Onboard Vision System for Detecting Emergency Situations on UAVs

Abstract

One of the main challenges in embedded systems for Unmanned Aerial Vehicles (UAVs) is executing high-computationalload tasks, such as those using neural networks and computer vision, in real time and autonomously, without relying on external processing. This article proposes a decentralized architecture that enables the implementation of a hand gesture recognition system onboard a drone. Initial results show that, at altitudes of up to 3.5 meters, recognition accuracies of 89% or higher are achieved for gestures such as OK, STOP or SOS (a combination of FOUR and FIST). Additionally, detection times for simple gestures like STOP and OK average no more than 370 milliseconds, demonstrating the system’s efficiency despite the computational demands. These results demonstrate the feasibility of performing real-time recognition tasks onboard a UAV while maintaining computational autonomy and an adequate response time for real-world applications.