How to Tune RTAB-Map for Multi-UAV Loop Closure Reliability?

[m-christen]

We are developing a multi-UAV mapping system where each drone builds its own SLAM map in real time, with occasional spatial overlap between flight paths. Our objective is to merge these maps into a single global RTAB-Map database, reprocess it, and recover accurate inter-session transforms while still detecting valid loop closures.

A central server collects simultaneous data streams from multiple drones, then merges them into the global database.
After merging, we reprocess the database to detect new loop closures and optimize the map.

Disabling all covariance validations (set to 0) allows inter-session loop closure detections but also introduces false positives.
Lowering covariance tolerances prevents to detect inter-session closures unless the sessions come from the exact same acquisition.

During reprocessing, which parameter settings can balance inter-session loop closure recall with robustness against false positives?

[matlabbe]

What you are describing is similar to https://github.com/MISTLab/DOOR-SLAM or more recently https://github.com/MISTLab/Swarm-SLAM

I am not sure how do you do the merging process, but if you would use rtabmap-process "map1.db;map2.db;map3.db;map3.db" output.db, you should not have to modify the covariances, unless they are poorly estimated. What you can do instead is to adjust RGBD/OptimizeMaxError parameters to accept more (increasing it) or accepting less (lowering it) loop closures. In multi-session mapping, I also suggest to use GTSAM optimizer (Optimizer/Strategy=2).

If you can share the output.db as computed above, I could be interested to look at issues.

[m-christen]

Thank you very much for the reply. I was using the default optimizer (TORO), and due to the high variance, rtabmap was automatically triggering a new map.
During reprocessing, it required the specific method optimizeBA() for Bundle Adjustment, which is not implemented for the TORO and GTSAM optimizers.

By disabling Bundle Adjustment and using GTSAM alone, the results seem to be much more consistent from an initial test.
I’ll run some validation sessions in the coming days to confirm.

Thanks again for the support!