Re: Improving a lidar-only setup

Posted by Robin on
URL: http://official-rtab-map-forum.206.s1.nabble.com/Improving-a-lidar-only-setup-tp11071p11373.html

Hi,

Sorry for the late response. I finally had time to dive into this again.
- The fix for the clean shutdown indeed works, thx!
- Our cameras are indeed behind some glass, so the emitter is creating reflections. So we'll just stick with no emitter and infra+depth.

Btw, in case this is helpful info for you, for getting to localize far from the mapped paths, I adjusted these parameters compared to the defaults in the lidar3d.launch.py example:
- RGBD/LocalRadius: 30.0
- RGBD/ProximityPathMaxNeighbors: 2
- Icp/CorrespondenceRatio: 0.1
Because the livox 360 has a large range and fov, the localization remains pretty accurate even when far from any mapped location.

Currently we're trying to get the most out of rtabmap in terms of repeatability. Do you have any suggestions which parameters to focus on? I know this is not a trivial question and depends on quite a few factors.

When localizing in pretty much the exact same scenario (same start/end location and same speed profile), we get roughly +-5mm deviation at most. When we change the start location and/or the speed profile though, this becomes something like +-15mm deviation. Our ideal goal would be +-3mm in all cases where the last 2 meters of movement can be at slower speeds.
Some notes about the use case:
- the above tests are movements to/from for a distance of about 5 meters, with varying turns and speeds and always containing a quarter pivot + 0.5m linear movement near the end.
- rotation speeds of the vehicle go up to at most 1.2rad/s, linear movement speeds go up to 2m/s
- 90% of the pointcloud features are ceiling/walls that are 20m+ away

As said in my previous posts here, we've already ensured very accurate odometry calibration, lidar deskewing, static tf for lidar and timestamp sync of the sensor.
Aside from parameter trial/error and tuning, we were thinking of maybe incorporating markers for critical locations as well, but I'm not sure if that's going to give more trouble than benefit. If the marker is obscured, I would assume that affects repeatability for example.

Any guidance on which parameters to focus on or anything else is greatly appreciated!