ORIGINAL ARTICLE
Figure from article: Zero-Shot UAV Person Re-ID...
 
KEYWORDS
TOPICS
ABSTRACT
Person re-identification (Re-ID) in unmanned aerial vehicle (UAV)-captured images presents unique challenges due to extreme variations in viewpoint, low image resolution, and frequent occlusions. This study develops and evaluates a zero-shot Re-ID framework that leverages contrastive language-image pre-training (CLIP)-based visual embeddings fused with soft biometric attributes, enabling person retrieval without requiring training data. This approach extracts interpretable semantic features such as clothing color, gender, and body proportions, and combines them with high-level vision-language representations generated by CLIP to enable robust identity encoding. By integrating visual and textual modalities, the framework supports flexible query-based person searches for identities not seen during training. Experimental evaluation and comparative analysis demonstrate the effectiveness of the method, achieving a Rank-1 accuracy of 75.87% on the Aerial-Ground Re-ID (AG-ReID.v2) dataset, which comprises person Re-ID across mixed aerial and ground scenarios, surpassing other deep-learning approaches under zero-shot conditions. These results confirm that combining soft biometrics with vision-language models enables scalable, training-free person re-identification in real-world aerial surveillance applications.
Journals System - logo
Scroll to top