Fast Bayesian graph update for SLAM

Abstract

Robust, accurate localization is crucial to many of the current challenges in robotics. In recent years, visual simultaneous localization and mapping (SLAM) approaches have been studied intensively. The majority of approaches realize an excellent performance through precise visual odometry (VO) and loop closing based on a graph optimization technique. However, as the graph size increases, the computation time required for the graph optimization increases significantly. This makes it difficult to implement on mobile devices and causes a decrease in online performance. We propose a novel Bayesian graph update (BGU) to accelerate the loop closing of graph-based SLAM. The proposed method comprises sequentially updating the graph each time a new edge is obtained instead of optimizing the entire graph. Consequently, the update latency of one loop closing is reduced. The fundamental idea is to update whole edges on the graph using a Bayesian filtering method. We propose a fast algorithm based on a decoupled extended Kalman filter; its calculation time is proportionate to the number of edges at most. We confirm the speed and accuracies of our BGU using a simulated dataset. We also demonstrate that our BGU works well with actual visual SLAM methods by equipping it on ORB-SLAM2.