Unicamp-UAV: An open dataset for human detection in UAV imagery

Human detection in UAV imagery is a critical task that has a wide range of applications, including search and rescue operations, public safety, and other scenarios, such as safe UAV path planning to avoid individuals positioned along the trajectory. However, detecting individuals in aerial images remains challenging due to the variability in human poses and sizes, differing UAV camera perspectives, and dynamic environmental conditions. Moreover, the lack of sufficiently large and accurately annotated datasets for training deep learning models for this specific task further complicates detection. To address this limitation, this study introduces a new dataset — Unicamp-UAV — comprising 6,500 images captured with a DJI Phantom 4 UAV and 58,555 manually labeled instances. The quality and usefulness of the Unicamp-UAV dataset were evaluated through experiments with different YOLO models — YOLOv7, YOLOv8m, YOLOv9m, YOLOv10-M, and YOLO11m — trained and tested on the dataset. All deep learning models employed in this study achieved higher AP50 values when trained with the Unicamp-UAV dataset than when trained with the Manipal-UAV dataset. The YOLOv8m and YOLO11m models demonstrated the best inference speed and mean accuracy, respectively. With AP50 exceeding 70%, these models prove suitable for real-time applications. The Unicamp-UAV dataset is publicly available and well-suited for training YOLO models to detect individuals positioned in front of the UAV in urban areas during daytime, with UAVs operating at altitudes of approximately 30 meters. Future research may leverage the Unicamp-UAV dataset to compare YOLOv8 and YOLO11 with Transformer-based detectors, which have also demonstrated high precision and real-time performance.