CASIA OpenIR  > 毕业生  > 博士学位论文
变视场情况下的视觉跟踪技术研究
Alternative TitleResearch on Visual Object Tracking under Changing Field of View
宋翼
Subtype工学博士
Thesis Advisor杨一平
2015-05-29
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计算机应用技术
Keyword目标跟踪 视场渐变 视场切变 尺度变化 位移跳变 前景占有比 Mean Shift Object Tracking Gradually Changed Fov Fov Switching Scale Change Abrupt Motion Foreground Occupation Ratio Mean Shift
Abstract视觉目标跟踪是计算机视觉中最活跃的研究内容之一,其在无人机侦察、机器人自动控制、汽车自动驾驶等方面有着重要意义。在相机拍摄过程中,相机和感兴趣目标的运动不确定性,相机内部参数的任意性,以及拍摄环境的复杂多变性,都对视频中的目标在线实时准确跟踪算法提出了重重挑战。由于拍摄主体对被摄物体分辨率精度的要求,相机视场发生变化的情况非常普遍,使得目标在相机视野中的尺度发生大范围缩放变化,以及随之发生外观变化和大位移运动。国内外研究者当前还未针对相机视场变化情况下的目标跟踪进行系统性地研究,现有跟踪算法在这种情况下的准确性、稳定性和效率存在很多问题,难以满足实际应用需求。本文系统性地针对相机在两种变视场情况下的视觉目标跟踪和定位方法展开深入研究,主要工作包括: (1)针对相机视场渐变情况下的目标实时跟踪问题,提出了一种基于Mean Shift与帧间局部特征快速匹配的尺度适应性目标跟踪算法。相机视场渐变情况下,目标的尺度和位移在连续帧间发生平缓变化。所提方法通过帧间目标的局部特征匹配来解算帧间仿射变换模型,获得目标在帧间的尺度缩放系数。将该尺度计算结果融入Mean Shift定位算法后,加入尺度估计的调整策略,能使其在跟踪时准确估计目标的尺度。由于使用快速的FAST角点检测以及LHOG描述方法进行局部特征提取,算法计算效率较高。实验表明,该算法能快速、准确地定位目标,并很好地进行目标尺度估计。 (2)提出了基于前景占有比率不变性的Mean Shift算法,在相机视场渐变情况下进行实时目标跟踪。文中提出了图像的前景占有比率(FOR)的概念,并总结了其具有的四个简单性质。根据局部图像FOR在帧间的不变性,可以获得帧间前景目标的粗尺度缩放系数;加入合适的尺度修正及调整策略后,可以得到前景目标的精尺度。将之应用到Mean Shift跟踪算法中,使其能在多种环境下均能准确定位目标,同时很好地估计目标尺度大小,并达到实时跟踪的要求。相比(1)中方法,其在目标局部特征难以被提取的条件下同样适用。 (3)针对相机视场切变处的目标锁定问题,提出了一种基于级联Mean Shift算法的目标定位器。在相机视场切变处,目标的尺度和位置都发生了较大变化。为解决这种情况下的目标定位问题,所提方法根据这种特殊应用中的先验信息将切变后帧进行下采样处理,并在处理后的图像中使用多核窗Mean Shift定位器级联定位方式来求取大位移变化下目标的准确位置。该方法同时使用了背景加权直方图建模以及权值改进技术,使得目标的定位更为快速准确。此外,方法根据获取到的候选目标与切变前帧中目标模型的相似性,来确认目标是否已消失于切变后帧中。实验表明,该方法能够有效地处理目标在视场切变处发生的大位移变化,快速准确地定位切变后帧中的目标。 (4)针对相机视场切变后目标消失于相机视野中,需要进行相机回切来重新定位原目标的情况,提出了一种基于自适应前景占有比率的目标定位器来实现目标重定位。这种情况中,目标在相机回切后帧中的位置相对于切变前帧发生了剧烈变化。所提方法首先通过评估切变前帧中局部图像区域的FOR值来...
Other AbstractVisual object tracking is one of the most active research topic in the field of computer vision, and it plays a significant role in applications such as unmanned aerial vehicle surveillance, robotics and driver assistance. Accurate online visual tracking faces great challenges due to the uncertainty of the motion between the camera and the target, the previously unknown parameters in the camera model, and the clutters in the circumstances for video capturing. During the capturing, the field-of-view (FOV) of the camera may be changed frequently to obtain different resolutions of the target. In these situations, the scale change of the target occurs a lot in the captured image sequences, along with the appearance change and the abrupt motion of the target. By now, researchers have not yet systematically studied the online visual tracking in situations where the FOV of the camera changes, resulting in unsatisfactory accuracy, robustness and efficiency of the tracking algorithms when dealing with these situations in real applications. In this dissertation, we lucubrate object tracking algorithms in dealing with two situations of FOV change of the camera. The main contributions are listed as follows: (1) A scale adaptive tracking method using Mean Shift and efficient feature matching is proposed to deal with the tracking problem in the situation where the FOV of the camera gradually changes. In this situation, the scale and the location of the target gradually change between consecutive frames. The scale estimation module calculates the affine model by matching local features extracted in the target between consecutive frames to obtain the scaling factors. This module is modified and integrated with the Mean Shift tracker to give accurate results for the target scale in tracking. The proposed method utilizes FAST corners and LHOG description for the good efficiency in feature matching. Experimental results show that the tracker locates the target accurately and efficiently, and gives satisfying results for the scale of the target. (2) A scale adaptive Mean Shift tracker based on invariant foreground occupation ratio is proposed to deal with the tracking problem when the FOV of the camera gradually changes. The thesis of foreground occupation ratio (FOR) of the image is proposed, and its four simple properties are summarized. With its invariance between two frames, the coarse scaling factor of the foreground is obtained, and it is modified and adjusted prope...
Other Identifier201218014629097
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6727
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
宋翼. 变视场情况下的视觉跟踪技术研究[D]. 中国科学院自动化研究所. 中国科学院大学,2015.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20121801462909(8244KB) 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
[宋翼]'s Articles
Baidu academic
Similar articles in Baidu academic
[宋翼]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[宋翼]'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.