Pedestrian re-identification based on multi-stream modal

Abstract

The goal of PReID (Person re-identification, PReID) is to maximize the function of distinguishing pedestrians in order to deal with the influence of uncertain background. The multi-layer learning method based on the neural network is applied to the background of extracting the feature information of the image, according to the composition method of the pedestrian structure, the problem of pedestrian rematching is considered in detail. Designed and proposed a multi-flow pedestrian re-identification algorithm, based on the neural network framework of global and local feature extraction, combined with multiple loss functions to optimize the pedestrian re-identification process, and achieve the purpose of pedestrian re-matching. Use long and short-term memory (LSTM) networks to model pedestrian images, focusing on the combination of information between global and local features. Using triple network multi-channel feature extraction method to expand the breadth of image extraction. In addition, the introduction of multiple losses to optimize the fusion of local information reduces the adverse effects of this part on network discrimination, realizes the complementarity of local body features and global features, and the complementarity of recognition tasks and ranking tasks to achieve higher feature learning accuracy. Experimental results show that the proposed method is much better than the latest method on three challenging data sets (including Market-1501, CUHK03 and DukeMTMC-reID).