Person Search Based on Improved Joint Learning Network

Person re-identification has received more and more attention in recent years. However, the pedestrian images used in most existing algorithms are always produced by cropping the integral surveillance images in manual or machining ways, and there is usually only one person in each of the cropped images. In this paper, person re-identification based on the integral surveillance images is researched, which is more close to the real-world scenario. The challenge of person search mainly comes from: (1) unavailable bounding boxes for pedestrians, (2) large consumption on time and hardware. To address these two issues, we propose a multi-level feature fused framework (MLF), which can deal with pedestrian detection and person re-identification in a unified network. The first module of the framework is served as the common module for both pedestrian detection and person re-identification. The second module is designed for pedestrian detection, in which three scales of feature maps from different layers are fused to get precise pedestrian bounding boxes; In addition, we use appropriate anchors and introduce soft-NMS into our algorithm to reduce missed-detections and false-detections, especially for small pedestrians. The third module is designed for person re-identification, we use an aggregated residual transformation for deep neural network in some convolutional layers, in which way the convolutional layers' parameters and training time could be decreased. Experiments based on CUHK-SYSU and PRW show the effectiveness of the proposed method.