Research Of Person Re-identification Algorithm Based On Deep Learning

Person re-identification(Re-ID)is a technology that uses visual information to determine whether a specific person exists in an image or video.This technology does not need to monitor the structural information of the network to enable person identity association across camera devices.As one of the key technologies of intelligent visual surveillance system,person recognition plays an extremely important role in criminal investigation and security.Affected by factors such as shooting scene,shooting angle,person attitude,person clothing,etc.,the appearance of the same person is greatly different,while the appearance of non-travelers may be similar,which greatly affects the use of visual information Accuracy of person re-identification.This thesis makes in-depth research on this problem,and proposes an optimized person re-identification method from the perspective of feature construction and sample constraints.Specifically,this article does the following:(1)This thesis proposes a person re-identification method based on part feature fusion.This method uses a dual branch network structure combining global and local feature extraction as the overall architecture of the model.In the local feature extraction branch network,the individual body parts of the person are located by explicit semantic segmentation methods,and the part features with rich details are accurately obtained to improve the recognizability and robustness of person features.Secondly,this thesis proposes to construct a feature weight allocation module based on the attention mechanism,highlighting highly distinguished part features of the person,and further boost re-ID performance.(2)This thesis proposes a feature extraction network based on the dual attention mechanism.Based on the structure of the two-branch network,two attention modules(channel attention module and spatial attention module)are added to the middle layer of the local feature extraction network to dynamically obtain highly differentiated,Local features with strong robustness are finally combined with two attention features to improve the discrimination ability of person features.(3)This thesis proposes a joint optimization framework based on multi-task learning.From the perspective of enhancing sample constraints,the two learning tasks of matching and classification are integrated into a unified framework,and two loss functions are used for joint optimization,which strengthens the constraints of the model on sample changes within and between classes and improves performance of the model.(4)In this thesis,experiments are performed on two large-scale person re-identification datasets Market1501 and DukeMTMC-reID.The experimental results verify the effectiveness of each module designed in this thesis and the entire network model.