Implementation of an UAV autonomous mission planning aimed at collecting geo-tagged images from a multispectral camera

During the last decade, the amount of research that involves unmanned aerial vehicle (UAV) has spread out by offering great number of challenges for artificial intelligence and knowledge representation. Autonomous vehicles have gained a huge relevance in different fields and they have been proposed for a wide range of indoor and outdoor applications, including searching, surveillance, infrastructure inspections and serial aerial image acquisition. The most common mission scenario involves placing on the UAV different sensors like GPS, camera, telemetry for data collection tasks where the data could be processed using off-line or real-time applications supervised by a ground station. The purpose of the thesis gets inspiration from the Spectral Evidence of Ice (SEI) project, aimed at providing innovative tools to inspect and identify the ice on airplanes, using techniques of photogrammetry and spectral analysis, to support operators in improving the efficiency of the deicing process. The autonomous capability of UAV structure can be useful to facilitate this procedure thanks to the usage of a multispectral camera. To achieve this goal, the developed topics are navigation, data analysis, communication protocol and devices interface. ROS (Robot Operating System) plays a key role in the development of the system helping to build the robot application and enabling the communication between the pc companion, meant to be the control unit for mission planning and management, and the autopilot, intended to be the flight manager of the UAV. Consequently, the goal of this study is testing the drone maneuverability, the application program interface and providing a collision free trajectory to reach the desired destination. Localizing, detecting obstacles, direction correcting, and tracking the trajectories with reasonable accuracy are features examined in the path planning problem. Therefore, it is crucial to further extend path planning algorithms and methods for autonomous drones. The algorithms used to develop the path belong to the graph search algorithm which, given a fixed point, generates a circular trajectory around a specific object placed within the test environment. Another accomplished task is the integration of a multispectral camera in the architecture system in order to take geo-tagged images each time the drone reaches a desired waypoint for inspecting and reconstructing purposes. The collected images are involved in several tasks. In the first phase they are exploited to get information about aircraft’s location in the apron. They are used to feed an already developed convolutional neural network (DeepWay approach) to create a dataset of segmented pictures. The information extrapolated from those images provides an occupancy grid map of the analyzed sector in order to achieve a new list of waypoints which optimize the path planning. So, in the second phase, the data obtained by the CNN allowed to navigate closer to the aircraft and have a better view and prospective to detect the ice through the multispectral acquisitions. Finally, all this work shall be tried out on a real case to validate and compare the results gained in the simulation environment.