Pose guided adaptive graph convolutional network based on pose for obscured pedestrian re-identification

In the Reid method which is based on deep learning, the combination of local features and global features can provide a robust representation for person retrieval. Human pose information can provide the location of the human skeleton, thus effectively directing the network to focus more on these critical regions, while also helping to reduce noise interference from background or occlusion. However, previous pose based approaches may not take full advantage of the pose information and rarely consider the different contributions of independent local features. In this paper, we design a pose guided multi-branch graph convolutional attention network (PG-MBGCAN) for the problem of re-identification of occluded pedestrians, which exploits the pose relationship to improve the robustness of pedestrian features and considers the contributions of different local features. Experimental results show that the proposed method in this paper has good performance in different datasets.