Monocular Depth Estimation for UAV Obstacle Avoidance

In this paper, we present a novel method for obstacle avoidance of a quadrotor equiped with a single front camera. The proposed method is composed of three parts: depth estimation, obstacle detection, and obstacle avoidance controll. We use convolutional neural networks (CNN) to estimate depth from RGB image. Then the depth image is fed into the obstacle avoidance system, in which proposed control algorithm steers the quadrotor to fly away from obstacles, and after that, continue towards the destination. We conduct a lot of experiments, either in virtual environment with a simulated drone, or in real world with a quadrotor Parrot Bebop2, to verify the effectiveness of our method.