CASIA OpenIR  > 毕业生  > 博士学位论文
自然场景下视觉目标跟踪的关键技术研究
Alternative TitleResearch on Visual Object Tracking in Natural Scenes
马丽莉
Subtype工学博士
Thesis Advisor卢汉清
2010-06-07
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模式识别与智能系统
Keyword目标跟踪 均值漂移 粒子滤波 轮廓跟踪 视觉注意模型 Object Tracking Mean Shift Particle Filter Contour Tracking Visual Attention Model
Abstract目标跟踪是计算机视觉领域的研究热点之一,在视觉监控、人机交互、智能交通等诸多领域中有着广泛的应用前景和商业价值。尽管相关研究已经进行了几十年时间,但是在复杂的自然场景下进行目标跟踪,仍然有许多理论与技术问题亟待解决。这些问题包括:跟踪过程中的遮挡、背景干扰、目标的复杂运动和目标形状与表观变化等。本文针对以上问题,首先从目标表观模型及目标运动约束等角度进行了研究,提高了基于核的目标跟踪方法在遮挡、背景干扰和目标复杂运动下的跟踪鲁棒性;另外,通过考虑形状先验,解决了遮挡及目标尺度与形状变化时的轮廓跟踪问题;最后将视觉注意机制引入目标跟踪中,强化目标与背景的差异,为目标跟踪问题提供有效的先验信息。论文的主要贡献归纳如下: 1. 针对传统均值漂移方法中直方图目标模型的空间信息缺失问题,提出了一种基于分块直方图的均值漂移目标跟踪方法。分块直方图的目标表观模型可以保留目标的空间信息,提高在背景干扰和部分遮挡时的跟踪鲁棒性。另外针对均值漂移的目标跟踪方法对初始点敏感的问题,本文提出了一种融合局部模式搜寻的均值漂移目标跟踪方法。该方法通过在目标区域选择合适的模式,并进行局部模式搜寻来进行均值漂移的初始化,提高了在目标快速运动时的跟踪鲁棒性。最后,将以上两种方法融合,进一步提高了在复杂的自然场景下的跟踪鲁棒性。 2. 针对自然场景下目标的复杂运动,本文提出了一种基于交比不变约束的多核协同跟踪方法。该方法通过多核的策略将目标的复杂运动分解为多个简单运动,并通过加入交比不变约束提高了多核系统的可观测性。交比不变约束在适应目标复杂运动的同时,提供给多核系统一个足够严格的约束,从而提高了跟踪复杂运动目标时的鲁棒性。 3. 提出了一种融合粒子滤波与三维图割的目标轮廓跟踪方法。在该方法中,粒子滤波跟踪目标的位置,而三维图割在目标位置附近进行目标分割。传统的基于直方图目标模型的粒子滤波跟踪方法,损失了目标的空间信息且引入了背景像素的干扰,而图割方法具有全局性质,亦即背景中与前景分布类似的区域会被分割为前景,不利于目标轮廓的提取。为了克服这些问题,本文将目标的形状先验,即指导模板,分别融合到粒子滤波和三维图割中,提高了跟踪的鲁棒性和目标分割的准确性。同时将指导模板进行实时的选择性更新,解决了遮挡问题,并适应了目标的形状与尺度变化。实验证明,该方法提高了在遮挡及目标尺度与形状发生较大变化时的跟踪鲁棒性。 4. 提出了一种基于视觉注意模型的目标跟踪方法。该方法融合了自上而下、任务相关和自下而上、刺激驱动的视觉注意机制。首先根据自下而上的视觉注意模型,将图像分解为不同的特征图,然后利用目标跟踪中的先验知识,即目标区域比背景区域更受关注的假设,采用逻辑回归的方法对特征图进行调制,加强目标区别于背景的特征,并抑制背景相关的特征,从而得到目标与背景差异更大的显著度图,之后通过一种有效的搜索策略在显著度图上定位目标位置。实验表明该方法可以比较鲁棒地跟踪目标。另外,将该方法得到的显著度图融合到均值漂移等传统的目标跟踪方...
Other AbstractObject tracking is one of the hottest topics in computer vision community. It has been widely applied in video surveillance, human-computer interface, intelligent transportation and so on. Although research on object tracking has been conducted for several decades, there are still many challenges and difficulties, for example, occlusion, clutter background, complex scenes and object appearance and shape changes. In this dissertation, novel object appearance model and motion constraint are employed to improve the robustness of kernel based tracking method in the cases of occlusion, clutter background and object complex motion; shape prior is exploited in contour tracking to deal with occlusion and object shape changes; visual attention model is introduced into object tracking to enhance the difference between object and background, which provides effective prior information for object tracking. The main contribution of this dissertation is as follows. 1.A robust mean shift object tracking algorithm via fragment based representation is proposed to compensate for the spatial information loss in histogram. The spatial information of the object is reserved in the fragment based representation, which makes tracking much more robust especially when the object is partially occluded or there are attractive objects in the background. Meanwhile, a mean shift object tracking algorithm integrating local mode seeking is proposed to deal with the problem that mean shift is sensitive to initialization. The proper modes are chosen in the object region and local mode seeking is used to initialize the mean shift iteration, which makes tracking much more robust even when the object is moving fast. Finally, combining the above two algorithm can make tracking even more robust in complex natural scenes. 2.A multiple collaborative kernel tracking algorithm with cross ratio invariant constraint is presented. The complex motion of object is decomposed into simple motions using multi-kernel strategy and the cross ratio invariant constraint improves the observability of the system. Object under complex motion can still be tracked by the algorithm. 3.We present a novel contour tracking approach that combines particle filter and 3D graph cut. Particle filter predicts the location of the object while 3D graph cut segment the object near the predicted location. Traditional histogram based particle filter does not reserve the spatial information and is easy to be affected by the backg...
shelfnumXWLW1453
Other Identifier200618014628073
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6299
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
马丽莉. 自然场景下视觉目标跟踪的关键技术研究[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_20061801462807(5277KB) 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
[马丽莉]'s Articles
Baidu academic
Similar articles in Baidu academic
[马丽莉]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[马丽莉]'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.