Reward-driven U-Net training for obstacle avoidance drone

作者：

Highlights：

• Monocular vision based drone with obstacle avoidance capability in the mazes.

• Segmentation model with U-Net infers the next moving direction for a flying drone.

• Reward driven Actor-Critic provides a depth map to U-Net, saving manual labeling.

• Our drone can fly not only in the trained maze but also in the untrained mazes.

摘要

•Monocular vision based drone with obstacle avoidance capability in the mazes.•Segmentation model with U-Net infers the next moving direction for a flying drone.•Reward driven Actor-Critic provides a depth map to U-Net, saving manual labeling.•Our drone can fly not only in the trained maze but also in the untrained mazes.

论文关键词：Autonomous drone,Reward-driven training,Actor Critic networks,U-Net,Policy gradient method,Reinforcement learning

论文评审过程：Received 14 August 2019, Revised 26 October 2019, Accepted 26 October 2019, Available online 29 October 2019, Version of Record 6 November 2019.

论文官网地址：https://doi.org/10.1016/j.eswa.2019.113064