CASIA OpenIR  > 毕业生  > 硕士学位论文
基于中层特征表达的目标识别技术研究
Alternative Titlemid-level feature representation learning for object recognition
郑帅
Subtype工学硕士
Thesis Advisor黄凯奇
2011-05-24
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目标识别 特征表达 步态识别 多摄像机跟踪 Object Recognition Feature Representation Gait Recognition Multi-camera Tracking
Abstract目标识别作为计算机视觉中的核心问题以及智能视频监控中的关键技术,得到了广泛的关注和重视,该技术的目标是让计算机能够智能识别出图像中出现的物体。具体而言,在智能视频监控系统中,目标识别技术就是理解并报告摄像头获取得到的图像及其图像序列中出现的感兴趣人,并进一步给出其身份信息。为了实现这些目标,典型的计算机视觉算法实现包括了训练和测试两个阶段。训练阶段,给定图像中的目标数据的特征表达后,通过分类器学习得到在特征空间划分样本分布的超平面。测试阶段,根据待测样本在特征空间中和超平面模型相对位置,实现对于样本类别的预测。这样,典型计算机视觉算法性能好坏取决于数据特征表达是否有足够区分能力,分类器学习是否能够准确估计出样本真实分布。随着智能视频监控系统获取得到的数据越来越多,精确但又复杂的分类器模型已经不能适用,能否学习得到一个鲁棒且具有足够区分能力的特征表达逐渐成为解决现实条件下大规模图像数据分析的关键。本文针对视频监控背景下的目标识别技术中的中层特征表达学习这一问题展开深入而又广泛的研究。具体内容涉及目标分类 多摄像机跟踪步态识别以及多模态行为生物特征融合等方面。在本文中,主要工作和贡献有: 1)针对无监督视觉词典方法,本文研究了目标分类中的快速编码过程以及有监督字典学习方法,在PASCAL VOC数据库上取得了一定的效果。 2)针对当前多视角步态识别中出现的鲁棒性问题,提出了一种基于低秩稀疏编码的视角变换模型,在CASIA数据库上去的了最好的性能。 3)提出了一种基于L1-inf的群稀疏编码方法,在多视角步态识别以及多摄像机跟踪两个数据集上取得了当前最好的效果。 4)针对人体行走产生的地面反应力和步态姿势图像序列,提出了一种基于步态和足印的人体身份识别系统,建立了大规模的算法评测数据库。提出了一种基于CCA的级联特征融合策略。 总的来说,本文结合智能视频监控背景下目标识别技术中的中层特征表达问题进行了深入的研究,在构建中层特征表达上作了一定的探索。
Other AbstractObject Recognition is of fundamentally importance to an intelligent visual surveillance system, and has received great attentions from both academic and industry communities. The final goal of the system is to help computer work as human being, to recognize object of interests. To achieve the goal, typical computer vision algorithm implementation has two stages: training stage and testing stage. During training stage, given the representation of object in image, we employ learned classifer to separate the data using learned hyper plane. During the testing stage, given the position of samples in feature space and the learned hyper plane in feature space, the system need to predict the label of given sample. In that case, the performance of typical computer vision algorithm depends on the discriminative ability of feature representation, and the discriminative ability of learned classifer. As increase in the data obtained by visual surveillance system, it is not practical to learn a complex classifer like nonlinear classifer. To learn the robust and discriminative feature representation is the key to address the large scale image understanding problem. Under the background of intelligent visual surveillance, we proposed to address the issue of learning mid-level feature representation for different tasks in intelligent visual surveillance system including object recognition, multi-view gait recognition, multiple camera tracking and multiple biometric fusion based pedestrian recognition.
shelfnumXWLW1685
Other Identifier200828014628067
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/7609
Collection毕业生_硕士学位论文
Recommended Citation
GB/T 7714
郑帅. 基于中层特征表达的目标识别技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2011.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20082801462806(4526KB) 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
[郑帅]'s Articles
Baidu academic
Similar articles in Baidu academic
[郑帅]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[郑帅]'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.