Research On Cross-domain Adaptive Person Re-identification Algorith

In intelligent video surveillance,a person is the main body of intelligent surveillance video analysis,and cross-domain person re-identification(re-ID)is one of the important challenges in the field of public security intelligence.The purpose of cross-domain person reprofiling is to predict whether two images from different cameras belong to the same person.Compared with the great success achieved by single-domain person re-ID,cross-domain person re-ID using a re-ID model trained from the source domain to target domain will result in significant performance degradation due to the data deviation and scene difference between the source domain and target domain.In this paper,the spatial mismatch of person images and the quality of cross-domain false labels in cross-domain person re-ID are mainly addressed.The following two cross-domain person re-ID methods are proposed as follows:(1)This paper propose an adaptive grouping based on local semantic clustering of cross-domain pedestrian recognition method,for the person image background clutter and spatial mismatch problem,proposed semantic parsing is utilized to extract the person’s local semantic characteristics of top and bottom,and set up the control field improve the analytical accuracy,proves that the semantic alignment module for cross-domain heavy person re-ID model is effective.At the same time,the adaptive clustering module is proposed to improve the adaptability and accuracy of the unsupervised clustering algorithm,and the semantic alignment module and the adaptive clustering module are jointly trained to improve the performance of the person re-ID model.Experiments on market-1501 and Duke MTMC-re ID datasets show that the proposed method is 7.8% and9.6% better than D-MMD in Rank-1,which are currently state-of-arts cross-domain person re-ID methods.(2)In this paper,a cross-domain person re-ID method based on multi-label collaborative learning is proposed.First,a semantic parsing model is used to construct a multi-label data representation based on semantic alignment,to guide the construction of local features that pay more attention to person foreground regions,achieve semantic alignment,and reduce the influence of background on cross-domain person re-ID.Furthermore,based on the global feature of personal image and the local feature of a person after semantic alignment,the collaborative learning average model is used to generate a multi-label representation of the person re-ID model to reduce the interference of noise hard label in cross-domain scenes.Finally,a collaborative learning network framework is used to combine the multi-label semantic alignment model to improve the recognition ability of the person’s re-ID model.Experiments show that on the cross-domain person re-ID dataset of Market-1501→Duke MTMC-re ID,Duke MTMC-re ID→Market-1501,Market-1501→MSMT17 and Duke MTMC-re ID→MSMT17,Compared with the current state-of-arts cross-domain person re-ID method NRMT,m AP improves by 8.3%,8.9%,7.6% and 7.9% respectively.The proposed method improves the generalization ability of the cross-domain person re-ID model.Compared with the current state-of-art cross-domain person re-ID methods on multiple people re-ID data sets,performance of the proposed method on Rank-1 and m AP,the proposed method has better superiority.