SIA OpenIR  > 机器人学研究室
图像光照处理及目标跟踪算法研究
Alternative TitleResearch on the Illumination Processing of Image and Object Tracking Method
杨大为1,2
Department机器人学研究室
Thesis Advisor唐延东
ClassificationTP391.41
Keyword光照补偿 Retinex理论 稀疏表达 粒子滤波 贝叶斯理论
Call NumberTP391.41/Y27/2014
Pages117页
Degree Discipline模式识别与智能系统
Degree Name博士
2014-11-28
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract人脸识别和目标跟踪是计算机视觉中的热点问题,具有重要的理论研究价值和广泛的应用前景。尽管该课题的研究已取得了长足的进步与发展,提出了若干先进的算法及应用系统,但仍然面临着一些具有挑战性的问题,影响其算法的鲁棒性及应用,如光照变化问题、识别中的遮挡问题等。本文分别从人脸识别中的光照补偿问题、目标跟踪中的光照变化及遮挡问题出发,对基于Retinex理论的人脸图像光照补偿算法、基于稀疏表达、粒子滤波器及朴素贝叶斯理论的目标跟踪算法进行了深入研究。针对现有算法存在的一些问题,本文主要完成以下研究工作: 1)针对目标跟踪过程当中的光照变化及部分遮挡问题,提出了一种在粒子滤波框架内,基于LBP纹理特征,通过稀疏表达进行目标跟踪的方法。算法中采用基本的LBP算法,利用其对光照不敏感的优点来减弱跟踪中光照的影响。稀疏表达算法的性能取决于字典的构造,本算法中的完备字典由两部分组成:n个模板图像的LBP特征列向量和为了克服噪声和部分遮挡引入的微小模板,并在跟踪过程中自适应地更新字典,来减少累积误差。算法中采用通用的重采样方法,采样后的粒子独立分布,新粒子重新分配权值,有效地避免了粒子滤波算法中的粒子退化现象。 2)针对跟踪中的部分遮挡问题,提出了一种结构化的加权联合特征表观模型。该模型将目标图像划分为若干个图像子块。这些图像子块保持固定的空间结构信息。发生遮挡时,根据每个图像子块对跟踪结果的作用不同,为每个图像子块设置空间位置权值;在每个图像子块内分别计算局部的颜色特征和纹理特征,并分别设置颜色特征权值和纹理特征权值;将这些加权后的局部特征向量化,作为目标的表观模型。以该模型为基础,利用贝叶斯理论进行跟踪。在仿射变换空间内,利用随机函数获取不同尺度的候选目标图像,较好地解决了目标跟踪过程中的目标尺度变化问题。 3)提出了基于双边滤波器的Retinex光照补偿算法和基于轮廓波变换的Retinex光照补偿算法。Retinex方法光照补偿的基本假设是光照缓慢变化,但人脸上的阴影违反了这一假设,在不连续的阴影边缘存在晕轮效应。经双边滤波器和轮廓波变换滤波后,得到的光照估计图像能很好地保留了原图像中阴影边缘信息,在Retinxe理论框架下得到的反射图像,有效地弱化了投影阴影的影响。 4)提出了一种结合光照补偿的目标跟踪方法。先用基于Retinex的光照补偿方法对图像进行预处理,然后将结构化的表观模型与模板匹配相结合进行目标跟踪。采用稠密采样的方法在前一帧跟踪结果周围获取不同尺度、不同位置的所有候选目标图像,用扩展的最小中位数平方估计法进行模板匹配,获得当前帧的跟踪结果,并采用逐步更新模板的方法来适应表观模型的的变化,减少跟踪中的漂移现象。
Other AbstractFace recognition and object tracking are the important issues in the computer vision field and have comprehensive applications. Researchers have paid extensive attentions on them, and made much progress and development. A number of algorithms and application systems have been proposed and developed. However, for the robust performance of these algorithms and applications, there are also some challenges to be solved, such as illumination change and occlusion in the recognition. Aiming at these problems, the further study in this dissertation contains two aspects: one is illumination compensation based on Retinex theory, and the other is object tracking algorithm based on sparse representation, particle filtering and Bayesian theory. The main contribution of this dissertation is as follows. 1) Aiming at the problem of illumination variation in the object tracking, an object tracking method with sparse representation in particle filter frame is proposed based on LBP textual feature of object. The method eliminates the illumination effect in the object tracking using the illumination insensitive of LBP. The performance of the sparse representation depends on the formation of the dictionary, the dictionary of the method is consist of two parts: LBP feature vectors of model images and trivial templates to overcome noise and partial occlusion. The dictionary is adaptive updated to reduce the accumulate error. To avoid particle regeneration, the common resampling method is used, thus the particles is independent distributed and assigned new weighted values. 2) To deal with the questions of partial occlusion in object tracking, a structural appearance model with weighted associated features is proposed. The tracked object image is divided into some small image blocks which reserve fixed space structure information. When there is partial occlusion, the role of each block is different and is assigned different space position weighted value. In each block, the color features and textural features are calculated. These features are weighted and a vector is composed, which is taken as the appearance model of the tracked object. Applying the Bayes theory, a weighted Native Bayes tracking method based on the appearance model is proposed. To deal with the scale change in tracking, different scale candidates are obtained using random transform function in affine transformation space. 3) Based on bilateral filter and contourlet transform, a Retinex illumination compensation algorithms is proposed, respectively. The basic assumption of Retinex illumination compensation is that the illumination change slowly and smoothly, while the shadow in the face image violates the assumption, and there is halo effect in the discontinuous edge. After filtering by bilateral filter and contourlet transform, the information of shadow edge in original image can be well preserved in the estimated illumination image. Then the reflection image based on Retinex theory can weaken the effect of cast shadows. 4) A tracking method based on illumination compensation is proposed. Firstly, Retinex based illumination compensation method is applied to the sequence image. Then, the structural appearance model and template matching are used to track the object. The candidates with different scales and different positions are obtained around previous tracking result using dense sampling, then, the current tracking result is obtained using templates matching with extended least median square. To eliminate shift in tracking, a step by step template updating method is applied, which update a small part of the template to adapt to the appearance change.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/16804
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
杨大为. 图像光照处理及目标跟踪算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2014.
Files in This Item:
File Name/Size DocType Version Access License
图像光照处理及目标跟踪算法研究.pdf(6138KB) 开放获取CC BYApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[杨大为]'s Articles
Baidu academic
Similar articles in Baidu academic
[杨大为]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[杨大为]'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.