Design Flow and Implementation of a Vision-Based Gesture-Controlled Drone
Arthur Findelair, Xinrui Yu, Jafar Saniie · 2022
The increasing efficiency of complex Neural Network architecture and the continuous improvement of embedded edge computing has reached a point that allows the deployment of advanced computer vision tasks on some of the most critical embedded applications, such as aerial drones. While similar applications were already possible by moving heavy processing on a ground station, an autonomous and centralized system significantly improves usability and security. The machine is thus self-sufficient and less prone to network attacks. Three main challenges stand-out during the deployment of our complex gesture recognition pipeline: (1) allowing user-defined controls, (2) ensuring robustness, and (3) on-board deployment. These challenges are tackled through handcrafted features to avoid the curse of dimensionality, neural network optimization on GPUbased companion computers, and data augmentation to cover reallife edges cases such as partial inputs.