Indoor topological mapping with place recognition and scan matching

Map building is one of the key tasks of autonomous mobile robots’ navigation. Traditional mapping methods build dense metric map (e.g. as an occupancy grid). Maintaining such map in case of long-term navigation is difficult because of high computational costs and odometry error accumulation. Representing the environment as a sparse topological structure (e.g. a graph of locations) lets us eliminate these drawbacks and provide fast path planning. In this work, we propose a topological mapping method which builds and updates a graph of locations without use of global metric coordinates. For localization, the proposed method uses neural network-based place recognition in pair with 2D projection-based scan matching. We carry out experiments with our method in several photorealistic simulated scenes and on data from a real robot. In simulation, we compare our method with some state-of-the-art topological mapping methods. According to the results, the proposed method significantly outperforms competitors in terms of navigational efficiency, keeping graph connectivity, high scene coverage, and low part of inconsistent edges.