Metric monocular SLAM and colour segmentation for multiple obstacle avoidance in autonomous flight

Abstract

We propose an obstacle avoidance system based on image segmentation of the obstacles to be avoided in combination with a control policy for autonomous flight for which the MAV's position is obtained through a visual SLAM approach. The latter, however, utilises images captured from a monocular camera onboard the MAV, hence pose and map can be only obtained up to a scale factor. To address this, we incorporate metric via fixating the MAV's height and its onboard camera angle, which is set at 30 degrees in downwards direction to the floor. This enabled us to obtain a depth image of the floor observed by the camera that can be recorded only once and passed to a RGB-Depth SLAM system. Thus, MAV's pose can be estimated with metric, which is then considered into the avoidance rules. We carried out 36 flights with an 86 % of successful flights with no collisions, where obstacles were placed in different settings. Our approach is intended to solve one of the missions in the Indoors Category of the International Micro Air Vehicles Conference and Flight Competitions (IMAV) 2017, but we are certain that our approach could be extended to more general scenarios.