Research On Person Re-Identification Algorithms Based On Deep Learning

In recent years,the rapid expansion of camera networks has provided massive amounts of visual data for intelligence video surveillance,but it also greatly increases the difficulty of video processing.By identifying and searching a specific person image over a large camera network,person reidentification can effectively associate person identities across camera views,therefore they play an important role in many surveillance applications,such as object tracking,motion analysis and behavior understanding.However,person images usually contain many variations(such as pose,occlusion,background,illumination and resolution),and then those variations make the face features change greatly,leading to the performance degeneration of person re-identification.Based on the study of pose variations,modality discrepancies and longtailed distributions,a series of novel methods are proposed to improve the effec-tiveness and robustness of person re-identification models.The main contents and contributions of this thesis are summarized as follows:1.hierarchical attention and group attention high-order person re-identification frameworks are proposed to learn pose-invariant high-order person features and mitigate the pose misalignment problem.The Kronecker product is adopted to aggregate multi-level global and local convolutional features,and a count sketch function is used to transform Kronecker product into Hadamard product.Therefore,the dimension of the higher-order features is reduced without significantly damaging the representational capabilities.A group shuffle Kronecker product algorithm is designed to adopt channel group and channel shuffle strategies to fully learn intra-group and inter-group high-order interactions,which greatly reduces the time and space complexity of the higher-order features.Besides,convolutional features of multiple images and foreground regions are aggregated by group shuffle Kronecker product in order to obtain the corresponding high-order features.Experimental results prove that those two frameworks can effectively align person poses without relying on pose estimation and feature partition.2.A hard modality alignment network framework is proposed to solve the modality discrepancy problem.Considering that the modality discrepancies of different dimensions are unevenly distributed,this framework mines the hard subspaces with large modality discrepancies,and then focuses on aligning the modality distributions of those hard subspaces,which is helpful to learn modality-invariant person features.In addition,the framework can simultaneously eliminate global and local modality discrepancies,which helps to further improve the cross-modal generalization ability of re-identification models.Experimental results clearly demonstrate the superiority of the proposed framework over existing cross-modality person re-identification methods.3.A multi-patch matching network framework is proposed to mitigate the modality discrepancy problem.This framework can simultaneously align the modality distribution of coarse-grained and fine-grained patch features,and transfer semantic knowledges among different patches to learn the complementarity between coarse-grained and fine-grained patch features,so as to boost the robustness of person features.In addition,this framework can adaptively assign higher weights to hard patch tasks and lower weights to easy patch tasks,in order to highlight the priority of hard patch tasks.Experimental results show that proposed framework brings remarkable performance improvements and outperforms the state-of-the-art methods by a large margin.4.A novel multiple variation feature generation framework is proposed to solve class-imbalance and hard-imbalance problems.Different from the previous long-tailed re-identification methods,this framework equalizes the sample number of each person from the perspective of component decomposition and feature generation,so as to learn class-balanced person features.In addition,the adversarial learning method is adopted to improve the recognition difficulty of fake features,then these hard fake features can provide more opportunities to traverse the parameter space and obtain the optimal model parameters.Experimental results demonstrate the advantage of the proposed framework in terms of both effectiveness and efficiency.