Video-based person re-identification via spatio-temporal attentional and two-stream fusion convolutional networks

Person re-identification aims at matching the identity of a same person that is captured from non-overlapping cameras. Most of the existing person re-identification methods are still focused on image-based solution. The problem of video-based person re-identification is better handled since video includes spatial and temporal information of pedestrians rather than a separate image. In this paper, we propose a novel approach for video-based person re-identification using spatio-temporal attentional and two-stream fusion convolutional networks, which consist of the two-stream fusion convolutional neural networks (TSF-CNN), the long short-term memory networks (LSTM), the spatial attention subnetwork and the temporal attention subnetwork. Specifically, the TSF-CNN learns the temporal and spatial characteristics simultaneously, and performs two fusions to achieve better feature representation. The spatial attention network is to automatically select the important part of pedestrian image in each frame. The temporal attention model assigns different weights according to the importance of different frames. Experiments on benchmark datasets demonstrate that the proposed approach is superior to existing methods of video-based person re-identification. (C) 2018 Elsevier B.V. All rights reserved.