Research On Metric Learning Based Methods For Person Re-identification

With the rapid development of computer vision technology and the construction of video monitoring network,intelligent monitoring system has received extensive attention.The technique of pedestrian re recognition is a hot topic in the field of intelligent monitoring.The existing data sets are small,resulting in small sample problems,making the learning model appear overfitting.Therefore,in this paper,based on the existing metric learning algorithm,an improved algorithm is proposed to alleviate the problem of small sample and over fitting.First of all,this paper studies the related theories of the re recognition of the people.The pedestrian re recognition can form the expression of the image by extracting the color and texture features of the image,and then classify the sample feature data and measure the distance.As the resolution of the pedestrian image is low,the details are not obvious,and the color and texture information are more significant.In this paper,we study the existing algorithm of image feature expression,elaborate the theory of the related feature expression model,and select the LOMO feature combined with the mission requirements and difficulties.Secondly,the pedestrian recognition technology based on metric learning method is introduced in this paper,and the excellent pedestrian recognition algorithm is introduced in detail.The existing method of pedestrian recognition based on metric learning is a more accurate measure of sample similarity by learning a Mahalanobis distance.In the process of distance learning,the distance of the metric is optimized by restricting the distance between the positive sample and the distance of the negative sample.According to the different constraint conditions,the recognition accuracy of the algorithm is different.In this paper,we compare the algorithms based on VIPeR database,and analyze the performance of the algorithm through the results.Finally,we propose a novel approach based on semi-supervised method to deal with the over-fitting problem and small sample size problem in the task of person re-identification.The learning method is improved on the traditional metric learning algorithm based on semi-supervised.On the one hand,it increases the number of model training samples.On the other hand,it can effectively restrain the over fitting phenomenon.The effectiveness of this method is verified by comparing the most advanced measurement learning algorithms in three common public data sets.