Research On Application Method Of Individual Re-identification Model Based On Video

With the development of intelligent video surveillance,more and more attention has been paid to the technology of individual re-identification,which is mainly applied to the criminal investigation and intelligent tracing.At present,most researches on individual re-identification are mainly based on images(including in-depth learning),but the information extracted from image is limited.This thesis focuses on the research of video-based individual re-identification,mainly including video-based individual appearance feature expression,individual spatio-temporal characteristics expression and individual re-identification distance metric model.Individual appearance feature expression.The appearance of the individual under different cameras will be changed after being affected by the light and a suitable method to find the rich texture features in the video.In order to solve the above problems,this thesis combines Lab color feature with spatio-temporal texture feature,extracts the histogram of Lab color feature from each frame sampling,combines the spatio-temporal texture feature,and uses the local average histogram to express the appearance feature vector.The similarity measurement of Top-push distance metric model proves that the combined feature with the single feature.Individual spatio-temporal characterist.ics expression.In order to solve the problem of time time.offset in video matching and retain all possible dynamic information in the individual motion.Based on the spatial pyramid model,this thesis uses the temporal pyramid and the spatial pyramid to obtain the spatio-temporal pyramid feature.This method enrich representation of person videos by imposing a temporal pyramid structure,motivated by pyramid match kernel and its spatial extension which is closer to the actual scene.The similarity metric based on the Top-push distance metric model verifies that the matching results.Individual re-identification distance metric model.For video-based research,individuals not only have similar appearance characteristics,but also may have similar motion characterislics,so they need stronger constraints.For this,Top-push idea is introduced into Mahalanobis distance metric learning algorithm,and Top-push distance metric model is obtained.In this model,the distance between the extracted features is constrained by increasing the class spacing and reducing the intra-class distance,so as to optimize the rank of individual re-identification.In this thesis,experiments are carried out on PRID2011 and iLIDS-VID datasets,and effective results are obtained by comparing them with different feature extraction methods and different models.