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

• Main Author: Carmon, Daniel
• Other Authors: Alex Chichung Kot
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/176553
• Institution: Nanyang Technological University

Person Re-Identification (Person Re-ID) is a challenge which main goal relates to the matching of person images obtained from various cameras. Person Re-ID is growing in importance in several key fields relating to homeland security, surveillance, and sports performance. Cross Perspective Person Re-ID delves specifically into when the matching of person images when the images are gathered at different viewpoints from one another. Currently datasets linked to Person Re-ID do not consider issues within the cross-perspective realm. As it is primarily focused on recognising people from a similar vantage point. This report builds upon a previously collected dataset from Nanyang Technological University, expanding upon this with the collection of a new dataset. This new cross-perspective dataset is approximately twice as large as the preliminary dataset. This can be used as a more comprehensive testing dataset where models trained upon the existing public datasets can be experimented in cross-perspective specific challenges. The three targeted problems that are discussed in this report concerns differences in Visual features, where from a higher viewpoint some characteristics intrinsic to a person can be obscured. A second challenge is in person alignment- at high drone perspective person images are tilted at a steep angle which lowers efficiency of models. Lastly as drone images are taken physically at a further distance to ground images the images of people are lower in resolution as they consume less on-screen pixels. Utilising a DEX framework resulted in an improvement of 14% Rank-1 and 24% mAP for mitigating view changes challenges. Recorded improvements of 3-6% in Rank-1 score by aligning non-orthogonal angles to a vertical position, addressing the challenge of person alignment at high drone perspectives. Switching from a ResNet-50 backbone to a HRNet backbone improvements of 6-12 % and 1-4% for Rank-1 and mAP metrics respectively have been attained. By looking at these issues an optimal model has been suggested along with future work to further improve cross-perspective person re-ID results.