域通用和域分离字典对学习的行人重识别算法

Domain-Common and Domain-Separation Dictionary Pair Learning for Person Re-Identification

为克服不同相机视角之间的域偏移问题,提出一种基于域通用和域分离字典对学习的跨视角行人重识别算法。具体地,基于来自同一相机视角下的行人共享相同的域,并且同一视角中每个行人图像所携带的域信息在短时间内具有一致性,将同一视角下的行人图像分解为特定视角的域信息分量和域分离的行人外观特征分量,提出一个判别字典学习模型以创建用于描述域信息分量的域通用字典和描述行人外观分量的域分离字典。由于来自同一相机视角下的图像具有域相似性,因此通过低秩正则化来细化用于表示域信息的字典。为了进一步提高学习字典的判别能力,在算法中约束相同视角、相同身份的多幅图像的编码系数具有很强的相似性。此外,采用一种新颖的扩展正则化方法来解决不同行人相似外貌特征和同一行人不同外貌特征的视觉外观歧义问题。在四个具有挑战性的数据集上进行实验,结果表明域通用和域分离字典对学习的算法相对于一些现有最新算法更具有效性和优越性。

In order to overcome the problem of domain-shift between different camera views,a domain-common and domain-separation dictionary pair learning approach for cross-view person re-identification is proposed.Specifically,based on the fact that pedestrians from the same camera view share the same domain,and that the information of domain carried by each pedestrian image in the same view has consistency for short time,the person image from the same view is decomposed into specific perspective domain information components and domain-separation pedestrian appearance feature components,and a discriminative dictionary learning model is developed to create a domain-common dictionary for characterizing the domain information components and a domain-separation dictionary for characterizing the pedestrian appearance components.Since images from the same camera view have domain similarity,dictionary used to represent domain information is refined by low-rank regularization.Besides,to further improve discrimination ability of learned dictionaries,certain constraint is proposed in the algorithm,that is,the coding coefficients of multiple images with the same view and identity to have strong similarity.Furthermore,a novel expansion regularization is used to solve the visual appearance ambiguity of similar appearance features of different pedestrians and different appearance features of the same pedestrian.Finally,experiments on four challenging datasets demonstrate the effectiveness and the superiority of domain common and domain-separation dictionary pair learning approach against some state-of-the-art methods.