BROWSE
Detail
Journal Article
LAttE: A label-free and multimodal framework for context-aware person re-identification
Cited 0 time in 
Share 
Share
- Authors
- Dasom Ahn, Sangwon Kim, Kwang-Ju Kim, Byoung Chul Ko
- Issue Date
- 2025-11
- Citation
- Neurocomputing, v.653, pp.1-9
- ISSN
- 0925-2312
- Publisher
- Elsevier
- Language
- English
- Type
- Journal Article
- Abstract
- Person re-identification (Re-ID) refers to the task of identifying the same individual across multiple non-overlapping camera views, and it is considered a critical task in surveillance, security, and smart city applications. We introduce label-free attributes and pose embedding (LAttE), a novel Re-ID framework designed for intelligent surveillance systems that eliminates the need for manual attribute annotation. LAttE constructs a rich attribute bank by using GPT-4o to synthesize diverse human-centric descriptors, which are then embedded using a contrastive language–image pretraining encoder. These textual attributes are fused with visual and pose features through a cross-modal attention mechanism, resulting in a comprehensive representation of pedestrian appearance and structure. To further enhance robustness, we incorporate a feature alignment strategy based on the maximum mean discrepancy, improving consistency across varying viewpoints and sensor conditions. Experiments on benchmark datasets show that LAttE achieves state-of-the-art performance in both mean average precision and Rank-1 accuracy. These findings highlight its potential as a scalable and label-free solution for Re-ID tasks in intelligent vehicle applications, including pedestrian detection, in-cabin monitoring, and cooperative driving systems.