕士学位论文
Alternative TitleObject Tracking based on Online Learning
Thesis Advisor卢汉清
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机技术
Keyword目标跟踪 协同跟踪 排序支持向量机 Object Tracking Collaborative Tracking Ranksvm
Abstract视觉目标跟踪是计算机视觉领域中研究热点之一。它在视频监控、机器人导航、人机交互、军事应用等领域中有着广泛的应用。本文在对国内外存在的方法以及相关技术进行分析与总结的基础上,针对复杂场景、遮挡、光照、消失-再现等问题展开了深入分析和探讨,提出了两个算法和方案,设计并实现了一个实时目标跟踪系统。论文的主要成果和贡献如下: (1) 提出了一种基于短期跟踪器与长期检测器相结合的协同跟踪算法。该方法通过将两种具有不同特性的分类器进行协调融合和动态更新,充分挖掘短期目标外观渐变信息和长期目标外观稳定信息,可以有效处理目标遮挡、消失-再现、渐变等问题。实验结果表明,该方法在一些复杂的环境中能够获得鲁棒的跟踪效果。 (2) 提出了基于改进的样本排序学习目标跟踪算法。该方法对原来的正-负样本对进行扩展,引入了稳定正样本库,提出了介于正样本和负样本之间的零样本概念,组成了正-负样本对、正-零样本对、零-负样本对,从而更有效地刻画目标与背景之间上下文关系,并提出了相应的更新策略。在公开数据集上测试结果表明,本方法能够更好地处理目标遮挡和旋转问题,具有较小的平均中心位置跟踪误差。 (3) 搭建了一个主动视觉目标跟踪平台。通过合理地设计相关数据结构和算法优化,该平台能够实时、鲁棒地进行主动目标跟踪,并能有效处理目标漂移问题。另外,该平台也具有较强的可维护性和可拓展性。
Other AbstractVisual object tracking on real-time is one of the hottest research point in computer vision. It has broad applications and a wide range of future in video surveillance,robot navigation, human-computer interface, military application and so on. Based on the survey of existing methods and techniques, this paper gives further analysis and research, designs and achieves the real-time object tracking system. The contributions of this thesis are three fold: Firstly, in order to mitigate model drift problems, the tracking algorithm of collaborative tracker based on short-term tracker and long-term detector is introduced.It could make use of the characteristics of tracker and detector so as to mine the gradient information of appearance model in the short-term and the stable information for a long time and can handle the problems of object occlusion, disappearance-appearance, gradual change. The experiment results show that this method can be obtained robust tracking results in a complex environment. Secondly, advanced algorithm based on sample rank learning is proposed.The method expands the positive-negative pairs in original paper, and introduces stable positive template library,the zero sample between positive sample and negative sample which will generate the positive-negative pairs, positive-zero pairs, zero-negative pairs for effectively depicting the context relationship between object and background. Testing results on the public data sets show that this method can better deal with the problem of target occlusion and choice, with a smaller average central location tracking error. Lastly, we build up an active visual object tracking platform.By rational design of data structures and algorithm optimization, the platform can be tracking object robustly in real-time and be able to deal effectively with the target drift shift.In addition, the platform also has a strong maintainability and scalability.
Other Identifier2010E8014669002
Document Type学位论文
Recommended Citation
GB/T 7714
朱贵波. 基于在线学习的目标跟踪[D]. 中国科学院自动化研究所. 中国科学院大学,2013.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20102801462800(7010KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[朱贵波]'s Articles
Baidu academic
Similar articles in Baidu academic
[朱贵波]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[朱贵波]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.