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

As the field of intelligent surveillance is developing towards the direction of working all day even under poor illumination conditions or extreme weather conditions,infrared cameras are also being widely used based on the deployment of visible cameras,RGB-infrared cross-modality person re-identification has attracted more and more attention.On the one hand,the person images with accurate bounding box images obtained through pedestrian detection as input is the significant basic for the excellent performance of person re-identification in practical applications.However,most of the current pedestrian detection research focuses on pedestrian detection in visible imagery while ignoring thermal imagery.Due to the differences in the images of different modalities,it is not ideal to transfer pedestrian detection results in visible imagery directly to pedestrian detection in thermal imagery.In the meantime,pedestrian detection in thermal imagery still has high research value constrained by lots of conditions such as changing of scales and occlusion.Therefore,pedestrian detection in thermal imagery still has high research value;on the other hand,suffering from large intra-and cross-modality variations,RGB-infrared cross-modality person re-identification still faces huge challenges.From the perspective of constructing an RGB-infrared cross-modality person re-identification system,this thesis conducts research on pedestrian detection in thermal imagery and RGB-infrared cross-modality person re-identification.The main work is as follows:(1)This thesis proposes an effective pedestrian detection network in thermal imagery.Aiming at solving the problem of great changes in scale of pedestrian in the detection scene,this thesis proposes an improved multi-scale feature fusion module,which makes it possible to fully combine low-level position information and high-level semantic information,obtain a richer combination of gradient information,and reduce the calculation amount;For the problem of imbalance between positive and negative samples,this thesis encourages more flexible continuous representation of predictive labels,and proposes to use Quality Focal Loss as classification loss and confidence loss.In the evaluation of the pedestrian detection network in thermal imagery proposed by this thesis,in terms of the evaluation of the KAIST dataset,LAMR(Log-Average Miss Rate)reached 21.64%,LAMR Day reached 27.48%,and LAMR Night reached 8.72%;in terms of the evaluation of the FLIR dataset,mAP(mean Average Precision)reached 78.1%,and the pedestrian AP reached 82.9%.(2)This thesis proposes an RGB-infrared cross-modality person re-identification network based on the dual-scale attention residual module.The network designs a novel dual-scale attention feature module,which can be used to focus as much as possible on the discriminative and robust features of the input sample person images in different modalities from the local scale and the global scale.On the SYSU-MM01 dataset,taking the evaluation in the single-shot+all-search mode as the example,the Rank-1 and mAP of the network reached 61.99%and 59.80%respectively;on the RegDB dataset,taking the evaluation in the Infrared to Visible mode as an example,Rank-1 and mAP reached 84.29%and 80.84%respectively.(3)This thesis proposes an RGB-infrared cross-modality person re-identification network based on dual attention constraints between intra-and cross-modality.The network contains two constraint modules,namely the adaptive graph structure constraint module and the dual-scale attention weighted constraint module.At the same time,Focal Loss is introduced as identity loss to improve the sample imbalance problem.On the SYSU-MM01 dataset,taking the evaluation in the single-shot+all-search mode as an example,the Rank-1 and mAP of the network reached 63.58%and 60.25%respectively;on the RegDB dataset,taking the evaluation in the Infrared to Visible mode as an example,Rank-1 and mAP reached 84.29%and 80.84%respectively.