Research On Person Re-identification Based On Multi-granularity Feature Fusion

Person re-identification refers to the technology of using computer vision to retrieve the specific pedestrian across regions and cameras,which is often used in the field of intelligent security and criminal investigation.In recent years,with the development of economy,public safety issues have become more and more important,and person reidentification has become a research hotspot in the field of computer vision.However,due to the impact of many factors such as the low image resolution,pedestrian posture changes,illumination and background changes,and occlusion in real scenes,how to extract discriminative pedestrian features is still a huge challenge.The traditional person re-identification method recognizes by hand-crafted feature.However,due to the limited representation ability of hand-crafted feature,it is difficult to train a robust person reidentification model.Therefore,this thesis research on person re-identification based on deep learning.The main work and innovation of this thesis are as follows:(1)Aiming at the problem of camera style difference and occlusion,a data augmentation method based on CycleGAN and random erasing is proposed.Due to the different camera models and placement positions,the styles(such as brightness,color,etc.)of the images captured by different cameras tend to be quite different.In order to reduce the impact of camera style differences,CycleGAN is used to transfer the camera style of training image,that is,while keeping the original training image content unchanged,the image is converted from the original camera style to other camera styles,and the original training image and the generated image are mixed as a new training set.Aiming at the occlusion problem,the random erasing algorithm is used to simulate the occlusion of the training image to enhance the generalization ability of the model.(2)Combining global feature and local feature,a person re-identification method based on multi-granularity feature fusion is proposed.Firstly,the feature extraction network is replaced by Res Ne St50,and Res Ne St50 introduces attention mechanism.The attention mechanism makes the model pay more attention to the pedestrian in the image and reduce the interference of background changes.Then,the global average pooling and fully connected layer of Res Ne St50 are removed and replaced by global and local branches.The global branch focuses on the overall information of the image and extracts coarse-grained global feature.The local branch divides the image horizontally to extract more fine-grained local feature.In the training phase,the above data augmentation method is used to enhance the generalization ability of the model.In addition,the re-ranking algorithm is used to optimize the initial ranking list obtained by the similarity measurement to improve the accuracy of recognition.Finally,experiments are conducted on the Market-1501 dataset,Duke MTMC-reID dataset and MSMT17 dataset.The experimental results show that the accuracy of the proposed person re-identification method based on multi-granularity feature fusion is better than that of existing methods such as MGN and Pyrimad.Rank-1 accuracy achieves96.3%,91.2% and 83.8% respectively,and m AP achieves 91.2%,80.9% and 61.5%respectively,which proves the effectiveness of the proposed method.