RETRACTED CHAPTER: Person Re-identification Based on Transform Algorithm

Abstract

Person Re-identification (Person ReID) is a new technology emerging in the field of intelligent video analysis in recent years, aiming to solve the problem of person re-identification and retrieval under cross-lenses and scenes. It is also a hot research topic of computer vision in recent years. However, the study of Person ReID faces many challenges, such as low image resolution, visual angle change, attitude change, light change and occlusion. Research on target ReID has focused on Person ReID and vehicle ReID, and most state-of-the-art methods are based on Convolutional neural networks (CNN) structures. CNN, although successful, process only one local region at a time and can cause the detail loss of the data due to the existence of convolutional and down sampling operations. For the above problems, this paper designs a new goal-oriented Person ReID architecture, called Trans ReID. In this approach, the images are first converted into several patches, and strong baselines are built, which is advantageous against CNN-based method studies on several Person ReID datasets. To further enhance the robust features under Transformer for learning, two new modules are proposed. We uses the movement and patch scrambling operations to rearrange the embedding of the patch for better identification performance and wider coverage, and we integrate and analyze some non-visual cues to reduce feature bias during camera view changes. The experimental results show our method achieves excellent performance on Person ReID datasets.