Person Re-identification Based On Deep Learning

Person re-identification is used to determine whether the pedestrian image acquired from different cameras is the same person,which has a prominent position in the field of public safety.As computer performance improves,new data sets continue to emerge,related research continues to emerge.This article describes the research methods in the field of pedestrian re-identification in recent years.Based on this and combined with the relevant theories of deep learning,we do some work as follows:For person re-identification based on tag data,this paper designs three kinds of implementation algorithms.Firstly,based on the original network which aims at single image classification,we add images similarity information by regarding it as a two-category problem.And the weighting the two classfication loss in model training;secondly,taking into account the similarities and differences between images,we use triples as model optimized goal.It aims to improve the expression ability of the entire network by constraining the distance relationship between similar images and heterogeneous images.Third,based on the general classification model,we use convolution to block the acquired image feature space,and then study image space continuity information by adding LSTM,and complete the model optimization.For person re-identification without tag data,we use pre-trained model to extract samples feature,and the extracted image features are clustered to obtain image tags.Then we do a screen to clustering result,which makes the data used in the training more reliable.At last,image pairs and triplet strategy were used for training and optimization of the obtained tag samples to complete the pedestrian re-identification task.For the algorithm proposed in this paper,in order to verify its validity,this article has carried on the experimental test to many pedestrian re-identification data sets,and through the comparison with some existing related methods,it has verified the effectiveness and superiority of this algorithm.