Autonomous Quadrotor Control Using Convolutional Neural Networks

Quadrotors are considered nowadays one of the fastest growing technologies. It is entering all fields of life making them a powerful tool to serve humanity and help in developing a better life style. It is crucial to experiment all possible ways of controlling quadrotors, starting from classical methodologies to cutting edge modern technologies to serve their purpose. In most of the times quadrotors would have combination of several technologies on board. The attitude angles and altitude control used in this thesis are based mainly on PID control which is modeled and simulated on MATLAB and Simulink. To control the quadrotor behavior for two different tasks, Obstacle Avoidance and Command by Hand Gesture, the use of Convolutional Neural Networks (CNN) was proposed, since this new technology had shown very impressive results in image recognition in recent years. A considerable amount of training images (datasets) were created for the two tasks. Training and testing of the CNN were performed for these datasets, and real time flight experiments were performed, using a ground station, a Arduino microcontroller and interface circuit connected to the quadrotor. Results of the experiments show an excellent error rates for both tasks. The system performance reflects a major advantage of scalability for classification for new classes and other complex tasks, towards an autonomous flying and more intelligent behavior of quadrotors