基于特征融合及差异矩阵的行人再识别  

Person re-identification based on feature fusion and discrepancy matrix

作　　者：李大湘 费国园[1,2] 刘颖 LI Da-xiang;FEI Guo-yuan;LIU Ying(College of Communication and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;Ministry of Public Security Key Laboratory of Electronic Information Application Technology for Scene Investigation,Xi’an University of Posts and Telecommunications,Xi’an 710121,China)

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121 [2]西安邮电大学电子信息现场勘验应用技术公安部重点实验室,陕西西安710121

出　　处：《计算机工程与设计》2020年第4期958-963,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(61571361、61102095);陕西省国际合作交流基金项目(2017KW-013);西安邮电大学研究生创新基金项目(CXJJLY2018040)。

摘　　要：针对行人目标在不同摄像机下外观显著性变化的问题,提出一种基于特征融合及差异矩阵的行人再识别算法。串联融合显著颜色名描述符(SCNCD)和微调的卷积神经网络(FTCNN)特征来描述行人图像,采用K-means算法获取包含典型行人图像的参考集以优化目标与参考身份相对应的重建关系,运用差异矩阵度量(DMMM)算法进行度量学习。在VIPeR和PRID450s行人再识别数据集上的实验结果表明,所提行人再识别算法具有良好的匹配率和有效性。Aiming at the saliency variation of pedestrians’appearance under different cameras,a person re-identification algorithm based on feature fusion and discrepancy matrix was proposed.The salient color names based color descriptor(SCNCD)and fine-tuned CNN(FTCNN)features were tandemly fused to describe pedestrian images.K-means algorithm was adopted to obtain a reference set containing typical images to optimize the reconstruction relationships among the target and the reference identities.Discrepancy matrix and matrix metric(DMMM)algorithm was applied for metric learning.Experimental results on the VIPeR and PRID450s person re-identification datasets show that the proposed algorithm has good matching rates and effectiveness.

关 键 词：行人再识别 特征融合 差异矩阵 卷积神经网络 度量学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]