Obstacle avoidance by combining background subtraction, optical flow and proximity estimation

Abstract

For Micro Aerial Vehicles (MAVs), robust obstacle avoidance during flight is a challenging problem because only light weight sensors as monocular cameras can be mounted as payload. With monocular cameras depth perception cannot be estimated from a single view, which means that information has to be estimated by combining images from multiple viewpoints. Here we present a method which focuses only on regions classified as foreground and follow features in the foreground to estimate threat of potential objects being an obstacle in the flight path by depth segmentation and scale estimation. We evaluate results from our approach in an outdoor environment with a small quadrotor drone and analyze the effect of the different stages in the image processing pipeline. We are able to determine evident obstacles in front of the drone with high confidence. Robust obstacleavoidance algorithms as presented in this article will allow Micro Aerial Vehicles to navigate semi-autonomously in outdoor scenarios.