Abstract: Abstract: visible-infrared cross-modal person re-id problem means recognizing and retrieving person in mixed day and night scenarios that means some images are taken in visible light while the others are in infrared. A cross-modal person re-identification relation network named as IVRNBDS (Infrared and visible relation Network based on dual-stream structure) is proposed. The network has two branches to extract the features of the visible light modal pedestrian image and the infrared modal pedestrian image respectively. And then the pedestrian image is divided into six segments horizontally to extract relationships between each segment and the other segments of the pedestrian, and the relationship between the core feature and average feature of the pedestrian is also extracted in the relation model. In addition, a triplet loss function based on heterogeneous centers is introduced to relax the strict constraints of the ordinary triplet loss function, so that images of different modalities can be better mapped into the same feature space. Experiments on public dataset indicate a superior recognition effect for the IVRNBDS network.

Key words: Keywords: person re-identification, visible-infrared cross-modal, dual-stream structure, Hetero-center triplet loss, relation model

摘要： 摘 要: 针对可见光-红外跨模态行人重识别模态差异导致的识别准确率不高的问题，提出了一种基于双流结构的跨模态行人重识别关系网络IVRNBDS。该网络利用双流结构分别提取可见光模态和红外模态行人图像的特征，然后将行人身体分为6个片段，提取行人的每个片段与其他片段之间的关系，以及行人的核心特征信息和平均特征之间的关系，设计形成关系抽取模块。在损失函数设计时，引入异质中心三元组损失函数，放松普通三元组损失函数的严格约束，使不同模态的图像可以更好地映射到同一特征空间中。实验表明IVRNBDS模型具有较好的识别效果。

关键词: 行人重识别, 跨模态, 双流结构, 异质中心三元组损失函数, 关系模块