基于度量学习的行人重识别综述

Review of Pedestrian Re-identification Based on Metric Learning

行人重识别的主要目标是在多个摄像机拍摄的图片或视频中识别同一行人,在智能安防等领域的应用前景广阔,是近年来计算机视觉方向的热门研究课题之一。随着深度学习的高速发展以及行人数据集的多样化,行人重识别的研究取得了显著进展,涉及的技术主要包括两个部分:特征提取和度量学习。已有的行人重识别研究更加关注特征提取方面,对于度量学习的系统论述不多。有效的度量学习对于提高行人重识别的准确性至关重要,故对基于度量学习的行人重识别技术进行梳理与分析具有重要价值。本文对近年基于度量学习的行人重识别方法进行总结,主要归纳为两部分:度量方法与度量学习算法。其中度量方法可分为距离度量与基于超图的相似性度量,将两种方法在行人重识别公开数据集上进行性能对比;度量学习算法总结为经典度量学习算法与深度度量学习算法,对深度度量学习算法中的损失函数进行性能总结与对比。最后,分析了基于度量学习的行人重识别中存在的问题及发展方向。

The main objective of pedestrian re-identification is to identify the same pedestri-an in pictures or videos taken by multiple cameras,holding great promise for various fields,including intelligent video surveillance and intelligent security,and it is one of the popular areas of research in the field of computer vision in recent years.With the rapid development of deep learning and the diversification of pedestrian datasets,significant progress has been made in the research on pedestrian re-identification,and the technologies involved mainly in-clude tw o parts:feature extraction and metric learning.Existing literature on pedestrian re-i-dentification focuses on feature extraction,and has less systematic discourse of metric learn-ing.How ever,an effective distance metric is essential in enhancing the accuracy of pedestrian re-identification.Therefore,sorting out and analyzing pedestrian re-identification based on metric learning is valuable.This paper intends to summarize the pedestrian re-identification methods based on metric learning in recent years,mainly including metric methods and met-ric learning algorithms.The metric methods can be divided into distance metric and hyperg-raph-based similarity metric,and the performance of the tw o methods on public datasets of pedestrian re-identification can be compared.Metric learning algorithms are summarized into classical metric learning algorithms and deep metric learning algorithms,and the perform-ance of the loss function in the deep metric learning algorithm is summarized and compared.Finally,the remaining challenges and prospect of pedestrian re-identification based on metric learning are discussed.