Research On Cross-modality Person Re-identification Based On Deep Learning

Person re-identification has important applications in the field of security,and cross-modality person re-identification matches pedestrians with the same identity in heterogeneous data to solve the problem of 24-hours monitoring.Compared to handdesigned feature,the deep learning-based cross-modality person re-identification algorithm has greatly improved in terms of accuracy and robustness,so it has important significance for cross-modality person re-identification.This paper aims at cross-modality person re-identification tasks,analyzes and designs cross-modality person re-identification algorithms based on deep learning from the aspects of neural network structure and loss function.Compared with existing papers,the method designed in this paper has achieved leading accuracy in two public datasets(SYSU-MM01 and Reg DB)for cross-modality person reidentification.The main work of this article includes the following three aspects:(1)A feature extraction framework based on hard pentaplet was designed.The framework has the characteristics of high accuracy and high scalability.The framework uses the improved single-modality person re-identification network as feature extraction module.Because the existing single-modality person reidentification model has stronger feature extraction capability than classification network,it also has high scalability.In addition,this paper designs hard pentaplet loss and a corresponding hard pentaplet sampling method.The hard pentaplet loss consists of a hard global triplet and a hard cross-modality triplet loss.The former handles both intra-modality and inter-modality variations.Since the inter-modality variations are more serious,the latter enhances the former's focus on inter-modality variations,and the loss function significantly improves training efficiency and accuracy.In this framework,we cascade the hard pentaplet loss and identity loss functions,further improving the accuracy.(2)A dual-channel network based on multi-granularity features was designed.The two channels are composed of two feature extraction modules,which has the same structure and different parameters.The two feature extraction modules extract the pedestrian features of the infrared images and the visible images.Each feature extraction module consists of a global feature branch and two local feature branches.The strategy of composition and fusion of global features and local features makes the network have the ability to extract multi-granularity features,so that the model can pay attention to the detailed features of different parts of the pedestrian.Finally,the discriminability of feature vectors is improved through identity loss and heterogeneous center loss functions.(3)The role of the gradient harmonizing mechanism in the field of cross-modality person re-identification was studied,and the effects of easy samples and outliers caused by excessive modal differences on model training are addressed.The identity loss based on the gradient coordination mechanism can improve the training results while keeping the model structure unchanged.This method is scalable and universal.