ConstrID: a hybrid benchmark of real and synthetic images for construction worker re-identification

Computer vision-based person re-identification (ReID) technology enables cross-camera pedestrian detection by extracting and matching visual features to determine identity consistency. In construction safety management, ReID is applied to identify and continuously track specific construction workers. However, mainstream ReID datasets, collected in public environments, cannot adapt to construction sites characterized by complex backgrounds, heavy occlusion, and high inter-person similarity, thereby limiting their practical application. To bridge this domain gap, this paper constructs and evaluates the ConstrID-Dataset, a specialized dataset comprising 5,137 images of 138 construction workers, innovatively combining real-world data with AI-generated synthetic data. Four ReID models were employed to comparatively evaluate multiple cross-domain training strategies on the ConstrID-Dataset and public datasets. Experimental results indicate that integrating construction-specific data enhances the models’ feature learning capabilities in industrial settings, achieving an optimal mAP of 77.95%. The proposed dataset provides a robust foundation for worker identification and tracking, significantly advancing construction safety management.