Pouso autônomo de VANTs baseado em rede neural artificial supervisionada por lógica fuzzy

Abstract

Unmanned Aerial Vehicles (UAVs) are shown as promising technology because of their high applicability and low costs. Thus, these vehicles are engineers and researchers studies targets that aim, in addition to applying them, to improve their performance, safety and make them autonomous and easily interaction. Flight stages such as takeoff, ascent, cruise, descent and landing are objects of studies to improve these aircrafts performance. Landing is a delicate stage for the vehicle, whose improper operation can result in accidents and losses. With this purpose, the present dissertation proposes a technique for the UAVs autonomous/assisted landing onboard the vehicle, without the use of ground control stations. As a sensing, the Ar Track Alvar computational vision algorithm is used to identify artificial markers used as a landing site. The landing site configuration aims the application of landing at higher altitudes, as different artificial markers are used for its composition. The landing algorithm is also a contribution of the present work, in which the execution is performed by an Multilayer Perceptron Artificial Neural Network (ANN) whose training is supervised by a logic fuzzy that uses the Mamdani inference. The use of fuzzy becomes viable due to non-deterministic characteristic and is less susceptible to sensing noise. Another important point is the no need to adjust gains for the procedure for each aircraft used, making the process dangerous and laborious. This setback is seen in classic controllers like the PID. Despite the advantages of fuzzy logic, this is computationally costly due to its Mamdani process. As a trained RNA is a set of matrix operations, it is proposed to train it supervised by the already functional fuzzy algorithm. This reduces the computational complexity of the embedded algorithm, facilitating image processing. The proposed landing firmware is developed on the Robot Operation System (ROS) and is focused on replication on real and embedded devices. The results are presented in Software in the Loop (SITL) and in real experiments at outdoor environments for static and dynamic landing spots. Comparison of algorithm performances is also shown. The performance was satisfactory and the RNA capacity and computational complexity reduction were verified.