Research On Visible Light-infrared Light Cross-modal Pedestrian Re-identification Based On Deep Learning

Person Re-identification,whose aim is to match images belonging to the same person across multiple disjoint cameras,is an image retrieval task that can help with surveillance video retrieval.The general Person Re-identification task is to match between color person images,but many surveillance cameras will enter the infrared mode to capture infrared images when the lighting conditions are not good,such as at night.In order to do Person Re-identification between visible images and infrared images,it is necessary to study the Visible-Infrared CrossModality Person Re-identification.Existing Visible-Infrared Cross-Modality Person Reidentification models are not effective.To improve the Cross-Modality Person Re-identification,this thesis had done two works:This thesis proposed Modality Batch Normalization module to alleviate the modality distribution differences that exist in Cross-Modality Person Re-identification.In CrossModality Person Re-identification,the same piece of clothing may look completely different in the infrared image and in the visible image.But it has to be mapped into the same feature,then the data distribution of this feature in the two modalities may not be the same,i.e.,there is a modality distribution difference.In order to reduce the modality distribution differences,this thesis designed Modality Batch Normalization module to adjust the data distribution of each modality using statistical information so that the data distribution of the two modalities is approximately the same.This thesis unified the loss function metrics used in Person Re-identification.When training Person Re-identification models,two loss functions are generally used: the proxy loss function and the pairwise loss function.However,there is a problem of inconsistent optimization metrics among the commonly used loss function combinations.That is,the metric for the optimization of the proxy loss function is the dot product,while the metric for the optimization of the pairwise loss function is the Euclidean distance.This inconsistent optimization objective problem makes the model ineffective.This thesis rewrote the form of the loss function so that both loss functions optimize the cosine similarity uniformly.In addition,for Cross-Modality Person Re-identification,this thesis introduced the random grayscale data augmentation method to increase the amount of training data while reducing the modality differences between visible and infrared images.This thesis also eliminated the hard sample pair selection of the pairwise loss function,allowing more cross-modality sample pairs to be optimized in each iteration.Based on these works,on the SYSU-MM01 dataset,proposed Cross-Modality Person Reidentification model achieved 64.78% and 62.46% of the Rank-1 and m AP evaluation metrics,respectively,with nearly 10% improvement compared to the baseline model.