Research And Implementation Of Non-overlapping Multi-camera Pedestrian Re-identification

Recently,with the development of safe city and smart city,video surveilance network has been widely used in various industries.As a core technology in the intelligent surveillance,person re-identification across non-overlapping camera views has drawn great interset.Person re-identification has a great application value in target tracking and person retrieval across cameras.Due to significant changes in low resolution,lighting,view angle,human pose and camera properties,the apperance of one person often undergoes large variations across different camera views,thus the problem of person re-identification faces a huge challenge.To cope with these challenges,we propose a novel method for person re-identification based on multi-features fusion and independent metric learning.The main work and innovation of this thesis are follows:1.Aiming at the limitation of single feature to pedestrian image description,this paper describes the pedestrian image with outstanding features from the point of view of different features with different abilities.Firstly,the improved Retinex algorithm is used to deal with the original image s to reduce the impact of illumination changes.Then,processed images are extracted from four color festures including HSV,RGS,CIE LAB and Ycb Cr feature and a texture feature of SILTP which has robustness to the illumination and shadow.At the same time,on the way of feature extract ed,according to the distribution regularity of the effective information of the image,a non-uniform segmentation method is adopted to keep the complete information of the pedestrian image and reduce the interference of the background area as much as possible.2.The person re-identification based on distance metric learning is usually to combine the extracted features in series and establish the feature model to obtain the metric matrix.This method can not make full use of the advantages of different attribute features,resulting in the lower discriminatory validity of the metric matrix,which can not accurately describe the difference and similarity between samples.Therefore,this paper proposes a method based on independent metric learning.It is used to learn the metric matrix in different feature spaces,and then the similarity function is obtained in the respective feature space.Furthermore,the similarity is weighted according to different weights.Finally,using the ultimate similarity matches the person image.3.Based on above work and innovations,a simple multi-camera person re-identification system is designed to achieve person re-identification and target tracking.