Reg/Strategy = 1 & NeighborLinkRefining in lidar challenging environment

[kabuto]

Hi Mathieu,

My setup consists of dual 2D LiDARs and 2 RGBD cameras. What I mean by 'challenging environment for LiDAR' is that, due to uneven ground, the point cloud from the 2D LiDAR can be very inconsistent. Even small changes in ground angle can cause significant variations in the 2D LiDAR's point cloud.

I'm currently using Reg/Strategy = 1 along with NeighborLinkRefining = True, based on my understanding. In the environment I've mentioned above, NeighborLinkRefining can barely help because it's very hard to perform ICP, so it provides a bad initial guess for Reg/Strategy = 1, which only uses wheel odometry information.

I can see from the log that the CorrRatio is very low; it can range from 0.04 to 0.07. I've tried decreasing the ICP/CorrespondenceRatio to 0.04, but I think it's not a very good solution because it might cause incorrect registrations for ICP.

Is the VisRegistration also working in this scenario?

Could you please provide some help for my case? I appreciate it!

Best,
Kabuto

[matlabbe]

Hi,

How high over the ground are the lidars, and what are their range? One approach could be to reduce the lidar range used by rtabmap to a range that would not be affected by the ground (bad points), e.g, Icp/RangeMax.

Looks like ICP odometry would be difficult in that scenario, is your wheel odometry relatively accurate? You could do loop closures only with cameras too (Reg/Strategy=0) and just ignore the lidar.

cheers,
Mathieu