Research On Person Re-identification Methods Based On Deep Metric Learning

Image retrieval is an important research task in the computer vision field.Person re-identification(Re ID)is a subtask of image retrieval,which uses computer vision technology to determine whether there is a specific pedestrian across multiple non-overlapping cameras.Person re-identification has great application prospects in criminal investigation,security,and missing person searching,and has gradually become a hot issue in the computer vision community.In this dissertation,we conduct deep research on the deep metric loss function in person re-identification,design a new loss function for unifying pair-based loss and classification loss,and propose a person re-identification network based on the attention mechanism.The main contributions of this dissertation include the following aspects.(1)Aiming at the inconsistency problem of the gradient direction of the multi-task loss function,the deep metric loss function form that unifies classification loss function and the pair-based loss function is derived.The new loss function form solves some problems such as hard sample mining,exploding gradient,and pair weighting.Experiments show that the loss function form improves the performance of the person re-identification model in a certain extent.(2)In order to unify the cosine distance and Euclidean distance,a polar project distance,optimizing the vector distance and the vector angle at the same time,is designed.For reducing the cost caused by manual searching of hyperparameters,a threshold adaptive module that automatically learns the hyperparameters of the loss function is constructed.For integrating polar project distance,threshold adaptive module and deep metric loss function form,an adaptive polar project loss function is proposed.Experiments show that the method has better performance in person re-identification tasks,compared with the existing loss function.(3)Aiming at the problem that the common backbone network only pays attention to the global coarse-grained features and ignores the local fine-grained features,a person re-identification network based on the attention mechanism is constructed.Aiming at the differences caused by different camera domains of pedestrian images,an instance batch normalization(IBN)module that increases the generalization and robustness of the model is embedded in the network.The experimental results show that the constructed network achieves great result,compared with the traditional backbone network.