Deep Learning Approach For Person Re-identification

In recent years,artificial intelligence technology has developed by leaps and bounds,and people's life has gradually become more intelligent.Computer vision-related technologies have played an important role in the field of public safety.Especially in intelligent monitoring,the environment of public area is complex,and the field of view of a single camera is limited.It is often necessary to find a specific pedestrian in the field of view of multiple different cameras.Due to the large amount of data involved,manual processing is often inefficient.At this time,research on person re-identification come into being.Its main goal is to complete retrieval of pedestrian image under the cross-camera at high speed and accuracy.Because pedestrian images are all taken from surveillance camera,the environment they are in is very complex,background,lighting and other factors interfere seriously.In addition,neural networks pay different attention to attribute information of different scales in pedestrian image.For these problems,this paper designs a person re-identification method based on Siamese-Verification network.The main research content includes the following two parts:(1)A person re-identification method based on multi-scale hybrid attention is designed.The feature extracting network consists of two branches.The multiscale hybrid attention branch is composed of a multi-scale learning module and a hybrid attention module.The multi-scale learning module implements multi-scale learning of pedestrian images by means of dilated convolution stacking.In addition,from the perspective of multi-attention fusion,the spatial attention of the concise parameter-free layer and channel attention are merged.Introduce a batch drop block branch,where the feature erasing module randomly selects pixels in a certain area of the feature map in batch units to strengthen the network's attention to the characteristics of different areas.(2)Use the above model as the basic model in the Siamese network,a person re-identification method based on the Siamese verification network is designed.This network improves the original Siamese network by adding a fc layer after the subnet for softmax loss and fully considers the id of pedestrian images.A phased optimization strategy that pretraining the feature extraction network and then fine-tuning the Siamese verification network is proposed.And in the stage of testing,we use re-ranking as the distance measurement method.Design a comparative experiment,perform experimental verification on three mainstream pedestrian re-identification datasets,and compare the experimental results with the results of some mainstream methods.The results show that the multi-scale learning,mixed attention,and feature erasure module in this paper can improve the performance of person re-identification.And by introducing the Siamese-Verification network,the rank-1 and m AP of the final model is better than the existing mainstream methods to a certain extent.