"The use of cameras and computer vision algorithms to provide state estimation for robotic systems has become increasingly popular, particularly for small mobile robots and unmanned aerial vehicles (UAVs). These algorithms extract information from the camera images and perform simultaneous localization and mapping (SLAM) to provide state estimation for path planning, obstacle avoidance, or 3D reconstruction of the environment. High resolution cameras have become inexpensive and are a lightweight and smaller alternative to laser scanners. UAVs often have monocular camera or stereo camera setups since payload and size impose the greatest restrictions on their flight time and maneuverability. This thesis explores ORB-SLAM, a popular Visual SLAM method that is appropriate for UAVs. Visual SLAM is computationally expensive and normally offloaded to computers in research environments. However, large UAVs with greater payload capacity may carry the necessary hardware for performing the algorithms. The inclusion of general-purpose GPUs on many of the newer single board computers allows for the potential of GPU-accelerated computation within a small board profile. For this reason, an NVidia Jetson board containing an NVidia Pascal GPU was used. CUDA, NVidia’s parallel computing platform, was used to accelerate monocular ORB-SLAM, achieving onboard Visual SLAM on a small UAV. Committee members:"

Ingeniero Civil Eléctrico / Este trabajo consiste en la generación de un set de datos con un respectivo ground truth (medición más confiable) y el uso de los algoritmos ORB-SLAM (Orientated FAST and Rotated BRIEF (Binary Robust Independent Elementary Features) Simultaneous Location And Mapping) y LOAM (Lidar Odometry And Mapping) a modo de entender de mejor forma el problema de SLAM (localización y mapeo simultaneo) y comparar los resultados obtenidos con el ground truth. A modo de entender de mejor forma el set de datos generado, la funcionalidad de los diferentes sensores es explicada. Los sensores utilizados para generar los datos son LIDAR, cámara estéreo y GPS. Este trabajo posee dos mayores etapas, en primer lugar, el GPS es estudiado para establecer las diferentes formas de extraer los datos desde el dispositivo. Una forma es generar un nodo de ROS que mediante comunicación de Bluetooth otorga un mensaje que puede ser leído. Otra forma es presionar tres veces el botón de encendido del GPS, lo que inicia el almacenamiento de los datos en la tarjeta SD. Mientras el primer método entrega mayor cantidad de información, es menos confiable, existiendo la posibilidad de guardar mensajes vacios o perdida de ciertos datos, afectando la tasa de muestreo. Finalmente una combinación de ambos métodos es implementada. Un set de datos de prueba es generado cerca de la Universidad De Chile, para probar que los datos están siendo almacenados correctamente. En el test se concluye que a modo de obtener mejor resultado con el GPS es necesario tomar los datos en zonas con baja cantidad de edificios. Finalmente con los datos y el ground truth el Error Absoluto de la Trayectoria (ATE) es calculado como método de comparación de ambas trayectorias generadas con los algoritmos mencionados. El ATE s la cantidad de energía necesaria para transformar la trayectoria estimada en el ground truth. Dadas ciertas limitaciones en la extracción de los datos estimados, la comparación se realizo entre dos set de datos de prueba, con pequeña cantidad de loops en el camino recorrido. En esta situación los resultados dados por LOAM son mejores que los obtenidos con ORB.SLAM. Pero en un ambiente con mayor cantidad de loops y una trayectoria más larga ORB-SLAM entregaría mejores resultados. ABSTRACT This work consists of the generation of a data-set with ground truth and the use of ORB-SLAM (Orientated FAST and Rotated BRIEF (Binary Robust Independent Elementary Features) Simultaneous Location And Mapping) and LOAM (Lidar Odometry And Mapping) algorithms as a way to better understand SLAM and to compare the ground truth and the data-set generated. To fully understand the data-set generated, the functionality of the different sensors is explained. The sensors used to generate the data-set are LIDAR, Stereo Camera and a GPS. This work is divided into two stages, in the first place the GPS is studied to establish the different ways to extract the data from it. One way is to generate a ROS node that through Bluetooth communication generates a message which is published. The other way is to press three times the button of the GPS to store the data in the GPS micro SD memory. While the first method is capable of store more data per second, it is less reliable, existing the possibility of store an empty message or simply the loss of data in the process. In the end, a combination of the two methods is implemented, modifying the bag file with the data stored in the micro SD. A test-data is generated near the University Of Chile, to prove that the bag file (a type of file that can contain any kind of information such as images, video or text, between others) is correctly generated. In these tests, it was concluded that to obtain better performance of the GPS therefore, obtain a better ground truth, it was necessary to generate the data in a zone with a low quantity of high buildings. Finally with the data-set and the ground truth the Absolute Trajectory Error (ATE) is used as a method to compare the trajectories. The ATE is the amount of energy that would require to transform the estimated trajectory on the ground truth. Since certain limitations of the extraction of the estimated path, the comparison was made between two small data-set which counted with low quantity of closed loops. Therefore the LOAM algorithm shows better results in this trajectory. The ORB-SLAM algorithm shows better results in data-sets with a high quantity of loops in the path.