结合时序和质量评估的视频行人重识别

Video-based person re-identification combined with sequential and quality estimation

针对行人重识别中出现的行人遮挡、图像质量参差不齐的情况,提出一种结合时序和局域质量评估的网络模型。从时序信息和质量评估两方面出发,利用时序信息和注意力机制弥补空间上出现的噪音或者缺失,进一步提高基础特征表达的判别能力;将行人图像按比例分割为3段,分别进行局域质量评估;将得到的分数作为权重与行人特征加权求和得到最终的特征。将ResNet-50作为主干网络,在MARS、DukeMTMC-VideoReID和PRID-2011数据集上的实验结果表明,该网络模型在行人重识别准确率上有一定提升。

Aiming at the pedestrian occlusion and uneven image quality in person re-identification,a network model combining temporal information and the region-based quality estimation was proposed.From the two aspects of temporal information and quality evaluation,temporal information was used to make up for the noise or lack of space,to further improve the discrimination ability of basic feature expression.The pedestrian image was divided into three segments according to the proportion,and the quality was evaluated respectively.The final feature was obtained by weighted summation of the obtained score as weight and pedestrian feature.Taking Resnet-50 as the backbone network,through experiments on MARS,DukeMTMC-VideoReID and PRID-2011 data sets,experimental results show that the network model has a certain improvement in the accuracy of person re-identification.