Marker-Less Micro Aerial Vehicle Detection and Localization Using Convolutional Neural Networks

A relative localization system for micro aerial vehicles (MAVs), which is able to work without any markers or other specialized equipment, is presented in this letter. The system utilizes images from an onboard camera to detect nearby MAVs using a convolutional neural network. When compared to traditional computer vision-based relative localization systems, this approach removes the need for specialized markers to be placed on the MAVs, saving weight and space, while also enabling localization of noncooperating robots. The system is designed and implemented to run online, onboard an MAV platform in order to enable relative stabilization of several MAVs in a formation or swarm-like behavior, when operating in a closed feedback loop with the control system of the MAVs. We demonstrate the viability and robustness of the proposed method in real-world experiments. The method was also designed for the purpose of autonomous aerial interception and is a fitting complement to other MAV detection and relative localization methods for this purpose, as is shown in the experiments.