Research On Cross-modality Person Re-identification Method For Visible And Infrared

With the development of deep neural networks,the popularization of urban video surveillance systems and the improvement of surveillance networks,person re-identification(Re ID)technology has gradually become a hot research topic in the field of computer vision in recent years,and it plays an increasingly important role in the field of intelligent video surveillance and security.Traditional person re-identification technology is only suitable for cross-visible light camera person matching in daytime scenes,while cross-modality person reidentification(Cross-Modality Re ID)can match person cross-modality between the visible light mode during the day and the infrared mode at night.Recognition fills the application scenarios of traditional single-modality person re-identification.On the basis of traditional person re-identification,24-hour person re-identification is realized.In actual application scenarios,due to the modal differences caused by the high heterogeneity of color and infrared person images captured by visible light and infrared cameras,cross-modality person re-identification technology is facing a huge challenge.Therefore,this paper deals with the modal differences caused by visible light and infrared cross-modalities from the image level and the feature level to mitigate the negative impact on the cross-modality person re-identification task.The main work of this paper is as follows:(1)Aiming at the problem of the difference in the appearance of visible light and infrared person images at the image level and the insufficient samples of existing cross-modality data sets,a heterogeneous image augmentation(HIA)method based on a dual-stream network is proposed.In this paper,a lightweight heterogeneous image convolution generator is designed to convert the visible light image into a new sample similar to infrared,and the infrared image is color dithered to simulate the lighting changes in the actual scene.Input the original image and the generated image into the network together,based on the identity classification loss and the batch of difficult sample triple loss,use the designed positive sample pair constraint loss based on the heterogeneous image,and generate the heterogeneous image convolution The approximate infrared sample generated by the generator is the anchor point,and the distance between the anchor point and other positive samples is constrained,and the generator is optimized through the gradient back propagation.The heterogeneous image augmentation method based on the dual-stream network enriches the training samples,so that the person features in different modalities extracted by the network can have both modal common information and modality-specific information,effectively reducing modal differences.(2)Aiming at the problem of large modal differences in pedestrian features in different modalities at the feature level,On the basis of the method of heterogeneous image augmentation based on dual-stream networks,a cross-modality feature alignment(CFA)method based on modality classification is further proposed.This method designs a modal feature alignment module,which aims to guide the network to learn the common information of different modal features.This module contains two modal classifiers,which can classify different modal features using pre-built modal labels.Later in the feature network training,the visible/infrared modal features are classified as opposite modalities using the designed cross-modality feature alignment loss.During the network training process,the modal classifier and the feature network alternately update the parameters,so that the person features of different modalities have more commonalities in the high-dimensional feature space.Combining the two methods of image and feature level,cross-modality person re-identification is performed end-to-end.The method proposed in this paper achieves rank-1 57.82% and m AP 54.35% on the SYSUMM01 data set,and achieves the same on the Reg DB data set.The accuracy of rank-1 80.39%and m AP 75.05% proves the effectiveness of the method.