SlamDunk: Affordable Real-Time RGB-D SLAM

Abstract

We propose an effective, real-time solution to the RGB-D SLAM problem dubbed SlamDunk. Our proposal features a multi-view camera tracking approach based on a dynamic local map of the workspace, enables metric loop closure seamlessly and preserves local consistency by means of relative bundle adjustment principles. SlamDunk requires a few threads, low memory consumption and runs at 30 Hz on a standard desktop computer without hardware acceleration by a GPGPU card. As such, it renders real-time dense SLAM affordable on commodity hardware. SlamDunk permits highly responsive interactive operation in a variety of workspaces and scenarios, such as scanning small objects or densely reconstructing large-scale environments. We provide quantitative and qualitative experiments in diverse settings to demonstrate the accuracy and robustness of the proposed approach.