Localization does not improve and messes up move_base trajectories

[prasheelBB8]

Setup:

• ROS Melodic, RTABMap v0.20.18
• Differential drive vehicle (Max velocity in x - 1m/s)
• 2 Astra cameras for RGBD input (30Hz each)
• robot_localization's ekf (wheel+IMU) for odom input. (30Hz)
• ORB as Vis,FeatureType and Kp,DetectorStrategy

• The maps generated with this configuration look something like this -

  Then, if and when needed I perform a post processing using the rtabmap-databaseViewer to upgrade the 2D occupancy grid.

Issue: Now, when trying to localize in this map (even with the post processing mentioned above), the localization is really poor. I have tried the parameters mentioned in issues #300, #350, #719, etc. The /rtabmap/localization_pose are too less and at times there is a mismatch between the local costmap and the /rtabmap/grid_map in runtime that the difference does not let move_base to complete its trajectory.

I would like to understand

• Is the number of loop closures in mapping affecting the localization?
• Is the speed of our vehicle causing a motion blur for the feature matching?

Added our launch file - git_rgbd_slam_dual_camera_localize_launch.txt

[matlabbe]

If local costmap is using odom frame and global costmap is using map, if there is a difference between both is that the robot drifted against the map, the robot did not relocalize for a while.

Loop closures can help to track re-localization hypotheses during localization-only mode. The Bayes filter propagates re-localization hypotheses to linked nodes of the current best hypothesis. If /localization_pose topic is not published often, it is because the robot doesn't localize often.

If you are using rolling shutter cameras, there would be likely motion blur in image, which is bad for localization. You can browse images in your database using rtabmap-databaseViewer and see for yourself.