Enhancing UAV Autonomous Indoor Flight with Visual Odometry Techniques

In the rapidly evolving industrial landscape, Unmanned Aerial Vehicles (UAVs) have emerged as a pivotal component in enhancing operational efficiency and safety. These autonomous systems are particularly crucial in environments where direct human intervention is challenging or risky. The ultimate goal of this project is to assist human operators in specific missions by collecting and processing data. The drone should be able to provide stable flight in GNSS (Global Navigation Satellite System) denied environments exploiting the visual odometry algorithms and specific sensors. This thesis is a part of the FIXIT project, an initiative by the Competence Industry of Manufacturing 4.0 in Turin, Italy. The FIXIT project aims to establish a cooperative system between a UAV and an Autonomous Ground Vehicle (AGV), where the UAV can perform autonomous flights in industrial environments and dock on a moving rover. The drone is equipped with a Jetson Nano companion computer, a CubeOrange with Ardupilot, and a Lidar. The software stack includes Ubuntu 20.04,ROS Noetic, and other necessary software. An Intel Realsense d435i depth camera is also part of the drone’s equipment, which is crucial for the implementation of a visual odometry algorithm. The focus of this thesis is on the development of a localization strategy based on visual odometry exploiting the depth camera and other sensors to increase localization accuracy under various conditions. The RTAB-Map ROS algorithm is used for state estimation, and its parameters are tuned to optimize the speed of state estimation while maintaining an acceptable level of accuracy. To ensure smooth drone movement, a Python script was developed to fuse data from RTAB-Map, Lidar, and IMU using a filtering technique. This approach enables the drone to operate autonomously in indoor environments with high levels of magnetic fields, where GPS signals may not be reliable. Experimental tests were conducted with the drone in an indoor GPS-denied environment to validate the effectiveness of the proposed solutions. Future work includes testing more computationally expensive forms of data fusion algorithms and implementing a YOLO algorithm for object recognition. The potential applications of this work extend to security and inventory management in factories, among others. The findings of this thesis contribute to the ongoing efforts to enhance the adaptability and flexibility of UAVs in the context of Industry 4.0.