Visual based landing for an unmanned quadrotor

One of the main challenges for Unmanned Aerial Systems (UAVs) is to extend the endurance of small vehicles such as multi-rotors. Actually, Li-po batteries that guarantee a flight of about 20 min power this type of vehicles. The endurance can be extended by enabling vehicles to look for recharging station(s). In this paper, we propose a vision system able to detect and track a given pattern hosted on the target-landing platform. The pattern is also useful to estimate the UAV position while approaching the target or during the hovering close to the target. The paper focuses on an optimized adaptive thresholding technique that manages critical situations as changes in the scene’s illumination / shadows. The developed system runs at 90 Hz for processing a 752×480 grayscale image. Preliminary results on an NVidia Tegra Jetson K1 platform are also presented to distribute the computation between CPU and GPU.