Person Re-identification Based On Deep Learning

With the development of computer technology and the cost of camera reducing,the number of surveillance cameras in cities continues to increase,which plays an important role in the field of public safety.Cameras scattered throughout airports,malls,warehouse and residential buildings provide governments,companies and individuals with a large amount of image data,which facilitates the collection of evidence,prevention and tracing of crimes,and home safety.Person re-identification aims to retrieve a given pedestrian image from candidate pedestrian images,which plays an important role in intelligent surveillance system.However,surveillance cameras generate a large amount of image data and video data,and the processing of these data requires a lot of time and manpower,which increase the cost of public safety service.In recent years,the technology of deep learning has made great achievement,which provides methods for efficiently and accurately processing data generated by surveillance cameras.Person images with low resolution usually have small targets,and there is high similarity between pedestrians.Directly using traditional deep learning methods for person re-identification may cause misjudgement and poor cross-domain performance.Therefore,this paper focus on the application of deep learning methods in person re-identification.The main research work and contributions are summarized as follows:(1)Current deep learning methods for person re-identification usually extract global features or local features.But global features are difficult to avoid interference caused by background,and local features are easily biased due to the misalignment of person images.Aiming at solving the limitations of global features and local features,this paper proposed an attention network for person re-identification.Through spatial attention and channel attention,this method can assign more weights to salient regions and high-discrimination features,thereby improves the performance of the person re-identification model.This method has been tested on the Market-1501,DukeMTMC-reID and CUHK03 datasets.In contrast with recent methods,the proposed method gets better performance.The results of experiments have proved the effectiveness of this method.(2)The existing person re-identification models generally have achieved excellent performance when tested on the single domain,but when the models are trained on a dataset(source domain)and tested on another dataset(target domain),the performance will decrease a lot.In order to solve the domain adaptation problem of person re-identification and improve the cross-domain performance of the model,which helps reduce the time and manpower spent on annotating images,his paper proposed a part invariance network for cross-domain person re-identification.This method learnthe local invariance of the target domain,the invariance of the camera and the invariance of k-nearest neighbors with the assistance of a memory structure to measure the loss of the target domain without labels.So that the unlabeled target domain images with the same person identity can get close to each other during training.Finally,the method is tested on the existing dataset and compared with the current cross-domain person re-identification methods.The results of experiments have proved the effectiveness of this method.