Occluded area regularization and fine-grained feature extraction for occluded person re-identification

Person re-identification is of great significance in intelligent surveillance and other fields, but it often encounters the problem of occlusion. Existing methods mainly focus on feature extraction in the visible region and have issues such as ineffective fine-grained feature matching and neglect of regularization of feature extraction in the occluded region. To address these issues, we propose a method that enhances the model's ability to extract features of the visible regions from a more fine-grained pixel perspective. It involves four steps to handle occlusion: obtaining the occluded area, creating a mask for it, using a neural network to extract RGB features, and constructing the occluded area regularization loss. Experiments on datasets such as Occluded DukeMTMC show that our method outperforms previous methods and achieves good results on both occluded and non-occluded datasets, providing an effective solution for related research.