Research On Re-ranking Algorithm For Person Re-identification

With the development of economy and society,the improvement of infrastructure,information flow,material flow and energy flow have become the main flow characteristics of modern society.Along with the rapid growth of staff turnover,the demand for security is daily on the increase.It has become a new normal to extract visual information from surveillance cameras to assist security.However,surveillance data often has the characteristics of large amount and less critical period.The traditional way of relying on human investigation is often insufficient.Although there have been many researches which proposed lots of methods and have achieved advantageous results,there are still many shortcomings and deficiencies in practical application.Beginning with the important task of video analysis field,person re-identification,this paper summarizes the main research direction of the task,and its development,typical research theory and solutions.Also,the main problems and shortcomings of the current person re-identification methods in practical application are proposed:(1)it is difficult to obtain training data.At present,the most effective algorithms are supervised learning algorithms,which need to label a large number of pedestrian data under the camera.Due to the complexity of the actual scene,it is necessary to label high-quality data for training in specific scenes in order to ensure the application effect,which is often time-consuming and costly.(2)It is difficult to get effective results.Limited by the acquisition of training data,most of the current algorithm models are trained once and deployed permanently.Due to the deviation between the training set and the application scene,the algorithms usually can not achieve the expected effect in the application.In the face of the problems and defects,in order to improve the re-identification effect,this paper focuses on the post-processing of the initial results by studying the reranking algorithm of person re-identification.This paper focuses on two aspects,which are automatic re-ranking and human-computer interaction re-ranking,and proposes an automatic re-ranking algorithm based on multi-similarity and a human-computer collaborative re-ranking algorithm based on active feedback.The former can automatically mine the contextual similarity and graph-based similarity of initial results,and obtain better outcome under the constraints of the two kinds of similarity.The later introduces human supervision information,automatically recommends samples with large amount of information for users to mark through the algorithm,and optimizes the results based on annotated information by metric learning and weighted manifold ranking.In addition,this paper designs and implements a person retrieval system based on the two proposed re-ranking algorithms,validates the effectiveness and practicability in the actual scene.In conclusion,this paper mainly studies the person re-identification re-ranking algorithm.In the paper,we popose an automatic re-ranking algorithm and a humancomputer collaborative re-ranking algorithm.In addition,this paper also designs and implements a person retrieval system,which validates the algorithms and system in actual application scenario.