Analysis And Research On Person Re-identification Based On Person Attributes

In recent years,with the development of society and people's attention to public safety,more and more monitoring equipment has been deployed in major cities around the world.It has played a very important role in the cracking of criminal cases,the search for missing persons,urban security management and social stability.At the same time,with the massive amount of video data,the need for intelligent analysis and processing of video image data collected by monitoring devices by computers is increasing.Person re-identification,as an intelligent video analysis technique,has attracted extensive attention from many researchers.Person re-id aims to recognize person-of-interest from non-overlapping camera-views.Nowadays the most existing person re-id algorithm based on deep learning is mainly focus on matching cropped pedestrian images between queries and candidates,which is quite different from the situation in real-world scenarios.In view of the current problems,this dissertation aims to improve the person re-identification by the following three aspects of the work:1)In the practical application of person re-identification,if all the pedestrians in the candidate are manually cropped,it will take a lot of manpower.So the use of pedestrian detection algorithm becomes inevitable.In many of the previous work,pedestrian detection and pedestrian recognition were divided into two separate research sections.In order to improve the accuracy of the pedestrian detection and obtain the better rate of person re-id,we proposes an integrated person re-identification framework based on pedestrian detection.The pedestrian detection and person re-id are directly integrated into the same deep convolutional neural network,which can realize end-to-end training and testing.At the same time,we improve the Anchors in the Faster R-CNN and use a new loss function during the model training.Compared with other algorithms,the method proposed in this paper has the better performance in the two datasets.2)The common goal of person re-identification and person attribute learning is to describe pedestrians.Merely,person re-id will extract a person's global feature while attribute recognition focuses on local features.Considering the similarities and differences between these,this dissertation proposes a multi-task person re-identification framework.In the realization of the network structure,we adds a branch network of attribute predict into the integrated person re-id framework.Because of the parameter sharing,the computational complexity of the model is greatly reduced,and also can get the more effective pedestrian feature.Meanwhile,in order to comprehensively utilize the multi-scale information of the pedestrian images,we fuse the feature of the pedestrians attributes and identity.The experimental results prove that the proposed method can effectively improve the effect of the person re-identification.3)Since the research content of this dissertation is aimed at the real scene picture,and also will use the person attributes.Therefore,we present a 11 image-level attribute annotations for each image in the large-scale PRW and SSM pedestrian dataset.