基于统一划分的特征自适应行人再识别方法  

Feature adaptive person re-identification method based on unified partition

作　　者：张德磊 宋晓宁[1] 於东军[2] Zhang Delei;Song Xiaoning;Yu Dongjun(School of IoT Engineering,Jiangnan University,Wuxi 214122,China;School of ComputerScience and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《南京理工大学学报》2020年第3期266-271,共6页Journal of Nanjing University of Science and Technology

基　　金：国家重点研发计划(2017YFC1601800);国家自然科学基金(61876072);中国博士后科学基金(2018T110441);江苏省自然科学基金(BK20161135);江苏省“六大人才高峰”项目(XYDXX-012)。

摘　　要：为解决行人再识别(Re-ID)任务中,目标特征统一划分方法导致的部位信息关联性减少和识别率降低的问题,提出了基于密集连接卷积网络(DenseNet)的深度学习方法框架及基于统一划分方法的特征自适应Re-ID方法。在统一划分方法的基础上,为了保留部位特征的相关性,利用马氏距离公式计算相邻特征距离,自适应地选取信息相关性高的部位特征做融合,再对融合后的特征做行人分类。该文算法分别在Market1501数据集、CUHK03数据集以及DukeMTMC-ReID数据集上进行实验。平均精度均值(mAP)分别达到82.8%、70.3%、60.1%。该文方法与基于部位的卷积基准(PCB)以及部位对齐的行人再识别(AlignedReID++)相比,mAP均有提高。In order to solve the problems of reduced relevance and recognition rate of location information caused by unified partition of target features in person re-identification(Re-ID)task,a deep learning framework based on densely connected convolutional networks(DenseNet)and a feature adaptive Re-ID method based on unified partition method are proposed.On the basis of the unified division method,in order to retain the correlation of position features,the Markov distance formula is used to calculate the adjacent feature distance.The position features with high information correlation are selected adaptively for fusion.The fused features are classified into pedestrians.The algorithm proposed here is tested on the Market1501 data set,the CUHK03 data set and the DukeMTMC-ReID data set respectively.The average precision mean(mAP)reach 82.8%,70.3%and 60.1%respectively.Compared with the part-based convolutional baseline(PCB)and part aligned re-identification(AlignedReID++),the average precision of this method is higher.

关 键 词：统一划分 行人再识别 深度学习 马氏距离 平均精度均值 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]