基于场景迁移与区域对齐的行人再识别

Pedestrian re-identification based on scenes transition and region alignment

针对目前现有行人再识别方法训练样本不足的问题,利用语义分割方法将样本图像中的行人区域与背景分离,对背景区域使用生成式对抗网络(GAN)完成图像场景迁移,在保留行人特征的前提下对现有数据集进行扩充。针对数据集中行人区域未对准的情况,提出基于语义分割的滑动窗口行人对准方法,并根据数据集的扩充和对准在残差卷积神经网络结构ResNet-50中加入全局特征分支。实验中使用公共数据集Market-1501和DukeMTMC-reID对上述方法进行测试,在Rank-1指标上分别取得了91.4%和81.1%的准确率。

To solve the problem of insufficient training samples of existing pedestrian re-identification methods,semantic segmentation is applied to separate the pedestrian area from the background in the sample image.and the generative adversarial network(GAN)is used to generate the different backgrounds for image scene transition.This mechanism not only expand the existing datasets,but also retain the pedestrian characteristics.The sliding window alignment(SWA)method based on semantic segmentation is proposed to solve the problem of pedestrian area misalignment in datasets.At the same time,the global feature branch is added to the residual convolutional neural network structure ResNet-50.In the experiment,this method is tested on Market-1501 and DukeMTMC-reID datasets,and achieve accuracy of 91.4%and 81.1%in rank-1 index.