【摘要】 行人重识别(re-identification, Re-ID)是计算机视觉领域的研究热点，指在不同监控设备下对同一行人进行再识别。现有行人重识别模型大多独立提取行人不同部位的特征，缺乏对人体部位间内在联系的考量，导致行人在多姿态、视角变换场景下的识别率偏低。图神经网络(graph neural networks, GNN)对图结构数据具备强大的特征建模能力，能充分挖掘图中各节点间的关联信息，在一定程度上有效解决行人遮挡、视角与姿态变化等难题。首先系统梳理了当前行人重识别领域面临的关键问题与技术瓶颈，包括视角变化、遮挡干扰等核心挑战；继而概述了图神经网络的理论架构；在此基础上，再结合国际前沿进展，深入解析了GNN提升行人特征关联性的技术路径，并总结了现有方法的优化策略；最后指出了GNN及行人重识别技术的未来发展方向。更多还原

【Abstract】 Person re-identification(Re-ID) is a research hotspot in the field of computer vision, it is the re-identification of the same person under different surveillance devices. Features from different parts of the person are independently extracted by most existing Re-ID models, the intrinsic connections between human body parts are not taken into account, the recognition rate is relatively low when pedestrians have multi-pose and viewpoint changes. Graph neural network(GNN) possess strong feature modeling capabilities for graph-structured data, the associated information between various nodes in the graph can be fully mined by GNN, thus, problems such as pedestrian occlusion, viewpoint and posture changes can be effectively solved to a certain extent. The key issues and technical bottlenecks in the current Re-ID field were first systematically sorted out. Core challenges including viewpoint changes and occlusion interference were covered. Then, the theoretical framework of GNN was outlined. On this basis, combined with international cutting-edge progress, the technical paths through which GNN enhance the correlation of pedestrian features were deeply analyzed. Meanwhile, the optimization strategies of existing methods were summarized. Finally, the future development directions of GNNs and person Re-ID technology were proposed.更多还原