面向行人重识别的局部特征研究进展、挑战与展望

Research Progress,Challenge and Prospect of Local Features for Person Re-Identification

行人重识别(Person re-identification,Re-ID)旨在跨区域、跨场景的视频中实现行人的检索及跟踪,其成果在智能监控、刑事侦查、反恐防暴等领域具有广阔的应用前景.由于真实场景下的行人图像存在光照差异大、拍摄视角不统一、物体遮挡等问题,导致从图像整体提取的全局特征易受无关因素的干扰,识别精度不高.基于局部特征的方法通过挖掘行人姿态、人体部位、视角特征等关键信息,可加强模型对人体关键区域的学习,降低无关因素的干扰,从而克服全局特征的缺陷,也因此成为近几年的研究热点.本文对近年基于局部特征的行人重识别文献进行梳理,简述了行人重识别的发展历程,将基于局部特征的方法归纳为基于姿势提取、基于特征空间分割、基于视角信息、基于注意力机制四类,并详细阐述了每一类的原理及优缺点.然后在三个主流行人数据集上对典型方法的识别性能进行了分析比较,最后总结了目前基于局部特征算法的难点,并对未来本领域的研究趋势和发展方向进行展望.

Person re-identification(Re-ID)aims to achieve pedestrian retrieval and tracking in cross-region and cross-scene video.Its achievements have broad application prospects in intelligent monitoring,criminal investigation,counter-terrorism and riot control.Due to pedestrian images in real scenes having problems such as large illumination differences,different shooting angles,and object occlusion,the global feature is susceptible to interference from irrelevant factors,resulting in low recognition accuracy.The local feature-based method strengthens the model’s learning of key areas of the human body and reduces the interference of irrelevant factors by mining key information such as pedestrian posture,human body parts,and perspective features.Because the local feature method overcomes the defect of the global feature,it has become a research focus in recent years.In this paper,we combed the literature of Re-ID based on local features in recent years,and briefly described the development process of Re-ID.The methods based on local features can be classified into four categories:postural extraction,feature spatial partition,viewpoint information and attention mechanism.This paper first elaborates on the principles,advantages and disadvantages of each category.Then we summarize some typical methods in detail and compare their performance on three mainstream Re-ID data sets.Finally,this paper summarizes the difficulties of the method based on local features,and looks forward to the future research trend and development direction of this field.