UAV-Human is a large-scale human behavior understanding dataset. The dataset was acquired by drones over a period of three months, capturing videos in multiple urban and rural areas during the morning and evening, covering a wide range of subjects, backgrounds, lighting conditions, etc.

The dataset includes 67,428 multimodal video sequences and 119 action recognition objects, of which 22,476 frames are used for pose estimation, 41,290 frames and 1,144 identities are used for person re-identification, and 22,263 frames are used for attribute recognition.