Research On Person Re-identification Based On Deep Learning Methods

With the quick development of contemporary society,public security problems have received increasingly attention from masses.As one of the most important means for public safety management,surveillance video plays a significant role in the processes of city order management,missing person searching,and the cases' detection.In theses processes,the positioning and tracking of target people is often an important prerequisite.On the other hand,with the explosive growth of video data,the cost of using manual methods to find and locate target pedestrians will become increasingly higher,and inefficient manual monitoring methods will have the risk of missing key information for detection,thus endangering people's lives and property.Therefore,using machines to accurately and efficiently locate target pedestrians has become an urgent need.Person re-identification refers to pedestrian matching in multi-camera networks in non-overlapping views.In this paper,we research on the this task and focus on the two most critical issues of it: feature extraction and metric learning.For metric learning,we explore the performances of several loss function and propose a multi-task learning strategy for network training.And for feature extraction,this paper focuses on the design of network structures for persons' local feature extraction.Specifically,this paper mainly completes the following works:(1)In the study of metric learning methods,a multi-task learning strategy that combines classification model and retrieval model is explored.First,we explore different types of loss functions and verify theoretical feasibility of each of them;Secondly,we propose a multi-task learning network that can be applied for training using multiple loss functions.Based on this network,we further explore the combined performance of different kinds of loss functions.Finally,through the experiments on the person reidentification datasets,the optimal loss function's combination is selected,and then this design is further integrated into the feature extraction network proposed in this paper.(2)In the research of feature extraction methods,a multi-branch local feature extraction network structure is proposed according to the highly structured feature of pedestrian images,and it is merged with the global feature extraction branch to form a multi-scale fusion network.After that,the misalignment problem in person re-identification task,along with it's potential impact on our model is analyzed.Based on this,a solution using the space transformer networks to solve this problem is proposed.(3)We integrate the spatial transformation network,the feature extraction network and multi-task learning strategy into one model,completing the end-to-end design of the entire network so that all parts of the model can simultaneously perform optimization and parameter updating.We verify the effectiveness of each module as well as the entire network on the public dataset Market1501,and further analyze all experimental results.