A Study on Traditional and Modern Approaches for Person Recognition

CASIA OpenIR > 毕业生 > 博士学位论文

	A Study on Traditional and Modern Approaches for Person Recognition
	Muhammad Rauf1,2
	2017-04
学位类型	工学博士
中文摘要	Human recognition and identication are desirable yet challenging for many applications including surveillance, computer vision, and robotics. Many biomet- rics have been used in human recognition and identication applications. Gait biometrics is used to recognize and identify people from videos. The gait biomet- rics has its advantage of capabilities for recognizing people from distance. It is a very hot topic and has attracted lot of attention in recent years as we deployed hundreds of thousand surveillance cameras. These surveillance cameras can be used for analyzing data on the edge [1]. To deploy the deep learning based pro- cess on the smart devices, process need to be small in size and able to use low computing ability. This ability can be archived with the optimized learning tech- nique. Despite having a complex structure and limitations, the discrimination power of the gait make its a useful and unique component for person recognition at a distance. The work presented in this thesis aims to nd the new strategy for optimized machine learning and gait recognition and identication systems. The object of the work can be described as: Find the new strategy for gait recognition and identication process. Investigate the possible solution for deploying the machine learning tech- nique on the low processing power devices. In order to complete our goal we study dierent strategies to optimize machine iv A Study on Traditional and Modern Approaches for Person Recognition learning and human identication. This involved traditional bag-of-the-word framework and deep learning techniques. The goals that we achieved can be described as: Optimization of encoding methods used in Bag-of-Word framework: In this work, our goal is to draw a relationship between descriptors, and use this relationship to optimize the encoding methods. We used neighboring descriptors relationship during the encoding the words. We implemented this new optimization technique on dierent encoding methods and tested them on image classication and gait recognition. Retrieving data from videos on the base of their gait biometric: In order to address the challenges of large-scale video databases, we introduce new gait based retrieval. The gait based retrieval is used to extract the data of target person from the video database. We used deep hashing method to solve the retrieval problem, which outperforms the traditional methods that we implemented. On next stage this retrieved data can be used for further recognition and identication tasks. Reducing the computational cost by knowledge transfer technique: With the increasing use of smart devices and IP cameras, video analysis can be performed on the edge. Deep learning models are complex in size and costly in computation, it is not possible to use these models directly on smart devices and for edge based analysis. We use knowledge transfer technique to reduce processing and memory cost. We implemented this technique to transfer the functional knowledge of previous trained models to small Chapter 0. Abstract v models. Then these small models can be deployed on smart devices and for the analysis of data at the end terminal.
英文摘要	行人识别与鉴别在很多应用领域非常急需且具有挑战性，包括视频监控、计算机视觉、机器人等。一些生物特征技术己经应用于行人识别与鉴别中，其中步态生物特征常用于在视频中找出并鉴别行人身份。步态生物特征具有远距离识别行人的优势。近年来，由于大规模监控摄像头的部署，步态识别技术成为一个热门的研究问题并吸引了大量关注。这些监控摄像头可以在远程终端进行数据分析。为了便于在小型智能终端设备上使用深度学习技术来处理这些数据，就要求处理的过程要轻量化且能够使用较低的计算能力。这种能力可以通过优化学习技术来实现。虽然步态具有复杂的结构和局限性，但是其判别能力使其成为在远距离行人身份识别上是有效且唯一的方法，本文提出的工作致力于发现一种新颖方法用于机器学习和步态识别与鉴别系统。本文的目标总结如下： l 发现一种新的步态识别与鉴别处理的策略； l 研究将机器学习技术应用于低功耗设备的可能解决方案。为了完成这些目标，本文研究了一系列不同的策略以优化机器学习与步态识别系统。这其中包含了传统的词袋框架（Bag-of-Word,BOW）以及深度学习技术。我们实现的工作总结如下： l 优化Bag-of-Word,BOW词袋框架的编码方法：在本文中，我们的目的是得到描述子之间的关系，并且利用这种关系来优化编码方法。在编码单词时, 我们使用了相邻的描述子关系。我们将这种新的优化技术应用在不同的编码方法中，并在图像分类与步态识别的任务上进行实验。 l 基于步态生物特征在视频中进行数据检索：为了应对大规模视频数据库带来的挑战，我们提出一种新的基于步态的检索方法。这种基于步态的检索方法可用于在视频数据库中提取目标行人的数据。我们采用了深度哈希方法来解决检索问题，这种方法结果好于传统方法。在下一阶段这些检索到的数据可以进一步用于识别与鉴别的任务。 l 利用知识转换技术降低计算消耗：随着小型化智能设备和网络摄像头日益增长的使用，视频分析得以在远程终端设备直接进行。然而深度学习训练的模型通常占用较大空间且造成大量的计算消耗，因此难以直接将这些模型用在小型智能设备以及远程终端分析中。我们采用知识转换技术来减少处理过程以及降低内存消耗。我们利用这种技术将之前训练好的模型的函数化的知识转换为小型模型。然后这些小型模型可以在智能设备上运行并且可以用于远程终端的数据分析中。
关键词	近邻描述子步态检索卷积神经网络步态识别知识转移
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/15509
专题	毕业生_博士学位论文
作者单位	1.中国科学院自动化研究所 2.中国科学院大学
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	Muhammad Rauf. A Study on Traditional and Modern Approaches for Person Recognition[D]. 中国科学院大学. 中国科学院大学,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
行人身份识别的经典与现代方法研究 Muh（11367KB）	学位论文		限制开放	CC BY-NC-SA