基于全局特征改进的行人重识别

Improved Person Re-identification Based on Global Feature

由于行人重识别面临姿态变化、遮挡干扰、光照差异等挑战,因此提取判别力强的行人特征至关重要.本文提出一种在全局特征基础上进行改进的行人重识别方法,首先,设计多重感受野融合模块充分获取行人上下文信息,提升全局特征辨别力;其次,采用GeM池化获取细粒度特征;最后,构建多分支网络,融合网络不同深度的特征预测行人身份.本文方法在Market1501和DukeMTMC-ReID两大数据集上的mAP指标分别达到83.8%和74.9%.实验结果表明,本文方法有效改进了基于全局特征的模型,提升了行人重识别的识别准确率.

Person re-identification faces challenges such as posture change,occlusion interference,and illumination difference,and thus it is very important to extract pedestrian features with strong discriminability.In this paper,an improved person re-identification method based on global features is proposed.Firstly,a multi-receptive field fusion module is designed to fully obtain pedestrian context information and improve the global feature discriminability.Secondly,generalized mean(GeM)pooling is used to obtain fine-grained features.Finally,a multi-branch network is constructed,and the features of different depths of the network are fused to predict the identity of pedestrians.The mAP indexes of this method on Market1501 and DukeMTMC-ReID are 83.8%and 74.9%,respectively.The experimental results show that the proposed method can effectively improve the model based on global features and raise the recognition accuracy of person re-identification.