首页 | 本学科首页   官方微博 | 高级检索  
     

基于UV贴图优化人体特征的行人重识别
引用本文:徐智明,戚湧. 基于UV贴图优化人体特征的行人重识别[J]. 计算机工程, 2022, 48(11): 83-88+95. DOI: 10.19678/j.issn.1000-3428.0063288
作者姓名:徐智明  戚湧
作者单位:南京理工大学 计算机科学与工程学院, 南京 210094
基金项目:国家重点研发计划政府间国际科技创新合作重点专项(2019YFE0123800,2016YFE0108000);欧盟地平线2020科研计划(LC-GV-05-2019)。
摘    要:现有行人重识别研究大多关注人体在二维平面的特征表示,而在现实社会中,人体以一种对称的三维结构存在,三维人体结构相较二维平面像素含有更多的体型、姿态等特征,仅研究二维平面上的行人特征限制了计算机对人体特征的理解。利用人体是一种三维对称的刚体结构这一先验事实,提出一种基于UV贴图优化人体特征的行人重识别方法。对图像底库进行数据预处理,通过数据增广方法生成更多的训练数据,对预处理后的图片进行特征提取,将特征矩阵解耦为姿态、色彩、形状以及视角参数,利用参数信息调整预定义的人体模型以得到重构后的三维人体模型。将重构后的三维人体模型转化为UV贴图,即将人体特征从三维空间映射到二维平面,同时对UV贴图进行优化从而丰富人体特征。使用UV贴图训练三维行人重识别网络模型,利用后处理的方法对输出结果做进一步优化,以得到最终的行人重识别结果。在Market-1501数据集上的实验结果表明,该方法的rank-1准确率和mAP分别达到94.76%和82.53%,相较OG-Net模型分别提升13.82%和22.56%。

关 键 词:计算机视觉  深度学习  行人重识别  三维重构  UV贴图
收稿时间:2021-11-19
修稿时间:2021-12-30

Person Re-Identification Based on UV Map for Optimizing Human Features
XU Zhiming,QI Yong. Person Re-Identification Based on UV Map for Optimizing Human Features[J]. Computer Engineering, 2022, 48(11): 83-88+95. DOI: 10.19678/j.issn.1000-3428.0063288
Authors:XU Zhiming  QI Yong
Affiliation:School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
Abstract:Most previous person re-identification studies focused on the feature representation of the human body in the two-dimensional plane.However, in reality, the human body has a symmetrical three-dimensional structure.The three-dimensional structure of the human body shows more features, such as body shape and posture than two-dimensional plane pixels.Analyzing a person's features in the two-dimensional plane limits the computer's understanding of human characteristics.As the human body has a three-dimensional symmetrical rigid-body structure, a person re-identification method based on UV map for optimizing human features is proposed in this study.Data preprocessing on the image base is performed, additional training data are generated using the data augmentation method, the features of the preprocessed images are extracted, the feature matrix is decoupled into pose, color, shape, and angle of view parameters, and the predefined human model is adjusted using parameter information to establish the reconstructed three-dimensional human model.The reconstructed three-dimensional human body model is transformed into a UV map;that is, human features are mapped from the three-dimensional space to a two-dimensional plane and optimized to enhance human features.The three-dimensional person re-identification network model is trained using UV maps, and the output results are further optimized using post-processing methods to obtain the final person re-identification results.The experimental results show that on the Market-1501 dataset, the rank-1 accuracy and mean Average Precision(mAP) of the proposed method reach 94.76% and 82.53%, respectively.Compared with the OG-Net model, the rank-1 detection accuracy and mAP of person re-identification of this method improve by 13.82% and 22.56%, respectively.
Keywords:computer vision  deep learning  person re-identification  3D reconstruction  UV map  
点击此处可从《计算机工程》浏览原始摘要信息
点击此处可从《计算机工程》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号