Research On Cross-Modality Person Re-Identification Method Based On Deep Feature Interaction

In recent years,the position of intelligent security in urban planning and construction has become more and more important,and it shows broad application prospects in the fields of campus security,public security criminal investigation,and traffic safety.In such a background,person reidentification technology has drawn the attention from numerous scholars.The single-modality person re-identification has not been able to meet the application needs.Thanks to the deployment of infrared cameras in conjunction with visible cameras,a 24/7 surveillance system has been established.The purpose of cross-modality person re-identification is to match visible images and infrared images with the same identity in different scenes.Many excellent works have been conducted,but they still have shortcomings:(1)Some discriminative features are easily lost when extracting features,and the information between different modalities cannot be effectively fused.(2)The cost of data labeling is high,and a small amount of labeled data cannot meet the training requirements.(3)The person is occluded by other objects while walking,and some discriminative information is lost,which leads to the degradation of recognition performance.Therefore,the following solutions to these problems are proposed in this dissertation:(1)To address with the problem of information fusion from different modalities for the crossmodality person re-identification task,a Dual attention-aware Fusion Network(DAFN)is proposed.This method consists of two parts: a cross-modality feature extraction module and a self-attentive feature fusion module.The former is divided into modality-specific feature extraction and modalitysharing feature extraction to reduce modality differences from the feature extraction perspective.The latter preserves effective discriminative information through intra-modality self-attention embedding,and inter-modality self-attention enables inter-modality information to be fused interactively while suppressing unnecessary person background noise.(2)To address with the problem of semi-supervised scenarios with small amount of labeled data and large amount of unlabeled data for the semi-supervised cross-modality person re-identification task,a Cross-modality Pseudo Label Learning Network(CPLN)is proposed.The method consists of two parts,the identity alignment module and the pseudo label generation module.The identity alignment module performs identity alignment from global and local level,and includes a centerbased triplet loss.The pseudo-label generation module contains two pseudo-label determination conditions and a dynamic center-based cross-entropy loss to optimize the feature distribution for unlabeled samples.(3)To address with the problem of person occluded by objects for the cross-modality person reidentification task,a Multi-stage Feature Interaction Network(MSFIN)is proposed.This method consists of two parts,a multi-stage feature fusion learning module and a multi-scale attention module.In particular,the multi-stage feature fusion learning module aggregates features extracted from multiple stages of the network to enhance person semantic information and reduce the negative effects from occlusion.The multi-scale attention module acts on the features output from each stage of the network and effectively highlights the features in unconcluded regions by scaling the convolutional block perceptual field.Two widely cited datasets(Reg DB and SYSU-MM01)are selected for experimental evaluation and analysis of the above methods,and those methods are compared with other excellent crossmodality person re-identification methods.The experimental results show that all three crossmodality person re-identification methods proposed in this dissertation are effective.