Positioning system design for independent moving aircraft

Intensive work of scientists on the stable movement of unmanned aerial vehicles and especially drones in the air has largely overcome this problem. Indoor and outdoor GPS (Global Positioning System)-independent navigation systems have become a field of study on which today's scientists spend a lot of time. In this study, we tried to show that an aircraft using visual odometry with ArUco markers can perform localization based on instantaneous images with a high success rate. A simulation environment has been prepared in the CoppeliaSim simulation program to calculate the success rate of the localization. A warehouse space is planned as a closed space simulation. A series of racks are placed around the warehouse area and in the middle of the area, for a total of 27. In the simulation, each of the 27 racks in total is labeled using a 6x6 ArUco type label. ArUco labels are located at the upper right corner of each rack and are numbered counterclockwise. During the movement of the aircraft in the simulation environment, the location, IMU (Inertial Measurement Units) and camera information of the aircraft are obtained. From the image taken from the camera, the presence of augmented reality (AR) Tags are detected and the presence of the AR Tag, the coordinates of the AR Tag in pixels on the camera, and the area of the AR Tag is calculated according to the AR Tag ID. In addition to this data, the compass information of the aircraft is also taken. The actual location information of the aircraft is taken to be used as the output value in machine learning estimation. A regression model has been created for each location data. With the regression models, present location and pose estimates are produced in each region where the aircraft's camera saw AR Tags and the error value is calculated. R^2 performance values, which give information about the relationship between the x, y, z and γ angle regression estimates, which are the position information of the aircraft, and the actual values, were calculated. The scikit-learn library was used for the learning algorithms used and the default sub-parameters were used. Among the algorithms applied, AdaBoost gave 0.991 for the highest x value estimation for positioning, AdaBoost for 0.976 for the highest y value estimation, AdaBoost for 0.979 for the highest z value estimation, and AdaBoost for 0.816 for the highest theta value estimation. Based on the R^2 metric obtained in the current positioning and pose estimation; It has been seen that the prediction results are higher with the AdaBoost algorithm.