摘要： 行人搜索旨在全景图像中对目标行人进行定位和识别，可以看作目标检测和行人重识别任务的结合。然而，不同行人的着装相似性和同一行人在多变环境条件下的外观差异，增加了行人身份辨别的难度。为了解决这一问题，提出了一个原型分散网络，通过调整原型的分布情况，增强网络的辨别能力。首先，设计了一个原型增强模块，通过原型特征指导注意力网络的学习并利用K最大池化方法保留更多重要的行人特征，借助原型的指导使网络关注更重要的区域，学习更细粒度的行人特征，进而提高网络对相似行人的辨别能力。其次，提出一种自适应更新的原型学习策略，保证在原型特征更新时，检测精准的候选框做出更大的贡献。最后，通过分布稀疏损失保证所存储的原型尽可能分散，从而确保网络能识别到行人的可区分性特征。最终在公共的行人搜索数据集CUHK-SYSU和PRW上进行了实验，在mAP上分别达到了95.1%和49.8%，在top1准确率上分别达到了95.9%和88.5%，该方法有效地提高了行人搜索的准确率，证明了本文方法的有效性。

Abstract: Person search aims to localize and identify target pedestrians in panoramic images and can be seen as a combination of target detection and person re-identification tasks. However, the similarity in attire among different pedestrians and the appearance variations of the same person in diverse environmental conditions contribute to the increased difficulty of pedestrian identity recognition. To solve this problem, a prototype separation network is proposed to enhance the network's discrimination ability by adjusting the distribution of prototypes. Firstly, a prototype enhancement module is designed to guide the learning of the attention network through prototype features and retain more important pedestrian features using the K-max pooling method, so that the network focuses on important local regions and learns finer-grained pedestrian features with the help of the guidance of the prototypes, which in turn improves the network's discriminative ability of similar pedestrians. Second, an adaptive updating prototype learning strategy is proposed to ensure that the detection of accurate candidate boxes makes a greater contribution when the prototype features are updated. Finally, the distributed sparse loss ensures that the stored prototypes are dispersed as much as possible, thus ensuring that the network can recognize the distinguishable features of pedestrians. Finally, relevant experiments are conducted on the public pedestrian search datasets CUHK-SYSU and PRW, which reach 95.1% and 49.8% on mAP and 95.9% and 88.5% on top1, respectively. The results show that the proposed method effectively improves the accuracy of person search, which proves the effectiveness of the proposed method.