Study On Single-shot And Multi-shot Person Re-identification Algorithm

With the increasing emphasis on public safety of society,the video surveillance system has been deployed in major public places,providing technical support for social security and stability.In the video surveillance network,pedestrians are one of the most important focus.Pedestrian retrieval,especially for pedestrian retrieval under none-overlapping cameras,is an important research topic in the field of security.Therefore,how to establish a highly efficient and robust person re-identification algorithm is of great significance for research and application.Based on two different research scenarios in person re-identification,two methods are proposed to solve the problem of cross view person re-identification.The main contents of this paper are as follows:(1)For the single-shot,Most existing approaches seek for target by learning a global ranking model with the information of the entire image,which ignores the individuality of each person and the different contribution of each body part.Therefore,we proposed a person reidentification algorithm combining semantic attributes and body parts.Firstly,SVM is used to train the semantic attribute classifier,and then each attribute is weighted according to the probability of attribute prediction and the probability of attribute occurrence.Secondly,SVM learning is carried out for each individual upper and lower body part,and the classifier of each body part is obtained.Unlike the traditional classifier for the whole data set,we designed the classifiers for each body part of each individual respectively,so when calculating the similar distance can be well adapted to each individual.Then,we combine the distances of the upper and lower body part to obtain the final distance of the whole picture,and get the similarity results.Finally,experiments on three challenging datasets,VIPeR,QMUL GRID and PRID 2011,demonstrate that the person re-identification algorithm proposed in this work perform favorably against the state-of-the-art approaches.(2)For the multi-shot,the mainstream research direction is based on the set-based feature extraction and set-based metric learning.Most of the existing research methods focus on one direction.In this paper,we introduce the AdaBoost algorithm,and propose a framework of person re-identification integrating spatial-time feature and set-based metric learning.The proposed algorithm is tested on the PRID 2011 dataset and the iLIDS-VID dataset,respectively.Compared with the state-of-the-art methods,the experimental results show that the first recognition accuracy is improved with proposed algorithm.In this paper,the proposed algorithm is innovative.At the same time,it is of practical value to study the application of the algorithm for single-shot and multi-shot scenarios,and it can cover the most practical application requirements of person re-identification.