Person Re-identification Based On Attribute Learning

Person re-identification is an important research direction in computer vision.It has a wide application prospect in man-machine interaction and security preservation.In recent years,with the in-depth application of deep learning in computer vision,a lot of excellent research results,typically the person re-identification based on pedestrian attributes,have emerged in person re-identification using deep learning architecture.Existing research shows that,compared with other methods,deep learning framework recognition algorithm based on the attributes of the pedestrians has a better recognition effect and stronger robustness.Therefore,attribute learning based algorithm is used to conduct person re-identification with the research on the design of the neural network structure,using of labeled data information and how to overcome the influence of the imbalance in the distribution of pedestrian attribute data.The main research contents are listed as follows:1.A person re-identification algorithm based on the prior distribution of pedestrian attributes in datasets is proposed.This method is based on the deep neural network to deeply mine the dataset information.The main work can be conclued in two aspects.On the one hand,based on the unbalanced number of samples among the attributes of the dataset,unified normalization was carried out in the calculation of the loss function of the network to avoid the impact of quantity differences.On the other hand,based on the imbalance of positive and negative samples within each attribute of the dataset,the weight of positive and negative samples in the loss layer are adjusted in the network according to the amount of proportion,so as to enhance the influence of positive samples on the recognition results.In addition,pedestrian attributes and pedestrian ID are jointly identified in the network,and appropriate joint loss function is designed to optimize the results.Experiment results show that this method can effectively solve the problem of sample distribution imbalance and achieve high accuracy.2.A strategy of person re-identification based on the classification of heterogeneous pedestrian attributes is proposed.This method further analyzes the inherent differences of pedestrian attributes,establishes relevant classification standards,divides pedestrian attributes into four categories and designs identification sub-network for each attribute category,then,adopts different identification methods to identify different types of attributes.In view of the inconsistency of the loss algorithms used by different attribute recognition methods,a measure function for the loss of heterogeneous attribute is proposed,which enables different identification methods to be trained and learned in the same network model and realize the optimization of network parameters.The experimental results show that this method improves the recognition effect of some heterogeneous attributes,and then improves the overall accuracy of the person re-identification.3.A person re-identification framework based on hierarchical recognition of pedestrian attributes is proposed.In this method,the correlation information between attributes is mined in depth.Firstly,the attention model is used to extract the characteristics of pedestrian attributes.Then,attributes are graded according to the significance of pedestrian attributes and the amount of information these attributes contain.At the same time,the dataset is analyzed to obtain the attribute co-occurrence matrix that can reflect the correlation between attributes.According to the recognition results and co-occurrence matrix of attributes at the previous level,the recognition strategy of attributes at the next level is adjusted.Experimental results show that this method not only proposes a relatively novel network architecture,but also effectively improves the recognition accuracy of pedestrian attributes,especially small target attributes,and then improves the accuracy of person re-identification.4.Based on the above research results,a framework of person re-identification based on attribute learning is designed.The framework integrates the above three improvements into the same network architecture,solves the problems of connection and parameter adjustment among the network parts,and designs a perfect algorithm flow.The results show that the proposed method is more accurate and robust than the existing person re-identification methods.