Re: Mapping with multiple stereocameras, localization with multiple monochrome monocular cameras

Posted by matlabbe on
URL: http://official-rtab-map-forum.206.s1.nabble.com/Mapping-with-multiple-stereocameras-localization-with-multiple-monochrome-monocular-cameras-tp6647p6959.html

This makes me remember that we had a similar problem with a setup with a robot having a forward and a backward cameras. When the robot was traversing the same path but in inverted direction, transforms were rejected. We removed the check when having more than one camera: https://github.com/introlab/rtabmap/commit/bd979cca34f9d5ac0bc41c95230f5bcde59fde54

I think we can remove safely that check for good. It was added in January 2016 because of ICP problem described above. However, normally there was a check in RegistrationIcp to detect this problem, but was refactored on May 2016 to do correctly compare with the original guess transform. For example with that refactoring, if the visual guess tells that the transform is 180 deg (like in your case), ICP won't fail if it tells that it is 182 deg instead, but will fail if it tells 2 deg (it compares the change between input transform and the output, not the actual absolute transform angle). I removed the check in this commit: https://github.com/introlab/rtabmap/commit/c5158cade5f9b6f845c3df507c3032916d20ca20

frame_id should be the base frame indeed, not back_camera, with that commit the loop should now be accepted with your backward camera.

cheers,
Mathieu