基于深度学习的行人重识别算法设计.pdf

摘要 行人重识别主要应用在监控视频,是指给定一张特定行人的图像，在不重叠 区域的摄像头中检索出所有该行人图像的过程。传统的行人重识别算法主要基 于人工设计的特征，在精度、性能上有较大局限性。随着深度学习的提出和发 展，现已成功地应用到行人重识别领域。 本文主要就行人重识别的深度学习算法进行研究。首先介绍两种最基本的算 法：基于表征学习和基于度量学习的算法。针对两种算法的不足，介绍了一种 基于局部特征的方法：AlignedReID 。最后通过对AlignedReID 算法引入了随机 擦除的思想提出了改进算法。对以上四种算法进行实验探究，对实验结果进行 分析可以得出，改进的算法在性能和准确度上都有所提升，增强了网络的泛化 性能。 关键词：行人重识别；深度学习；表征学习；度量学习；局部特征 Abstract Pedestrian re-identification is mainly used in surveillance video, which means that given a specific pedestrian image, all the pedestrian images are retrieved in the camera in the non-overlapping area. The traditional pedestrian recognition algorithm is mainly based on the characteristics of artificial design, which has greater limitations in precision and performance. With the introduction and development of deep learning, it has been also applied to the pedestrian recognition field. This paper mainly studies the deep learning algorithm of ReID . First introduce two basic algorithms: characterization learning and metric learning algorithms. A new method called AlignedReID based on local features proposed on account of the defects of two algorithms. Finally, an improved algorithm is proposed by using the idea of random erasing to the AlignedReID algorithm. The above four algorithms are explored experimentally. It can be seen that the algorithm using the random erasing has improved performance and accuracy, which also enhances the generalization performance of the network. Keywords ：pedestrian re-identification ；deep learning ；characterization learning ； metric learning ；local feature 目录 1 引 1 1.1 研究背景与意义 1 1.2 目前研究现状1 1.3 研究内容以及组织结构2 2 行人重识别相关技术4 2.1 卷积神经网络4 2.2 行人重识别技术难点4 2.3 常用数据集和评价指标5 3