首页|针对大角度下视角差异的行人重识别方法研究

针对大角度下视角差异的行人重识别方法研究

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
行人重识别(Re-ID)也称为行人再识别，旨在给定一个目标行人，确定该行人是否出现在不同的摄像机下，或者是在不同的时间出现在相同的摄像机下。通常由于不同摄像机拍摄到的行人视角不同，在视角差异过大的情况下会对行人重识别准确率造成严重影响。因此，针对目标行人相对摄像机的视角不同而带来的识别率下降问题，提出一种基于外观-步态特征融合的行人重识别算法，使用视角信息对RGB图像与步态能量图(GEI)进行重要性权重估计后再加权融合，以此来克服视角不同而带来的影响。具体来讲，首先利用ResNet-50提取图像序列中每张图像的特征，采用时间池化的方式将其聚合为外观特征。其次使用另一个ResNet-50对GEI图像提取步态特征。然后对行人进行视角估计之后，映射函数将估计的角度映射为两种特征的重要性权重。最后基于自编码器结构将两种特征在重要性权重的指导下进行加权融合，生成对视角鲁棒的融合特征。在CASIA-B数据集上的实验结果表明，对于具有大角度视角差异的行人Re-ID，所提出方法在mAP和Rank-1评估指标上都表现出了显著的改进。在大角度差异情况下进行测试，准确率最高提升了 2。7％。
Research on Person Re-Identification Method for Large-Angle Viewpoint Differences
Person Re-Identification(Re-ID)aims to determine whether a person of interest appeared under different cameras or under the same camera at different times.Owing to the different viewpoints of people captured by different cameras,the accuracy of person Re-ID can be adversely affected.Therefore,in this study,a person Re-ID method based on the fusion of appearance and gait features is proposed to solve the reduced recognition rate caused by different viewpoints of people relative to a camera.Here,viewpoint information is utilized to estimate the importance weights of an RGB image and Gait Energy Image(GEI);subsequently,weighted fusion is performed to overcome the effects of different viewpoints.Specifically,ResNet-50 is first used to extract the features of each image in the image sequence,which are then further aggregated into appearance features via temporal pooling.Second,another ResNet-50 model is used to extract gait features from the GEI images.Third,after estimating the viewpoint of a person,a mapping function is designed to map the viewpoints to the importance weights of the two features.Finally,based on the auto-encoder structure,the two features are weighted and fused under the guidance of importance weights to generate fusion features that are robust to the perspective.Experimental results on the CASIA-B dataset show that the proposed method exhibits significant improvements in terms of the mAP and Rank-1 evaluation metrics for person Re-ID with large-angle viewpoint differences.When tested under large-angle differences,the highest accuracy improvement is 2.7％.

person Re-Identification(Re-ID)viewpoint differencesappearance featuregait featurefeature fusion

蔡毅翔、秦品乐、曾建潮、晋赞霞、秦佳、翟双姣

展开 >

中北大学计算机科学与技术学院,山西太原 030051

行人重识别 视角差异 外观特征 步态特征 特征融合

山西省科技重大专项计划"揭榜挂帅"项目山西省基础研究计划自由探索类青年科学研究项目山西省留学回国人员科技活动择优项目

202110101010101820220302122204920230017

2024

10.19678/j.issn.1000-3428.0067532

计算机工程
华东计算技术研究所　上海市计算机学会

计算机工程

CSTPCD北大核心
影响因子：0.581
ISSN：1000-3428
年,卷(期)：2024.50(5)
  • 39