iSAM2 using CUR matrix decomposition for data compression and analysis

Abstract

Abstract We introduce a factorization method to increase the calculation speed of incremental smoothing and mapping using Bayes tree (iSAM2), which is used in the back-end stage of simultaneous localization and mapping (SLAM), and to analyse the cause of the associated estimation error. iSAM2 is the method most commonly used to increase the accuracy of SLAM and shorten the calculation time required in real dense situations. In this paper, we describe the application of CUR matrix decomposition to iSAM2’s sparse linear system solver. CUR matrix decomposition is one of the low-rank matrix decomposition methods. It consists of matrices C and R, which are sets of columns and rows of the original matrix, and matrix U, which approximates the original matrix. Because of the characteristics of CUR matrix decomposition, it is possible to effectively approximate the sparse information matrix. Also, using principal component analysis, it is possible to identify the factors that increase or decrease the estimation error. We confirmed the feasibility of the proposed analysis method by applying it to real datasets and obtaining estimation errors similar to those obtained with iSAM2.