POSTER] ORBFusion: Real-Time and Accurate Dense SLAM at Large Scale

Abstract

Listen

Translate article

We present a new SLAM system capable of producing high quality globally consistent surface reconstructions with accurate real-time tracking and localization abilities. The system works on an off the shelf laptop with a typical GPU. This paper proposes an approach to unify feature-based keyframe techniques with fused volumetric surface reconstruction methods to overcome both of their limitations. On one hand, feature-based keyframe SLAM techniques have reached a level of maturity and can guarantee accurate and real-time tracking and localization ability, but their raw RGB-D point clouds are too noisy. On the other hand, volumetric surface reconstruction methods can produce a dense surface reconstruction of the environment, which will be helpful for Augmented Reality (AR) applications and scene understanding. However, current dense SLAM systems have limited tracking ability, which is vital for the quality of surface reconstruction. Moreover most of the current dense SLAM systems have to run on a powerful desktop PC to guarantee realtime performance. By unifying the feature-based keyframe tracking ability and adopting a multi-threaded design, our system improves both the tracking ability and the real-time performance. We present results of a wide variety of aspects of our system and evaluate it using the widely used TUM RGB-D and ICL-NUIM Datasets. Our system achieves unprecedented performance in terms of trajectory estimation, surface reconstruction, real-time and computational performance in comparison to other start-ofthe-art dense RGB-D SLAM systems.