CASIA OpenIR  > 毕业生  > 博士学位论文
引入视觉注意机制的目标跟踪方法研究
Alternative TitleVisual tracking using visual attention mechanism
黎万义
Subtype工学博士
Thesis Advisor乔红 ; 王鹏
2013-12-26
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控制理论与控制工程
Keyword视觉跟踪 视觉注意 显著性 稀疏编码 Visual Tracking Visual Attention Saliency Sparse Coding
Abstract目标跟踪是一个重要的计算机视觉问题,在智能监控、机器人、媒体制作与增强现实等领域有着广泛的应用。尽管近几十来,研究者们展开了大量的研究,目标跟踪技术取得了长足的发展,然而由于突变运动、遮挡、目标外观变化和背景干扰等因素的存在,使得复杂场景下的鲁棒实时的目标跟踪仍是一项极具挑战性的任务。视觉注意是人类视觉系统的一项重要的心理调节机制,在视觉注意的引导下,人类能够从众多的视觉信息中快速地选择那些最重要、最有用、与当前行为最相关的感兴趣的视觉信息,特别地,人类能够快速指向感兴趣的目标,从而可以轻松地实现对目标的稳定跟踪。因此,将视觉注意机制引入到复杂场景下的目标跟踪中,研究稳定且接近于人类认知机制的视觉跟踪算法,具有重要的理论意义与应用价值。 本论文对“引入视觉注意机制的目标跟踪方法”进行了深入研究。论文得到了国家自然科学青年基金项目“引入视觉注意机制的动态场景下视觉跟踪方法研究” (61100098)、国家自然科学基金重大国际合作项目:“视觉认知-反应”融合神经系统可计算模型及其在机器人中的应用——面向重大任务的智能决策研究”(61210009)和国家自然科学基金面上项目:“大型激光驱动器光机组件视觉检测与定位研究”(61379097)的资助。主要研究内容和贡献如下: (1) 对已有的“基于视觉注意机制的目标跟踪方法”进行了综述。 对该类方法进行归纳、总结和分类,对该类方法的特点和优势进行了讨论,同时对该类方法值得深入研究的问题进行了探讨。 (2) 针对“如何快速将视点引导到目标可能出现的区域的问题”,提出了基于频谱分析的自上而下视觉注意计算模型。该注意模型将自上而下的信息引入到频谱分析注意模型中,能够将处理资源分配到与目标最相关的显著区域。 (3) 针对“目标突变运动和长时遮挡导致跟踪方法失效的问题”,提出了一种基于频谱分析视觉注意模型的目标跟踪方法。自上而下的视觉注意计算显著图,在所得显著图的指导下,进行局部目标搜索和全局目标搜索,搜索结果的可靠性通过验证模块进行验证。该框架可以鲁棒地处理突变运动和目标经过长时遮挡或完全移出视野后的重新出现等问题。 (4) 针对“传统粒子滤波处理突变运动及长时遮挡所需粒子数量呈指数增长,导致指数增长的计算复杂度的问题”,提出了一种视觉注意与粒子滤波交叉改进的跟踪方法。使用视觉注意计算模型在跟踪之前先获得目标可能出现的区域,基于检测到的显著区域,提出新的采样提议分布,使得采样充分考虑目标很有可能出现的显著区域。该方法可有效处理突变运动和长时遮挡。 (5) 针对“目标外观建模可能引入噪声”的问题,提出了一种视觉注意加权的稀疏编码外观模型。使用基于谱滤波的自上而下和自下而上结合的视觉注意计算模型计算出图像的显著图,将计算得到的显著图对编码进行加权,从而得到候选区域的稀疏编码后的特征向量。显著性的加权操作,起到了突出前景物体而抑制背景的作用。对比实验表明,提出的方法优于没有使用显著性加权的稀疏编码方法,并能有效处理背景杂乱,光线变化等问题。
Other AbstractVisual object tracking is an important computer vision task which can be applied to many domains such as visual surveillance, robotics, media production and augmented reality. Despite extensive research on this topic in recent decades, achieving robust tracking performance still remains a huge challenge regarding abrupt motions, occlusions, changing appearance patterns of the object and background interference, etc. Visual attention is one of the key mechanisms of human visual system which directs the processing resources to the visual data of the potentially most relevant, specially directs our gaze rapidly towards objects of interest in our visual environment and as a result humans can easily achieve stable object tracking. Therefore, introducing the visual attention mechanism to the object tracking in complex scenes to achieve stable and humanoid tracking algorithms, has important academic significance and application value. This thesis focuses on visual attention based tracking methods. This research was partly supported by National Natural Science Foundation of China (61210009, 61100098, 61379097). The key research contents and contributions of this thesis can be summarized as follows: (1) The state-of-the-art visual attention based methods for tracking are reviewed. The attention-based visual tracking algorithms are classified into five categories and detailed descriptions of representative methods in each category are provided, and their pros and cons are examined. Besides, we highlight the advantages of attention-based tracking methods and provide insights for future. (2) To cope with the problem that how to direct the gaze to salient regions where target may appears, we propose a new computational model, i.e., top-down frequency analysis visual attention model which first introduces top-down information to frequency analysis attention model and can direct the processing resources to the salient regions which are most relevant to the target need to pop out. (3) We present a novel spectral analysis visual attention based object tracking method. For simulating top-down visual attention mechanism and calculating saliency maps, we first introduce top-down information to spectral analysis attention model. After calculating saliency maps, local target search and global target search are performed based on these saliency maps. The proposed method can cope with very difficult situations including abrupt motion and target reappearing after longtime oc...
shelfnumXWLW1952
Other Identifier201018014628008
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6574
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
黎万义. 引入视觉注意机制的目标跟踪方法研究[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_20101801462800(4624KB) 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
[黎万义]'s Articles
Baidu academic
Similar articles in Baidu academic
[黎万义]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[黎万义]'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.