Geometric Mapping for Sustained Indoor Autonomy

Abstract

Simultaneous localization and mapping (SLAM) is the first step for enabling autonomous operation in unknown and changing environments. Many applications such as service and assistive robots require constant movement between different regions along with accurate navigation and localization at any point in time. Algorithms for SLAM have matured greatly over the last few years and can accommodate different sensors, computing requirements as well as environments for use. However, for long-term autonomy indoors, reasoning with a large volume of RGB-D data is still a major challenge. In this work, we propose a pipeline that attributes semantics, more specifically cuboidal structure, to observed objects, uses them as landmarks for mapping and thereby reduces the dimensionality of the represented map greatly. We chose cuboids, because many common urban scenes (such as offices, homes, malls) contain cuboidal objects (such as cabinets, tables, shelves). We develop a metric to perform such attribution consistently so they can be used as landmarks for mapping/navigation. We have tested our pipeline on three different datasets and show that we can reduce the map representation significantly while maintaining localization accuracy in all of them. Our vision is that attributing low-level semantics such as one presented in this work would make long-term autonomy computationally tractable.