EdgeNeXt: A Lightweight Model for UAV-Based Gesture Recognition from Aerial Perspectives

Unmanned Aerial Vehicles (UAVs) increasingly collaborate with humans in scenarios requiring rapid, robust non-verbal communication, such as public safety, inspection, and logistics. Aerial gesture recognition faces key challenges: small and blurred subjects, cluttered backgrounds, and varying viewpoints , distances, and orientations. Prior work largely focuses on controlled or close-range settings, limiting real-world applicability. We introduce a novel outdoor dataset of five practical gestures (direction, acknowledge, land, stop, inspect) across eight orientations (0°-315°), altitudes up to 40 m, and diverse conditions, totaling 1,000+ RGB images, far exceeding prior datasets in scale and variability. We benchmark four state-of-the-art (SOTA) lightweight Convolutional Neural Networks (CNNs) (MobileNetV2 (M2), MobileNetV3-large (M3L), EfficientNet-D0 (E0), VGG16 (V16)) and propose EdgeNeXt (ENx), a custom lightweight architecture optimized for embedded UAV deployment. On gesture classification, M2 and M3L achieve 90.5% accuracy, E0 89.5%, and ENx 81.0% with 40× memory savings (0.87 MB vs. 33 MB) and 7× fewer FLOPs. For orientation estimation, ENx attains 97% accuracy, nearing the 100% of larger models. Our results enable efficient, real-time aerial gesture recognition for on-board UAV operation.