The Research On Generalization Method For Open-World Person Re-Identification In Aerial Imagery

Open-world person re-identification(Re-ID)aims to determine whether a given pedestrian appears in another camera scene through nonoverlapping video surveillance networks in an unknown spatial environment.The gallery may not contain the target pedestrian,making it a more challenging and practical application research than close-world person re-identification.Most of the existing research on Open-world person Re-ID is based on ground pedestrian imagery.In recent years,with the rapid development of unmanned aerial vehicle(UAV)technology,the visual monitoring system built by the UAV platform has received a lot of attention.The open-world person Re-ID based on the aerial images taken by UAV has become a popular research topic.Due to the flexibility of the UAV,the pedestrian images captured by the UAV are quite different from those captured by the fixed ground camera.Therefore,the person Re-ID task of aerial imagery is not only faced with the problems of illumination and clothing changes in the traditional Re-ID task,but also needs to deal with the serious occlusion,viewpoint and resolution difference caused by the interference factors such as inconsistent flying height and change of shooting angle of the UAV.Based on the deep learning technology,combined with the idea of integrity,this paper studies how to improve the generalization ability of open-world person Re-ID models under aerial photography.The main research contents and innovations include:1.A generalization method based on meta-transfer learning for person Re-ID in aerial imagery.Aiming at the problem that the pedestrian features constructed from aerial images are weak due to serious occlusion,perspective and resolution difference,a person Re-ID framework based on meta-transfer learning is proposed.The learning framework combines the advantages of meta-learning and transfer learning,and trains a feature extraction model with high generalization for aerial pedestrian images based on the idea of "learning to learn" and transfer learning to explore the relevance of knowledge.In addition,the framework introduces a training strategy based on curriculum learning to improve the utilization of hard samples in aerial images,and designs a measurement strategy based on Gaussian embedding to improve the model optimization direction,making the trained model more generalizable.2.Research on a generalization model for person Re-ID based on vision Transformer(Vi T)in aerial imagery.Aiming at the problem of weak generalization ability of models trained by existing deep learning methods due to the presence of multiple scales and grain of targets in aerial person imagery,a person Re-ID model based on Vi T is proposed.Based on the powerful Transformer mechanism,the model focuses on both global and local information of pedestrian images through multi-head self-attention modules.In addition,a feature fusion strategy based on view information and a feature augmentation strategy based on zero-padding and shifting are introduced to address the variable views and severe occlusion in aerial pedestrian images.Aiming at the sensitiveness to model settings such as optimizer and hyperparameters during Vi T training,a model structure optimization strategy based on standard convolution is introduced,which further improved the generalization ability of the model.To verify the feasibility of the method in this paper,several groups of experiments were carried out on a large-scale aerial person dataset PRAI-1581,and supplementary experiments were carried out on two commonly used ground person datasets Market-1501 and Duke MTMC-re ID.The experimental results show that the person Re-ID framework based on metatransfer learning and the person Re-ID model based on vision Transformer proposed in this paper can effectively improve the generalization ability of the person Re-ID model in aerial imagery.