Vision based real-time obstacle avoidance for drones using a time-to-collision estimation approach

Considering the growth of drone usage in various applications in safety, security, and rescue domains, the need to obtain a collision-free flight is a central problem in autonomous drones. In addition, being the environments in these applications mostly unstructured, dynamic, and unpredictable, the capability of reacting in real-time to the environment is paramount. Using the visual information from a monocular camera, this paper proposes an obstacle avoidance technique based on Time-to-Collision (TTC) estimations. An optimal control problem is formulated and solved in real-time using a Model Predictive Control (MPC) approach, in cascade with an attitude controller, to determine the rotors' actuation signals. The MPC algorithm uses the target information from the vision method to lead the drone to a position that ensures a correct tracking in real-time and, simultaneously, use the TTC data to generate potential fields in the vehicle reference frame ensuring obstacle avoidance. The results in simulation show that the drone can accurately compute the TTC, track a target, and maintain a safe distance from obstacles. Moreover, the strategy reveal to be computationally non-demanding, indicating that the obstacle avoidance approach presented can be easily integrated in a real-time solution.