Applying artificial vision techniques and artificial neural networks to autonomous quadcopter landing

Autonomous navigation of drones continues to be a very broad field of research in which different methods have been proposed, using technologies such as geolocation and artificial vision. This work presents a method that combines artificial vision techniques and artificial neural networks (ANN) to achieve the autonomous landing of a quadcopter based on references obtained by detecting and locating a helipad near the end point of its programmed path. This method lets the system be independent from GPS data by using a control based on visual characteristics in the last flight stage instead. Images from a camera mounted on the quadcopter will be used to locate the helipad. Later, two artificial neural networks will operate in cascade to identify the marker and determine its position. This information will allow the drone to locate itself on the helipad and update its landing routine autonomously. Additionally, the Lucas Kanade optical flow predictor has been implemented to track the marker as a function of the strong characteristics obtained from the region of interest delivered by the ANN. This process is performed with the aim of reducing the computational cost of the proposed method and improving its execution time. To validate the proposed method, flight tests were carried out in which the landing point was to be located in various types of terrain, achieving 70 percent of success on highly roughened surfaces and 100 percent in homogeneous surfaces.