Consistent Cuboid Detection for Semantic Mapping

Abstract

Building and storing efficient maps is an essential feature for long-term autonomy of robots. Modern sensors (such as Kinect) tend to produce a lot of data. However, long-term autonomy requires us to store this information in a succinct manner. One way to reduce dimensionality of information is to attribute semantics. Most indoor objects are cuboidal in nature. We conjecture that cuboids are a suitable semantic feature to attribute to indoor objects for efficient mapping. We adapt a cuboid fitting algorithm previously proposedfor object recognition, for indoor mapping. Our work stems from the observation that landmark detection for mappingrequires consistent detection of those landmarks. We implement several modifications to this cuboid detection algorithm that lead to consistent detection such as emptiness, orientation, surface coverage, distance from edges, and others. We incorporate these in the identification of the cuboid candidates in a scene, as well as an optimization algorithm for finding the best set of consistent cubes to cover a given scene. Our experiments show that in comparison, the set of cuboids detected by our algorithm are at least 50% more consistent based on our metrics.SLAM.