Map Slammer. Densifying Scattered KSLAM 3D Maps with Estimated Depth

Abstract

There are a range of small-size robots that cannot afford to mount a three-dimensional sensor due to energy, size or power limitations. However, the best localisation and mapping algorithms and object recognition methods rely on a three-dimensional representation of the environment to provide enhanced capabilities. Thus, in this work, a method to create a dense three-dimensional representation of the environment by fusing the output of a keyframe-based SLAM (KSLAM) algorithm with predicted point clouds is proposed. It will be demonstrated with quantitative and qualitative results the advantages of this method, focusing in three different measures: localisation accuracy, densification capabilities and accuracy of the resultant three-dimensional map.