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

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.