Reconhecimento de ações humanas para controle de VANTS: uma abordagem baseada em deep learning

Abstract

This work presents an architecture for the control of Unmanned Aerial Vehicles (UAVs) based on convolutional neural networks, aiming to provide a more natural and accessible human-machine interaction through human gesture recognition. YOLOv8 models, in both classification and detection variants, were trained on a custom-labeled dataset comprising 1,175 images collected in both real and simulated environments. The YOLOv8x-cls model achieved the best performance, with an accuracy of 95.83% and the lowest rate of critical errors. The solution was integrated into a Tello EDU drone and evaluated in practical tests, demonstrating an average response latency of 1,042 ms and stable behavior through a temporal validation mechanism (assertiveness filter) combined with safe landing protocols. Results indicate that applying deep learning techniques for gesture recognition is feasible, safe, and effective for visual drone control in controlled environments.