| Person re-identification is a sub-problem of image retrieval,which aims at retrieving the target person quickly in massive video sequences or images.In recent years,with the successful application of deep learning in various computer vision tasks,the performance of person re-identification has also been improved.However,in the face of the common occlusion environment in the actual scene,there will be problems like misalignment of parts,occlusion noise,occlusion and outlier of parts.The accuracy of the person re-identification methods without occlusion will drop significantly.The complex problem brings great challenges to the occluded person re-identification.In view of these problems,the main contents of this thesis are as follows:(1)Research data augmentation methods and CNN visualization methods.In view of the problem that invalid samples caused by related data augmentation methods are useless or counterproductive to model training,this thesis proposes a data augmentation method for occluded pedestrians based on key region retention.Improve the supervised single-sample data enhancement method.When randomly selecting the operation area,the class activation visualization technology is used to ensure that the key region is preserved,the complexity of the training sample is increased without losing key information,and reducing the impact of overfitting.Specifically,three improved occlusion pedestrian data enhancement schemes are designed,and the best scheme is determined through experiments.(2)Aiming at the problem that it is difficult for the model to learn robust pedestrian feature representation due to the misalignment of parts during part matching and the noise caused by occlusion,a pose-guided occluded person re-identification network model is proposed.The pose estimation module is designed to extract the features of human body semantic parts,and to divide the pedestrian images by locating the human body semantic parts,so as to solve the problem of parts misalignment during part matching.At the same time,in the backbone network,multi-scale branches are added to learn more discriminative deep features;a horizontal segmentation branch is added to divide the feature map into upper and lower parts of the body to improve the model’s ability to detect pedestrians in the upper and lower regions;design a multi-scale removal module for the noise of global feature.For global features of different scales,the attention mechanism is used to learn multiple sets of soft thresholds,and the noise features below the threshold are removed,so that the learning can be discriminative in the environment of redundant strong noise.global features that are robust and less intrusive.The experimental results show that the proposed pose-guided occluded person re-identification network model solves the problem of misalignment of parts,alleviates the noise problem caused by occlusion,and improves the global feature extraction ability of the model,making the retrieval accuracy of occluded person re-identification improved.(3)Aiming at the occlusion and outlier of parts during part matching,which leads to invalid part matching,an occluded person re-identification method based on part-part corresponding learning is proposed.On the basis of the occluded person re-identification network model proposed in(2),a new part matching branch network is added,which includes a part reweighting module and a part matching module.The part reweighting module generates adaptive interpretability weights for each part feature,attenuating meaningless part features and enhancing interesting part features.The part matching module uses graph matching technology to embed high-level part-part corresponding information into the branch network to enrich the information contained in the part features;and in the back-propagation process of the network,for the two graphs composed of the pedestrian key point features to be matched,the graph matching technique enhances the similarity of its corresponding keypoint features.The experimental results show that the proposed part matching branch network can effectively alleviate the problem of invalid part matching,and the result of part matching is better.(4)Combine the best occluded pedestrian data augmentation scheme in(1)and the occluded person re-identification method in(2)and(3),conduct comprehensive experiments on the full-body,partial,and occlusion dataset,then the results are compared with other excellent person re-identification methods.The experimental results show that the proposed method not only achieves the leading level in occluded person re-identification,but also achieves outstanding results in standard person re-identification. |
