Research On Person Re-identification Based On Sub-block Feature And Metric Learning Cluster Re-ranking

With the rapid development of artificial intelligence and computer vision,video surveillance technology has gradually become a research and application hotspot.Person re-identification task without overlapping vision is the key research content of video surveillance system,which has important application value in criminal investigation,intelligent traffic,person tracking and search.Due to the variable location of camera installation,the background of the person image taken is complex and changeable,and it is also affected by the change of camera perspective and the change of person posture,so the research on person re-identification technology faces many challenges.Therefore,more and more researchers have paid attention to how to establish a set of identification methods with high performance and robust to the above factors.This thesis focuses on the research of person re-identification technology for feature extraction,metric learning and re-ranking methods for monitoring person images,and mainly completes the following research work:(1)Based on the comparison of color features,several texture features,VGG16 network deep features,SIFT descriptors and other different features,a method of extracting coarse sub-block features based on human body parts was proposed.Firstly,the Retinex algorithm is used to preprocess and normalize the size of person images,then the head,upper body and lower body are divided into blocks in proportion and the LOMO features of these sub-block images are extracted.Then,the XQDA metric learning algorithm is used to learn the distance matrix according to each part,and finally the matching distance is obtained by weighted fusion of three parts.Experimental results show that the coarse sub-block feature extraction method based on human body parts can obtain higher recognition performance than the feature extraction method based on the whole person image.(2)A metric learning re-ranking method based on partition clustering was proposed.Firstly,the k-means clustering algorithm is used to cluster the person images in the training set to obtain the clustering center.Then,XQDA metric learningalgorithm is used to calculate the distance from the query person sample to the cluster center to obtain the distance matrix based on the cluster center of the test set.Finally,the original matching distance is weighted by the distance matrix based on clustering center.Experimental results on three public small sample person image data sets show that the metric learning re-ranking method based on partition clustering has obvious effect on improving person re-identification performance compared with the initial sorting results.(3)A person re-identification method based on the fusion of sub-block feature and metric learning clustering re-ranking is studied.Firstly,the LOMO features of different body parts are extracted.Then,the original comprehensive matching distance matrix is obtained through XQDA metric learning training and weighted processing for the sub-block features of each part.Meanwhile,the clustering centers are obtained for the LOMO features of the whole person image through k-means clustering,and the distance matrix based on the clustering centers are calculated.Finally,the weighted distance matrix based on clustering centers is used to correct the original comprehensive matching distance,and the final retrieval sort is obtained.The experimental results on three public person images of small sample data set show that the presented method based on the coarse sub-block features of parts of the body and learning based on partitioning clustering measure weight re-ranking,compared with other literature methods,can effectively improve the recognition performance of person re-identification task especially for small sample data set.