Occluded person re-identification

• Main Author: Li, Dichen
• Other Authors: Alex Chichung Kot
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/158080
• Institution: Nanyang Technological University

Person re-identification (person Re-ID) is defined as a matching process of a group of target people across different disjoint cameras. It can be applied to many public occasions for security use such as in the surveillance system of shopping centers. With the advancements in technology, person Re-ID has seen improvements over the last few years. However, the extreme assumption of full-body images is too limited for real-world applications. Existing person Re-ID algorithms face a serious performance drop in matching accuracy with the datasets with occluded images. This project focused on the occluded person re-identification task. After conducting a comprehensive survey of the existing public datasets and the state-of-the-art (SOTA) methods for the occluded person Re-ID problem, we categorized the existing methods into human key-point matching methods such as Pose-Guided Feature Alignment (PGFA) and attention-based methods such as Attention Framework of Person Body (AFPB). In this project, we have implemented several methods and conducted a comprehensive performance comparison using several benchmarks datasets such as, Occluded_REID, Partial_REID and Occluded_Duke. We evaluated the advantage and limitations of those methods. Finally, the findings of this project offer us some inspiration for the potential improvement in the future.