Online Global Loop Closure Detection for Large-Scale Multi-Session Graph-Based SLAM

TL;DR

The paper tackles large-scale, long-term SLAM by coupling online global loop closure detection with a memory-managed, graph-based SLAM framework to handle multi-session scenarios and the kidnapped-robot problem. It combines appearance-based loop closure (bag-of-words with SURF features) and geometry-based optimization (TORO) within a STM/WM/LTM memory hierarchy to bound online computation while enabling inter-session constraints. Experimental results on five indoor sessions with a Kinect and laser scanner show online processing within a fixed time budget and eventual merging of all sessions into a coherent global map, albeit with tradeoffs in map completeness due to memory constraints. The work yields a scalable, open-source solution for large-scale SLAM and points to future directions in exploration strategies to improve online performance and map robustness.

Abstract

For large-scale and long-term simultaneous localization and mapping (SLAM), a robot has to deal with unknown initial positioning caused by either the kidnapped robot problem or multi-session mapping. This paper addresses these problems by tying the SLAM system with a global loop closure detection approach, which intrinsically handles these situations. However, online processing for global loop closure detection approaches is generally influenced by the size of the environment. The proposed graph-based SLAM system uses a memory management approach that only consider portions of the map to satisfy online processing requirements. The approach is tested and demonstrated using five indoor mapping sessions of a building using a robot equipped with a laser rangefinder and a Kinect.

Figures (12)

• Figure 1: Memory management model.
• Figure 2: Illustration of a local map created from multi-session mapping.
• Figure 3: AZIMUT-3 robot equipped with a URG-04XL laser range finder and a Kinect sensor.
• Figure 4: Resulting local maps without (left) and with (right) graph optimizations for a) Map 1, b) Map 2 and c) Map 3. Loop closures are shown in red.
• Figure 5: Results for Map 4 (top) and Map 5 (bottom), with a) the map from all nodes still in WM (light gray) with the global graph (blue line) not optimized, b) the local map with local graph optimization and c) the global map with global graph optimization. Loop closures are shown in red.