CASIA OpenIR  > 模式识别国家重点实验室  > 生物识别与安全技术研究
监控视频开集行人再辨识与重排序研究
王宏升
Subtype硕士
Thesis Advisor雷震
2019-06
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Name工程硕士
Degree Discipline计算机技术
Keyword监控视频 行人再辨识 度量学习 开集行人再辨识 重排序
Abstract

       行人再辨识是智能视频分析中的一项重要技术,可应用于跨摄像机目标关联和图像检索等领域。在监控视频中,由于光照变化大,摄像机角度不同以及图像模糊等因素,行人再辨识仍是一个挑战性的研究问题。

       实际应用场景中,行人再辨识问题通常是一个开集测试问题,即待识别的行人不一定出现在注册图像集中。本论文针对开集行人再辨识展开研究,重点关注以下两方面:

       已有的大多数度量学习算法可以有效提升闭集行人再辨识性能,但面向开集场景时,原有的度量学习算法不是普遍有效,需要针对开集场景下的行人再辨识问题探索有效的度量学习改进方法,提升开集行人再辨识性能。

       重排序技术是基于初始获得的排序列表的一个后处理步骤。当前重排序在闭集行人再辨识问题上对再辨识性能的提升有效,其是否可以在开集行人再辨识问题中获得类似的有效性仍然是未知的。目前行人再辨识领域还没有关于重排序作用于开集问题上的研究工作,因此,重排序也是开集场景下再辨识问题的一个重要研究方向。

       本论文的主要研究贡献包括:

       在开集行人再辨识任务中,为解决开集数据间较大差异引起的跨视角二次判别分析法的正则化方法不能满足开集问题的需求,本文提出通过额外增加一个独立的正则化因子调整子空间的协方差矩阵的方法改进跨视角二次判别分析法,显著提升开集行人再辨识性能。

       通过评估几个主流重排序算法在开集行人再辨识数据库上的对比性能,本文发现现有闭集重排序方法对于开集行人再辨识问题不是普遍有效的现象。针对此现象给出了基于相似度分数分布的解释,进而提出基于得分归一化的重排序算法,能有效提升开集场景下行人的再辨识性能。

       现有大多数开集行人再辨识方法都是基于两两样本之间的表征相似度分数进行识别,没有充分挖掘候选样本之间的相似度信息。本文提出基于相似度分数分布的开集行人再辨识方法,利用待识别图像与候选样本之间的以及候选样本之间的相似度特征向量额外训练分类器,判断待识别图像是否存在于候选集中,进而提高开集行人再辨识的性能。

       综上所述,本文针对监控场景下的开集行人再辨识问题展开研究,提出一种改进的度量学习方法,基于得分归一化的重排序算法以及基于分类的开集行人再辨识算法,提升了开集行人再辨识的性能。

Other Abstract

       Person re-identification is an important technology in intelligent video analysis, which contributes to many areas such as target association from cross-camera views and image retrieval. This is a challenging problem because of large intra-class variations in illumination, pose or viewpoint, and occlusion in the video surveillance.

       In practical application scenarios, the person re-identification problem is usually an open-set test problem, that is, the subject of the probe image does not necessarily appear in the gallery set. This paper focuses on the following two aspects:

       Most of the existing metric learning algorithms can effectively improve the performance of the closed-set person re-identification. However, for the open-set scenarios, the existing metric learning algorithms have not yet perform generally well and it is necessary to explore an effective metric learning algorithm to improve the person re-identification performance.

       Re-ranking technique is a post-processing step based on the initial list. The current re-ranking is effective in improving the re-identification performance for closed-set problem. However, it is still unknown whether existing re-ranking algorithms can achieve similar performance in the open-set person re-identification. At present, there are few researches on the re-ranking for the open-set re-identification, therefore, re-ranking is also an important research direction for open-set re-identification.

       The main contributions of this paper includes:

       It proposes to introduce an additional regularization factor to adjust the covariance matrix of the obtained subspace, and significantly improves the open-set re-identification performance.

       With the evaluation of several popular re-ranking algorithms on challenging open-set person re-identification database, this paper finds that the existing closed-set re-ranking methods have not yet perform generally well. For this phenomenon, this paper provides explanations on the similarity score distributions and then proposes an open-set re-ranking algorithm based on the score normalization, which effectively advances the open-set person re-identification research.

        Most of the conventional open-set re-identification methods compute the feature distances between the query image and all the gallery images and return a similarity ranked table, which ignores the similarity information between candidate samples. This paper proposes a classification based open-set person re-identification method, which uses the similarity feature vector concatenated by the similarity scores between the probe image and the gallery image and the those between gallery images to train a classifier to determine whether the query image exists in the gallery set, thereby impressively improving the result.

       In summary, this dissertation is dedicated to improving the open-set person re-identification problem. An improved metric learning method, a score normalization method in re-ranking algorithm and an explicit genuine/imposter classification method are proposed to address this problem.

 

 

Pages62
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23913
Collection模式识别国家重点实验室_生物识别与安全技术研究
Recommended Citation
GB/T 7714
王宏升. 监控视频开集行人再辨识与重排序研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
监控视频开集行人再辨识与重排序研究.pd(2722KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王宏升]'s Articles
Baidu academic
Similar articles in Baidu academic
[王宏升]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王宏升]'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.