Autonomous drone navigation using deep learning and computer vision

In this study, a configurable autonomous UAV system and platform are presented for operations that require autonomy and inference ability. The system is capable of performing autonomous take-off, planned flight and landing while also making inference about objects. Autonomous flight system is built with open source flight stack of Pixhawk and Robot Operating System (ROS). If navigation task is known as waypoints, controller node in robot operating system sends predefined route to flight stack via communication protocol. That allows the drone navigates in open fields as desired. Also, object detection is implemented with widely used DNN algorithm. This algorithm is not only retrained by applying fine-tuning, also optimized with TensorRT engine so that system performance is improved in terms of accuracy and speed. That is to say, developed UAV is aware of environment during flight. Entire system runs on Jetson TX2 with high GPU performance. Also, SITL method is utilized within a simulator for debugging and fast control prototyping purpose so that most probable crashes can be prevented or mitigated before occurring during real tests. In addition to software a UAV platform is produced for running the developed system in real environment. Design consideration of UAV, hardware selection and production are presented.