Research On Person Re-Identification Based On Attribute And Identity Feature Fusion

Person re-identification is an essential component in the field of computer vision,which aims to match pedestrians across non-overlapping camera views.In recent years,with the improvement of computer performance and the increasing number of intelligent surveillance cameras,person re-identification has received more and more attention in industrial applications and scientific research.Based on previous research and existing challenges,we propose a network that fuses the attribute and identity features of pedestrians.There are four main contributions of the proposed method:1.A refined attribute prediction sub-network is proposed.The integration of traditional handcrafted features and deep neural network,combines the advantages of these two features.2.A weighted attribute prediction loss is designed.Considering the sample distribution ratios of different attribute categories,we set respective weights for each attribute loss to deal with the unbalanced sample distribution.3.A multi-region feature extraction sub-network is proposed.The global feature map is segmented to obtain more local information,then the features with stronger representation ability are acquired.4.A re-ranking method is proposed.The k-cross neighbors are defined to refined the initial distance rank.Experiments on three public datasets demonstrate the effectiveness of the four proposed points above,and the overall performance of the network is superior to the state-of-art methods.