SARD: A Comprehensive Dataset for Construction Site Activity Recognition

The Site Activity Recognition Dataset (SARD) is a construction-site video dataset designed for human activity recognition and spatio-temporal activity detection in real-world construction environments. It captures diverse multi-person and multi-activity scenarios and uses a human-object interaction label system to describe construction activities through the relationships between workers, objects, and interactions. SARD contains two subsets. The core subset provides video-level annotations for coarse-grained construction activity recognition, including 748 video clips across 11 activity classes and 27 human-object interaction categories. The supplementary subset provides fine-grained frame-level annotations for spatio-temporal activity detection, including 9 video clips, 67,098 annotated frames, and 1,018,232 annotated person bounding boxes. The dataset is intended to support research on construction site activity recognition, worker behavior understanding, safety monitoring, and intelligent construction management. This repository includes dataset annotations, label definitions, train/validation/test splits, and publicly releasable video files or source metadata.