A benchmark dataset and lightweight model for multi-distance gesture recognition in UAV control

Unmanned Aerial Vehicle (UAV) gesture recognition is an emerging form of human-computer interaction that enables intuitive UAV control across diverse and dynamic environments. However, progress in AI-based UAV gesture recognition has been hindered by the lack of large-scale, well-annotated datasets. To address this limitation, we present MD-UHGRD, a static gesture dataset containing 2280 annotated images collected from multiple individuals under varied environmental conditions and distances. This dataset supports the development of accurate and generalizable gesture recognition models. Building upon this foundation, we propose SA-YOLO, a lightweight and modular recognition framework that combines gesture detection with face and pedestrian tracking to enhance UAV responsiveness in real-world scenarios. SA-YOLO integrates Spatial Asymptotic Feature Pyramid Network (SAFPN), Scale Pyramid Pooling with Cross Stage Partial Convolution (SPPCSPC), and Space-to-Depth Convolution (SPD-Conv) to improve spatial feature extraction and computational efficiency. Experiments on the MD-UHGRD dataset, conducted under standardized training and testing conditions, show that SA-YOLO achieves 93.2 % 0.2 mAP with only 10.3 M parameters and 48
1 FPS. These results demonstrate not only high accuracy and efficiency but also stable and reproducible performance, positioning SA-YOLO as a benchmark in UAV gesture recognition with a strong balance between accuracy, speed, and model compactness.