Research On Several Issues For Person Re-identification

In recent years,with the growing popularity of video surveillance systems,people put forward new requirements on the automation and intelligent of video surveillance system.As one of the important objects of surveillance,pedestrian detection and re-identification are becoming a hot research field in the area of computer vision.The purpose of person re-identification is to match different images belonging to the same person correctly under different monitoring scenes.This paper focuses on the problems of person re-identification,from basic image preprocessing to multi-scale fusion framework,and makes a systematic research on this problem.The main achievements and innovations are as follows:1.Whether based on feature representation or machine learning,extracting discriminative image features is the basic work of person re-identification.Under different surveillance scenes,the illumination changes drastically,the image resolution is low,and there is background information interference.In this paper,three solutions are proposed respectively:(1)Multi-scale Retinex algorithm is applied for image brightness enhancement;(2)Upsampling based on feedback control to improve image resolution and detail information;(3)A FH segmentation algorithm based on symmetric information is proposed to extract pedestrian contours.The above preprocessing operation can effectively improve the original image quality,reduce the computational complexity and improve the recognition accuracy.In addition,this paper proposes a new feature extraction method based on asymmetric information to improve the measurement and matching accuracy of the feature vectors.2.In this paper,person re-identification is regarded as an image retrieval problem.The weight adaptive multi-feature Late-Fusion algorithm based on metric learning is proposed to minimize the number of pedestrian images,which are wrong recognized.The algorithm learns one optimized metric matrix for each underlying feature based on rank loss.Weight adaptive fusion algorithm is then proposed to fuse different features.Finally,the similarity measure function is obtained.Experimental results on two published datasets demonstrate the effectiveness of the algorithm.3.Finally,inspired by the principle of the decision of criminals,we propose an adaptive multi-scale fusion framework,which merges conventional methods weightedly.More stable and reliable reordering results are obtained.The experiment compares the framework with other fusion algorithms on two public datasets,and verifies the validity of the fusion framework.This paper starts from the basic image preprocessing work,and then proposes a person re-identification algorithm based on metric learning,and finally designs an adaptive multi-scale fusion framework.The research on person re-identification is carried out from three levels systematically.The final accuracy of re-identification is effectively improved.