大数据环境下深度学习行人再识别技术研究与应用

Research and Application of Pedestrian Re-identification Technology Based on Deep Learning in Big Data Environment

行人再识别(Re-ID)是网络匹配行人图像任务。它与属性识别在学习行人描述上有共同目标,而区别是粒度。通过基于属性标签和ID标签的互补性,多数Re-ID方法仅考虑行人身份标签,而包含详细本地描述的属性有利于Re-ID模型学习更多判别式特征表示。因此提出属性的人识别(APR)网络。APR网络是通过学习Re-ID嵌入并同时预测行人属性的多任务网络。在实验中考虑属性间的依赖性和相关性后对属性预测加权,则APR检索过程快了10倍,在Market-1501上准确度下降了2.92%。

Pedestrian Re-ID(Re-ID)is the task of matching pedestrian images with the network.Based on the complementarity of attribute tags and ID tags,most Re-ID methods only consider pedestrian identity tags,and the attributes that include detailed local descriptions help the Re-ID model to learn more discriminative feature representations.Therefore,an attributed person recognition(APR)network is proposed.The APR network is a multi-task network that learns Re-ID embedding and predicts pedestrian attributes at the same time.In the experiment,after considering the dependence and correlation between attributes and weighting the attribute prediction,the APR retrieval process is ten times faster,and the accuracy on Market-1501 drops by 2.92%.