SkiMap++: Real-Time Mapping and Object Recognition for Robotics

Abstract

We introduce SkiMap++, an extension to the recently
\nproposed SkiMap mapping framework for robot navigation
\n. The extension deals with enriching the map with se-
\nmantic information concerning the presence in the environ-
\nment of certain objects that may be usefully recognized by
\nthe robot, e.g. for the sake of grasping them. More precisely,
\nthe map can accommodate information about the spatial
\nlocations of certain 3D object features, as determined by
\nmatching the visual features extracted from the incoming
\nframes through a random forest learned off-line from a set
\nof object models. Thereby, evidence about the presence
\nof object features is gathered from multiple vantage points
\nalongside with the standard geometric mapping task, so to
\nenable recognizing the objects and estimating their 6 DOF
\nposes. As a result, SkiMap++ can reconstruct the geom-
\netry of large scale environments as well as localize some
\nrelevant objects therein (Fig.1) in real-time on CPU. As an
\nadditional contribution, we present an RGB-D dataset fea-
\nturing ground-truth camera and object poses, which may
\nbe deployed by researchers interested in pursuing SLAM
\nalongside with object recognition, a topic often referred to
\nas Semantic SLAM.