Abstract Discriminative feature learning is critical for pedestrian re-identification. Previous part-based methods mainly focus on the region of specific predefined semantics to learn local representations, ignoring the influence of posture changes, and the learning efficiency and robustness in complex scenes are poor. In this paper, a hypergraph video pedestrian re-identification method based on posture structure relationships and action constraint(PA-HVPReid) is proposed, which aims to make full use of pedestrian walking postures to obtain more discriminative features. The pose structure relationship feature solves the problem that the pooling operation destroys the feature structure relationship. This paper uses a graph convolution network (GCN) to preserve the structure relationship presented by the image feature map. The input of the GCN is the region at the joint points of the pedestrian to be detected, and the output is the color feature that retains the structural relationship. The structural relationship hypergraph is formed according to the structural relationship between the joint point regions. The action hypergraph can be constructed by constraining the action information between the joint point regions. The saliency score of the joint point region is calculated from the posture structure hypergraph and the action hypergraph. We convert the saliency score into a probability distribution problem and propose a relative entropy loss function based on regional saliency to measure the similarity of the two probability distributions. Experimental results show that the performance of our method is better than the existing method on three data sets.

中文翻译：

---

基于姿态结构关系和动作约束的超图视频行人重识别

摘要 判别特征学习对于行人重新识别至关重要。以往的part-based方法主要集中在特定预定义语义的区域学习局部表征，忽略了姿势变化的影响，复杂场景下的学习效率和鲁棒性较差。本文提出了一种基于姿态结构关系和动作约束的超图视频行人重识别方法（PA-HVPreid），旨在充分利用行人的行走姿态来获得更具判别力的特征。位姿结构关系特征解决了池化操作破坏特征结构关系的问题。本文使用图卷积网络（GCN）来保留图像特征图呈现的结构关系。GCN的输入是待检测行人关节点处的区域，输出是保留结构关系的颜色特征。根据关节点区域之间的结构关系形成结构关系超图。可以通过约束关节点区域之间的动作信息来构建动作超图。根据姿势结构超图和动作超图计算关节点区域的显着性得分。我们将显着性分数转换为概率分布问题，并提出了基于区域显着性的相对熵损失函数来衡量两个概率分布的相似性。实验结果表明，我们的方法在三个数据集上的性能优于现有方法。