Cross perspective person re-identification (drone and ground cameras)
The primary objective of person re-identification (Person Re-ID) is to address the task related to feature extraction and matching of person images captured from varied perspectives. Person Re-ID has numerous potential applications, such as in the domains of national security, intelligent retail, pu...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2023
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/168062 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | The primary objective of person re-identification (Person Re-ID) is to address the task related to feature extraction and matching of person images captured from varied perspectives. Person Re-ID has numerous potential applications, such as in the domains of national security, intelligent retail, public surveillance, and others. Contemporary datasets and algorithms for Person Re-ID primarily emphasize camera perspective, focusing on matching and recognizing individuals across the same perspective or even within multiple camera perspectives. With the rapid development and widespread application of drone technology, the Person Re-ID technology based on the drone perspective has attracted extensive research attention. Specially, continuous tracking and re-identification of target individuals between drone and ground surveillance cameras may encounter issues, as the same person may exhibit different visual attributes from different perspectives. To bridge the domain gap between drone cameras and surveillance cameras, this project aims to provide data support for person re-identification models across different viewpoints (from drone to ground camera or vice versa). This project utilizes a variety of loss functions and comprehensively evaluates the cross-perspective generalization performance of ResNet50 models trained with various public datasets. Additionally, it utilized NTU-Drone, a self-collected and annotated dataset, for experimentation. Ultimately, this project has established a standardized benchmark and provided detailed information for future research and improvement of Person Re-ID techniques from the perspectives of drones and high-altitude cameras. |
|---|