Cross perspective person re-identification (drone and ground cameras)

• Main Author: Zhou,Wenbo
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/168708
• Institution: Nanyang Technological University

Person Re-Identification(Re-ID) is a task to identify the same individual shown up in different camera video streams. It is getting more useful as the development of guard surveillance and camera technology. In the meanwhile, UAVs are also getting more and more frequently used, so combining Re-ID and Unmanned Aerial Vehicles (UAV) is a new area. Normally,the difficulties of person Re-ID is occlusion and background difference, low resolution, illumination changes, and viewpoint variations. In this project, different cameras at different perspectives are used to collect our own dataset(NTU-Drone). We use our ResNet baseline model to train Market- 1501, MSMT17-V2, DukeMTMC-ReID dataset, PRAI-1518, P-DRSTRE, and UAVHuman dataset and then test on UAV-Human, P-Destre, PRAI-1518 datasets, and NTU-Drone datasets. We are able to achieve the state of the art or at least idea accuracy. We classified our NTU-Drone dataset by the camera height to find the suitable big datasets to train before the test. Keywords: Person Re-Identification, UAV, Computer vision.