PERSON RE-IDENTIFICATION BASED ON DEEP LEARNING NETWORKS: A SURVEY
TL;DR
This survey offers an updated and focused review of deep learning-based ReID methods, encompassing research from 2020 to 2025, and investigates in-depth the engineering aspects, including system integration, real-time performance, and sensor constraints, which are often overlooked in reviews of earlier work.
Abstract
Person Re-Identification_(ReID) is a crucial task in computer vision with growing importance in security and engineering applications, particularly in surveillance and smart city systems. The hand-crafted feature-based existing approaches that consider texture and color struggle with complex real challenges involving lighting; person pose; and variable backgrounds. This survey offers an updated and focused review of deep learning-based ReID methods, encompassing research from 2020 to 2025. It investigates in-depth the engineering aspects, including system integration, real-time performance, and sensor constraints, which are often overlooked in reviews of earlier work. Techniques discussed in this study involve CNNs and transformers, triplet loss and contrastive learning, GANs, and methods that enhance matching accuracy and generalization. The paper compares recent methods; presenting their strengths and weaknesses, and setting directions for future research. The survey aims to provide a practical reference for engineers and researchers interested in developing robust and scalable ReID systems in real-world environments