Position based visual servoing with artificial neural network for quadrotor type UAVs

While Unmanned Aerial Vehicles (UAVs) perform certain tasks, their position and orientation during flight depend on the inertial measurement unit (IMU) and the global positioning system (GPS). Automatic landing in UAVs will alleviate the problems associated with adverse atmospheric conditions and eliminate possible errors as there is no pilot input, and without having to obtain permission from a control center for the area where the UAVs will land, they can land by using the GPS, IMU or VS already in place offers a solution to eliminate these problems. In the case of autonomous flight or autonomous landing missions, visual servoing systems ensure the completion of the mission in conditions where GPS systems are not available. In the visual servoing method, the camera mounted on the UAV is used to control the UAV according to the targeted mission. By providing the position or speed control of the UAV using the visual servoing method; low sensitivity, low cost, high durability system is obtained. The focus is on the PBVS system, which needs feature-dependent pose estimation for quadrotor-type unmanned aerial vehicles, as well as avoiding singularities originating from the interaction matrix. In this study, it is aimed to use field of view protection, unlike the classical PBVS. Although no prevention work is carried out, it is aimed to contribute to this study by checking whether the camera mounted on the rotary wing unmanned aerial vehicle, which is planned with a field of view protector, remains in the region so that it can see the features. In this thesis, it is aimed to use an artificial neural network for the stance prediction of KTGS, unlike classical approaches. In the first stage, a data set was created by trying 150 different scenarios to be used in the artificial neural network. Neural Network Toolbox was used to train the artificial neural network. With the trajectory planning, it is aimed to make the estimation of the center coordinates of the target features. It is aimed to test the convergence performances of both noise and trajectory planning on trained and untrained data by adding random noise to the artificial neural network posture predictor inputs. CoppeliaSim, which provides an easy and intuitive environment to create the virtual platform, was used in the simulation study.