Research On Pedestrian Re-identification Methods Based On Deep Learning And Attribute Learning

As one of the most important techniques in video surveillance in public places,pedestria n re-identification has received extensive attention of researchers.At present,existing methods extract low level features such as color,texture and shape for distinguishing pedestrians.However,these digits based artificial features may be not the best for non-rigid pedestrian objects,and lack of semantic expression ability,so that it is not easy to be understood by users in the practical application.Besides,most methods for pedestrian re-identification have taken supervised learning methods,relying on a mass of labeled training samples.But in practical application,it is impossible to obtain a large number of labeled images for each pedestrian.In order to solve the problems mentioned above,this paper proposes a novel method based on Deep Learning and Attribute Learning.The main contents of the article are as follows:(1)We propose a pedestrian re-identification method based on unsupervised Convolutional Neural Network and pedestrian's attributes.By combing the structure of Convolutional Neural Networks and the training principle of Convolutional Auto-Encoder,we extract features of pedestrian images without supervision,which avoids the dependence on labeled training samples,and obtains more representative characteristics of pedestrian so as to improve the accuracy of pedestrian re-identification.Adding attribute layer between features and categories,and judging the pedestrian category through the attributes make the pedestrian re-identification be more semantic and practical in application.Experiments on VIPe R dataset show the superior performance of our method on the accuracy of attribute classifiers(2)We propose a pedestrian re-identification method based on unsupervised Convolutional Neural Networks and hierarchical attributes.This method extracts features from blocks of pedestrian images and allocates classifiers for each block,solving the problem of informative redundancy and interference,and improves the accuracy of attribute classifiers.We propose hierarchical attributes,and re-identify pedestrians with coarse-grained and fine grained attributes,making pedestrian re-identification be more in line with people's cognitive rules,and can deal with the re-identification tasks in different conditions.We validate the method on the VIPe R database,and the results show that our model outperforms the state-of-the-art methods for pedestrian re-identification,and is robust to the problem of attributes absence.(3)We design and realize a prototype system for pedestrian re-identification with deep learning and attribute learning.A compact GUI is developed by using MATLAB,including two main function modules of re-identification with target pedestrian images and re-identification without target pedestrian images.The functions include inputting target pedestrian images,selecting pedestrian's hierarchical attributes,pedestrian re-identification and displaying candidate pedestrian images,et al.