Markerless monocular vision-based localisation for autonomous inspection drones

Abstract

This paper presents a markerless vision-based localisation algorithm to enable an autonomous inspection drone to determine its pose relative to a known inspection target. During an offline mapping phase, a 3D catalogue of the inspection target's persistent image features is created and stored. A neural network is trained on regions of interest of the images for fast segmentation. During the online localisation phase, the images are first segmented and the detected features in the segmented areas are matched with the stored features in the target's 3D catalogue. A pose estimation algorithm is applied to the matched features to determine the pose of the drone relative to the target. Practical experiments show promising results with small position and attitude errors.