Semantic map for indoor positioning system (floorplan enhancement with semantic SLAM)

Abstract

In navigation systems, the absence of GPS data poses an interesting challenge when it comes to managing drift. Systems such as ORB-SLAM manages this by performing Bundle Adjustment on loop closures, as such relatively accurate point-cloud maps may be generated from simple visual input only. In combination with semantic segmentation, semantically labelled point clouds are possible. A further visual enhancement may be made by comparing with a ground truth floorplan; not everything may be labelled in the floorplan and detections derived from the semantic cloud may be used for floorplan enhancement. Together, semantic slam and a ground truth floorplan may deliver a more visually appealing and accurate navigation visualisation. This project thus forms part of a system used to display a user’s position on a floorplan, as well as populate additional detections onto the floorplan using semantic SLAM by focusing on the frame matching problem between the machine generated semantically labelled octomap frame, and the floorplan frame.