Research On Robust Person Re-identification Algorithm Based On Part Aligned Network

Person re-identification is an important part of intelligent video surveillance and has become a hot research topic in the field of computer vision.It has important research significance and widely application value.Its main research content is to match person images with non-overlapping camera views.Based on deep learning,this paper focuses on the part misalignment of the human body caused by complex variations in viewpoint and pose.Three part-aligned network models are proposed.The main work of this paper is summarized as follows.(1)An Appearance and Pose-Collaborative Part-Aligned Network(APC-PAN)is proposed.The APC-PAN is constructed by stacking several PAMs which simultaneously capture appearance features and pose features of person images.Appearance features and pose features are then combined to explore the implicit complementary advantages.The experimental results on the Market-1501,Duke MTMC-re ID and CUHK03 datasets indicate that the R1 accuracy of the proposed APC-PAN are increased by 2.4%,2.8%and3.0%respectively,compared with the PABR algorithm.(2)A Global-Local Representation Part-Aligned Network(GLR-PAN)is proposed,which uses multi-branch structure to extract global features and local features without introducing additional information.In order to improve the ability of feature representation,GLR-PAN effectively combines global features and local features.The experimental results on the Market-1501,Duke MTMC-re ID and CUHK03 datasets show that the R1 accuracy of GLR-PAN are increased by 1.0%,3.3%and 11.9%respectively,compared with popular person re-identification algorithms like PCB model.(3)Different from the above two methods,a Segment Attention-Guided Part-Aligned Network(SAG-PAN)is proposed.It uses Res2Net as the backbone network to extract multi-scale appearance features.At the same time,it uses the human parsing model to extract part features,which can be used as an attention stream to guide the network to complete part features re-calibration from the spatial dimension.Additionally,in order to ensure the diversity of features,SAG-PAN effectively integrates the global appearance features of person image with the part fine-grained features.The experimental results on the Market-1501,Duke MTMC-re ID and CUHK03 datasets show that the R1accuracy of SAG-PAN are increased by 0.5%,3.8%and 5.0%respectively,compared with the state-of-the-art like~2-Net,which demonstrates that the effectiveness of SAG-PAN in the task of person re-identification.