CASIA OpenIR  > 毕业生  > 博士学位论文
基于移动机器人平台的动态视觉跟踪技术研究
Alternative TitleThe Research of Mobile Robot based Dynamic Visual Tracking Technology
郑碎武
Subtype工学博士
Thesis Advisor乔红
2010-06-04
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline控制理论与控制工程
Keyword机器人 动态视觉 目标跟踪 特征提取 流形学习 Robot Dynamic Vision Object Tracking Feature Extraction Manifold Learning
Abstract机器人动态视觉技术作为机器人的眼睛,是智能机器人的关键研究技术,也是目前机器人的研究热点。从应用上来说,该技术具有非常广泛的应用前景,应用平台包括移动机器人平台,车载平台,工业机械手,战斗机及无人机等平台,应用领域包括运动目标检测、跟踪和分析,人机交互,动态智能监控等领域。从技术上来说,不同与传统意义上的视频监控,由于视觉系统的动态性,基于机器人的动态视觉跟踪技术有其自身的特点: 1、机器人本体(视觉平台)的运动加快了视觉影像中背景和光线的变化,从而导致了稳定特征提取的困难; 2、机器人平台和被跟踪目标较大的相对运动增加了目标表征的变化及对分类跟踪算法的快速性要求; 3、移动机器人的实时控制要求视觉任务具有快速性和准确性。 针对以上特殊性,本论文对机器人动态视觉分类跟踪方法展开了深入的研究。 首先,论文从多特征融合角度出发,试图在动态复杂环境和目标自由运动情况下提取稳定有效的多特征组合,以实现鲁棒的机器人动态视觉技术。 其次,考虑到物理特征本质上容易受光线和复杂环境影响的缺陷,论文提出了一种抽象本质特征,并将流形学习框架首次引入到动态视觉中,获得了稳定有效的特征提取和单目标跟踪技术。 另外,考虑到机器人在复杂背景情况下对多目标跟踪的需求,论文基于保持本质变量连续性的流形学习方法,引入局部判别信息,实现了复杂环境下,稳定的多目标跟踪和遮挡处理技术。 最后,从机器人动态视觉系统对实时目标检测算法的需求和机器人动态跟踪系统的稳定性出发,论文提出了一种具有在线更新能力的分类检测算法,该算法可以对环境中新进目标进行实时检测和初始化,同时也可以和跟踪算法结合起来,实现更稳定的机器人动态视觉跟踪技术。 论文的主要研究内容和创新点如下: (1) 针对机器人动态视觉跟踪中稳定和有效的特征提取困难,本论文从多特征融合的角度出发,分析了基于移动机器人平台的视觉跟踪稳定特征提取的困难,提出了一种适用于人体头部跟踪的鲁棒的时空组合特征及其更新方法。该方法通过空间结构关系和运动连续性把多种特征组合起来,克服了传统多特征融合方法丢失空间结构信息,不能表征目标任意运动姿态等问题。进一步,基于该时空组合特征,本论文实现了贝叶斯框架下稳定的针对人体头部的机器人视觉跟踪算法。不同条件下的视频测试和复杂背景下机器人平台的跟踪实验证明基于该时空组合特征的贝叶斯跟踪算法稳定有效。 (2) 针对机器人动态视觉跟踪中物理特征在本质上容易受光线和复杂环境影响的缺陷,本论文分析了运动人体目标在跟踪过程中的本质运动流形,提出了一种通过非线性降维框架和重构高维空间的流形,获得可以最优保持本质变量连续性的低维特征子空间的流形学习方法(Intrinsic Variable Preserving Manifold Learning method: IVPML)。不同于现有的基于保持局部欧式距离的非线性降维方法,我们利用了本质运动变量构造了适用于目标跟踪的新流形空间,在非线性降维过程中,提取了保持目标本质变量连续性的低维特征子空间。该方...
Other AbstractAs the eyes of robots, Robotic Dynamic Vision Technology (RDVT) is one of the key techniques for intelligent robots, and is currently a hot topic in the research fleld of robotics. From the application point of view, RDVT has wide applications in various dynamic systems, which include mobile robot system, driverless vehicles, industrial robots, aircrafts and unmanned aerial vehicles. Its application range also covers the flelds of object detection/tracking/analysis, human-robot interaction and intelligent dynamic surveillance. From the technical point of view, difierent from traditional video surveillance techniques, visual tracking based on dynamic robots has its own characteristics due to the dynamic nature of vision systems, which are summarized as below. 1、The movement of the robot (visual platform) speeds up the changes of both the background and the illumination, which may cause difficulties in extracting a stable feature. 2、A large relative movement between the robot and the object may cause a large changes of object's location in the image, which will increase the computation complexity of tracking algorithms. 3、The real-time control of dynamic robots requires that the visual system should have high processing speed and accuracy. To address the above difficulties, in this thesis we have carried out a deep study on visual classiflcation and tracking methods for dynamic robots. First, in order to extract efficient fusion of multi-features in dynamic and complicated environments, we take the approach of fusing multi-features to achieve robust RDVT. Second, since physical features are sensitive to illumination and complicated environment, in this thesis we propose an abstract and intrinsic feature, and introduce manifold learning framework to dynamic visual tracking, which yields stable and efficient feature extraction as well as monocular visual tracking techniques. Third, considering that robots need to track multi-objects under complicated background, based on intrinsic variable preserving manifold learning method and the local discriminative information, we achieve a stable multi-objects tracking algorithm as well as a occlusion handling mechanism in complicated environment. Finally, motivated by the demand of real-time object detection algorithm and robust visual tracking techniques in robotic dynamic vision system, we propose a classiflcation/detectin algorithm which can be updated online. The proposed algori...
shelfnumXWLW1513
Other Identifier200718014628082
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6295
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
郑碎武. 基于移动机器人平台的动态视觉跟踪技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2010.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20071801462808(5649KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[郑碎武]'s Articles
Baidu academic
Similar articles in Baidu academic
[郑碎武]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[郑碎武]'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.