Algorithm or Theory of RTAB-MAP's Localization

[shiba_slam]

Hi,

I searched the intro lab's page and i found some papers of RTAB-MAP.
There are many talking about memory management.

But i want to know how does RTAB-MAP localization?
I know that loop closure using "Bag of word", and i want to learn something about localization.
Is it the same using loop closure(Bag of word)?
And i also want to know that whether "Lidar" help in localization or not.

Please give me some keywords of the localization method then i'll read the papers.
Thanks in advance!!

[matlabbe]

Hi,

Look at this paper (Section 3.4):

Loop closure detection is done using the bag-of-words approach described in [Labbé and Michaud,  2013]. Basically,  when  creating  a  new  node,  STM  extracts  visual  features  from  the  RGB  image  and  quantizes them to an incremental visual word vocabulary.
[...]
When a loop closure hypothesis reaches the fixed threshold “Rtabmap/LoopThr”,a loop closure is detected and transformation is computed.  The transformation is computed using the same Motion Estimation approach used by visual odometry (Section 3.1.1), and if accepted, the new link is added to the graph.  When a laser scan or a point cloud is available, link’s transformation is refined using the same ICP Registration approach than with lidar odometry (described in Section 3.1.2).

In localization mode, we do exactly the same as loop closure detection, but without adding a new constraint to graph, just correcting the current pose.

cheers,
Mathieu