Multicopters Obstacle Avoidance by Learning Optical Flow with a Balance Strategy

Obstacle avoidance using onboard sensors is an important part of the safe and reliable navigation of autonomous aerial vehicles. For Micro aerial vehicles (MAVs), due to the extremely limited payload, it is a better choice to equip only one monocular camera. Although much attention had been paid to using optical flow to avoid obstacles mimicking the behavior of flying insects, these methods have met only limited success. Here, we propose a recognize-and-avoid method drawing lessons from the reactive obstacle avoidance methods. To let MAVs recognize the environmental conditions, we build an optical flow dataset for obstacle avoidance in the simulation environment and use a deep neural network to classify optical flow images into 5 labels. Then an avoidance policy is designed to mimic the "optical flow balance" strategy of flying insects. We analyze the proposed method in different simulation scenes and demonstrate the generalization of our method.