The teleoperation of aerial vehicles can be onerous for naive operators unless the robot is endowed with some autonomy, including sense-and-avoid capabilities. This ensures a safe and smooth navigation even in case of users' lack of experience or distraction. In this paper, we propose a reactive collision avoidance strategy that allows a micro aerial vehicle (MAV) to autonomously avoid obstacles while being steered by an operator. We assume that the only available measurements come from an onboard RGB camera and we adopt a collision avoidance strategy based on virtual force fields. A U-Net is used to estimate the depth map starting from RGB images. Simulations conducted in several different outdoor environments validate the proposed approach.