The Key Technology Research On Person Re-identification

In recent years,with the continuous expansion of the video surveillance network,surveillance systems have been playing an increasingly important role in daily life.Person re-identification(re-id),as an important research task in the surveillance system,has received extensive attention.The aim of person re-identification is to search and match persons of interested under cross-camera scenes.Based on a general basic framework of person re-identification,the paper has carried out specific research on person re-identification from the feature representation of static images and dynamic sequences,and unsupervised cross-dataset learning.The main work of the paper includes the following aspects:In person re-identification based on static images,a spatial saliency learning module is designed to enhance the role of effective local regions and weaken the impact of noise regions.A re-ranking learning algorithm is proposed to improve the performance of person matching without adding additional data.In dynamic sequence-based person re-identification,a local spatial weighting subnet is designed to highlight the features in common for internal regions of the sequence.The temporal attention weighting subnet is presented to merge temporal saliency information.The sequence similarity L1 loss is proposed to combine identification loss and verification loss to optimize the network.In unsupervised cross-dataset learning,channel weighting,spatial weighting,multi-scale pooling and metric learning are introduced to enhance the generalization ability of features.Moreover,a generation scheme based on KNN for training samples is designed to learn the metric matrix.The experimental results on six datasets,i.e.,CUHK03,CUHK01,VIPeR,PRID-2011,iLIDS-VID,Mars,indicate that the proposed method achieves a very competitive performance with the state-of-the-art approaches.