【作者】 郑易；

【导师】 齐冬莲； 李超勇； 张建良；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2022， 硕士

【摘要】 行人重识别算法可进行跨摄像头的行人检索,在变电站智能视频监控、智能安防等领域具有广阔的应用前景。本文从理论研究与实际工程应用两方面着手,研究了一种基于在线层次聚类的无监督行人重识别方法,并将其应用于变电站人员重识别中。本文取得的主要研究成果如下:(1)提出了一种基于在线聚类的模型同步训练范式。根据特征空间局部范围内样本分布信息,每一次训练迭代后进行训练样本标签调整,实现了模型参数与伪标签同步更新,有效地解决了信息滞后的问题。通过实验对比了不同伪标签更新速率的效果,验证了使用在线聚类进行伪标签生成的有效性。(2)提出一种自上而下的多层次动态聚类方法。该方法包含簇拆分与簇聚合两个子模块,两者均基于标签传播算法进行调整与更新。簇拆分模块用于拆分伪标签簇,有利于提升伪标签的精确率。簇聚合模块用于聚合相邻伪标签簇,可提升伪标签的召回率。实验结果表明,多层次动态聚类方法可以更好地对行人的多层次语义信息进行捕捉,从而有效提升模型性能。(3)基于公开数据集与变电站人员数据集,对所提算法的有效性与鲁棒性进行了验证。本文算法在Market-1501、Duke MTMC-Re ID与MSMT17公开数据集取得了超越现有无监督行人识别算法的精度。此外,本文基于多目标追踪算法构建变电站人员数据集,并针对实际场景特点在算法上增加帧间平均融合模块,通过实验验证了所提算法的性能。更多还原

【Abstract】 Person re-identification aims at retrieving the same person’s images across different cameras.It has broad application prospects in both substation surveillance and intelligent security.This paper conducts research from both theoretical and practical engineering applications.For one thing,an online hierarchical clustering algorithm is proposed in order to tackle the remained challenges in person re-identification.For another,the proposed method is applied to the actual scene of the substation and realize the re-identification of substation personnel.The contributions of this paper are summarized as follows:(1)A synchronization training paradigm based on online clustering is proposed,in which feature learning and pseudo label generation proceed simultaneously.This structure solves the problem of lagged update information.The results with different pseudo label generation rates are compared through ablation study,and the effectiveness of the synchronization training paradigm is verified through extensive experiments.(2)A cluster dynamics method is proposed to enable pseudo label progression and refinement iteratively in a bottom-to-up hierarchical structure.It consists of two submodules: cluster split module and cluster merge module,both of which are updated based on label propagation algorithm.Cluster split module aims at splitting pseudo label cluster,which is beneficial to improve the precision of clusters.Cluster merge module is used to merge adjacent pseudo label clusters,which can increase the recall of pseudo labels.Extensive experiments demonstrate that hierarchical clustering method can better capture multi-level semantic information,thereby effectively improving the performance of model.(3)The effectiveness of the proposed algorithm is verified based on three public datasets and the real scene dataset of the substation.The method in this paper surpasses all current state-of-the-arts on Market-1501,Duke MTMC-Re ID and MSMT17.In addition,a dataset of substation personnel is constructed based on Tracktor algorithm.Aiming at the characteristics of the actual scene,this paper adds the temporal average fusion in the proposed algorithm and the performance is verified through experiments.更多还原