Tracking of Mobile Objects with an UAV and a DNN Controller

The success of the Ingenuity helicopter mission on Mars underscores the growing need for autonomous navigation technologies in Unmanned Aerial Vehicles (UAVs). However, detecting and tracking moving objects with unknown dynamics remains a challenge in planetary exploration. Current optimal control algorithms outperform classical controllers but struggle to generate control signals within the required operating time, leading to high computational costs. We propose a Deep Neural Network (DNN) architecture pre-trained with Model Predictive Control (MPC) for horizontal motion control, coupled with a Proportional-Integral-Derivative (PID) controller for the altitude and orientation of a UAV for mobile target tracking. This hybrid approach reduces computational costs, significantly improves the speed of control signal generation, and maintains performance similar to MPC even in scenarios where it was not trained on. In these cases, when the target vehicle increased its speed, the neural controller was able to follow it without the vehicle escaping the field of view. Control commands are computed from the estimated trajectory using visual information from RGB images and UAV states. Testing is conducted in Gazebo with the Parrot Bebop 2.0 and Husky Robot.