基于度量学习的行人再识别研究

发布时间：2018-05-21 20:07

本文选题：行人再识别 + 度量学习　；参考：《电子科技大学》2017年硕士论文

【摘要】：随着计算机视觉和模式识别的发展,行人再识别已成为防止潜在暴力事件发生的有力工具。行人再识别是在非重叠视域中匹配同一行人目标的过程。由于从不同视角采集的行人图像分辨率低,存在光照、姿态、背景变化及遮挡等问题,所以行人再识别一直是一项挑战性的课题。为了克服这些问题,行人再识别技术分别从两个不同的方面着手:提取鲁棒性的行人特征和学习合适的距离度量。在本论文中,我更多地关注后者。针对行人再识别问题,本论文的主要工作如下:1.对行人再识别技术进行概述。首先对行人再识别技术的背景、意义和发展历史进行了简单阐述。然后根据行人再识别技术的侧重点不同分别从特征提取和度量学习两个方面阐述了现有的行人再识别方法。2.研究基于核度量学习的行人再识别方法。核方法最大的优势就是在不知道具体的非线性映射函数的形式下,就可以将原始空间的数据向高维空间投影来提高分类能力。本文基于大间隔最近邻(LMNN)、局部费舍判别分析(LFDA)和Null Foley Sammon变换(NFST)提出了三种核度量学习方法。在三个具有挑战性的行人再识别数据库上的实验结果验证了核度量学习方法的有效性。3.提出基于非对称几何度量学习的行人再识别方法。它从几何关系的角度针对每个特定的视角学习投影变换来提高对称度量学习方法在行人再识别上的性能。对称的度量学习方法对所有的视角学习单一的投影变换,然而这往往忽略了不同视角之间的差异性。基于非对称几何度量学习的方法可以解决对称度量学习在行人再识别上的上述问题。在三个挑战性的行人再识别数据库上的识别性能,证明了其在行人再识别问题上的有效性。
[Abstract]:With the development of computer vision and pattern recognition, pedestrian recognition has become a powerful tool to prevent potential violence. Pedestrian recognition is a process of matching the same pedestrian target in a non-overlapping horizon. Due to the low resolution of pedestrian images collected from different angles of view, there are problems such as illumination, posture, background change and occlusion, so pedestrian recognition is always a challenging task. In order to overcome these problems, the pedestrian recognition technique starts from two different aspects: extracting robust pedestrian features and learning appropriate distance metrics. In this paper, I pay more attention to the latter. The main work of this thesis is as follows: 1. The technology of pedestrian recognition is summarized. Firstly, the background, significance and development history of pedestrian recognition technology are briefly described. Then, according to the different emphases of pedestrian rerecognition technology, this paper expounds the existing pedestrian rerecognition methods from two aspects: feature extraction and measurement learning. A method of pedestrian rerecognition based on kernel metric learning is studied. The biggest advantage of the kernel method is that it can project the data of the original space to the high-dimensional space without knowing the specific nonlinear mapping function to improve the classification ability. In this paper, three kernel metric learning methods are proposed based on large interval nearest neighbor LMNNs, local Fisher discriminant analysis (LFDA) and Null Foley Sammon transform. Experimental results on three challenging pedestrian recognition databases verify the effectiveness of the kernel metric learning method. A method of pedestrian rerecognition based on asymmetric geometric metric learning is proposed. In order to improve the performance of the symmetric metric learning method in pedestrian recognition, the projection transformation is studied from the angle of geometric relation for each particular angle of view. Symmetric metric learning methods learn a single projection transformation for all visual angles, but this often ignores the differences between different perspectives. The method based on asymmetric geometric metric learning can solve the above problem of pedestrian rerecognition in symmetric metric learning. The recognition performance on three challenging pedestrian rerecognition databases proves its effectiveness in pedestrian rerecognition.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】