Research On Person Re-identification In Cross-modality Scene

In recent years,with the increasing popularity of "safe cities" and "smart cities" construction,the demand for video surveillance systems grew.The person re-identification technology in the field of computer vision could meet the requirements of urban video surveillance systems and has attracted widespread attention.However,in practical scenarios,person re-identification technology was affected by low visibility or night conditions,resulting in poor performance.Therefore,researchers shifted their research direction to cross-modality person re-identification technology,that is,how to effectively use multiple modality data(such as visible images and thermal images)to identify pedestrians became the focus of research.Similarly,cross-modality person re-identification technology also involves various challenges.It not only needs to consider intra-modality changes such as viewpoint,pose,and low resolution but also needs to address the inter-modality differences caused by different image channel information.Therefore,this paper aims to reduce the impact of these issues on cross-modality person reidentification and improve the identification performance.The main works of this paper are as follows:(1)From the perspective of the intrinsic relationship between the local features and global features in pedestrian images,this paper proposes a dual-branch spatial attention network model,which aims to establish spatial dependencies between local features in the pedestrian feature map and effectively align the local features of pedestrians in two modalities to enhance feature extraction.Meanwhile,a cross-modality dual-constraint loss function was proposed,which adds center and boundary constraints for each class distribution in the embedding space to promote compactness within the class and enhance the separability between classes.Our experimental results showed that our proposed approach has advantages over the state-of-the-art methods on two public datasets SYSUMM01 and Reg DB.(2)This paper proposes a cross-modality person re-identification algorithm based on graph structures by exploiting context relationships in pedestrian images.The proposed approach includes both intra-modality and inter-modality graph structures.The intra-modality graph structure constructs structural information with samples as nodes to learn the dependencies among the graph features within each modality;the inter-modality graph structure aggregates and maps the samples of the two modalities,combining the relevant neighborhood structure information between the two modalities to learn the same feature representation.Finally,the multi-head attention mechanism was used to assign adaptive weights to both intra-modality and inter-modality graph structures,making the training process stable and efficient.After extensive experiments,it was found that our proposed approach achieves significant improvements over the baseline models.(3)Based on the two proposed cross-modality person re-identification algorithms,this paper designs and implements a cross-modality person re-identification system.The system not only provides a visual way to display the results of person re-identification,but also monitors the running status of the algorithm in real time,achieving the application of the research content of this paper and better meeting the needs of users.