Adaptive Sliding Window for hierarchical pose-graph-based SLAM

Abstract

We propose the Adaptive Sliding Window (ASW) which is a novel approach to solve the hierarchical pose-graph-based (PGB) simultaneous localization and mapping (SLAM) problem. We adjust the size of the sliding window (SW) for incremental optimization by eliminating the portion of the graph which has a low degree of similarity to the rest of the graph and by dropping poses which are not related to the latest robot pose. The decision is made by utilizing a graph-cut algorithm, where the weight matrix is created from the constraints' information matrices estimated by the front-end system. Our method provides the optimal window size to minimize information loss and linearization error. Moreover, due to the optimal SW size, our method produces the additional advantage of constructing an efficient hierarchical structure. To make a high-level graph, we create a high-level node (local map) by immobilizing the truncated part from the SW. The local maps can be efficiently matched in the front-end system to estimate the constraints between the high-level nodes. Therefore, our approach increases localization accuracy. We tested our algorithm on the indoor dataset obtained in an apartment environment to demonstrate the effectiveness of the proposed method. When our approach was applied to the hierarchical PGB SLAM back-end, we efficiently improved both localization accuracy (by reducing the information loss) and computational efficiency simultaneously.