Person Search Based On Joint Framework And Separated Framework

Person search is a technical study aimed to matching pedestrian identity from images taken from multiple surveillance videos.Because of its significant role in security system,the identification of pedestrians' individual identities has received extensive attention in research field and academic field.However,the pictures of current datasets used in the existing technical research are usually obtained through manual cropping or auto-detected by detector and labeling,while it is hard to obtain such similar pictures directly in practical applications.Moreover,pedestrian detection and person re-identification are explored as two independent research fields.Due to the complexity of outdoor environments,the current pedestrian detection algorithms can not perform well in practical application;Using standard labeled picture in person re-identification will inevitably cause a lot of accuracy problems.In order to make the research of the technology closer to the practical application,this article consider a comprehensive study of pedestrian detection and person re-identification,and proposed two pedestrian search algorithms based on joint framework and separate framework,respectively.The main contributions of this article are as follows:1.A person search algorithm based on inter-layer convolution fusion and pedestrians' anchor optimization under a joint framework is proposed.Dueing to the impact pf light and some other factor,one person can display different appearance in the same one dataset but different pictures,and light-processing was performed to decrease the difference.The basic network uses Resnet50 and the detection network uses Faster R-CNN,to optimize the detection effect on small targets,several output in different block of Resnet50 are fused.The statistics of the size and proportion of pedestrians' bounding boxes were made to design more suitable detecting bounding boxes for pedestrian.Besides,soft-NMS are used to reduce the missed detection on overlapping pedestrians.Finally,aggregated residual transformations are used to reduce the parameter training in the identification network.The verification results on two datasets are better than many algorithms.2.A person search algorithm based on image style learning and local feature response under a separation framework is proposed.Firstly,the Faster R-CNN algorithm is used for pedestrian detection alone,then the pedestrians in training set are cropped according to the data annotations,and the pedestrians in testing set are cropped according to the detection results.The cropped pedestrians are then sent to the identification network.In order to reduce the errors caused by multiple shots,a camera style learning method is used for image conversion,and then display person re-identification with multi-level local feature response,triple loss and classification loss are also applied to supervise the network.The algorithm is also trained and tested on two datasets,CUHK-SYSU and PRW.Experiments prove the effectiveness and feasibility of the algorithm.