Lidar + RGB

[Fortis]

Hello guys,

is it possible to use some different brands of Lidar scanners, such as LiVOX (Mid-40 or Horizon) with RTAB and use it together with RGB camera (which types are supported?).

Thanks.

[matlabbe]

Hi,

It depends what you want to do, but this combination is possible. I have more examples with lidars with large field of view though (like velodyne or ouster). I don't have experience with Livox sensors. It looks like we can download lvx files on their website, and convert them in rosbag with their livox_ros_driver. When I will have some free time, I'll check how it could work with rtabmap.

cheers,
Mathieu

[feta]

Hi,

I've been experimenting recently with Livox .bags and running them concurrently with RTAB-Viz so I wanted to chime in, so far what I've tried is accessing pointcloud data from the .bag (Livox didn't provide RGB data in this particular .bag).



This is basically just using scan_cloud data with ICP and without loop closure (since there are no images), so it doesn't mean much. I'd be interested in seeing how it performs with an actual Livox lidar since their acquisition method is different compared to most lidars. This blog mentions having issues with HDL Graph Slam due to the unique scanning pattern.

Edit: Also my attempt was using just ICP without odometry data (which is available with Livox Horizon) which breaks after a while. I'm having some trouble pinning down the exact topic for "odom" data.

[matlabbe]

Default rtabmap parameters would not work very robustly with this lidar. We would need to make a launch file example with parameters tuned for that kind of sensor. It could possible to accumulate a little more scans to help to get more overlap between consecutive clouds for ICP.

There is also LOAM-Livox: https://github.com/hku-mars/loam_livox with some rosbags.

cheers,
Mathieu

[tteresi7]

Hi I've been trying to run the rosbags you mentioned at the Livox LOAM repo.  These, as mentioned, only contain sensor_msgs/PointCloud2.

I've been trying to figure out a way to to run these in RTabMap, but I keep running into issues. (Without loop closure of course).  I've tried remapping scan_cloud to /livox/lidar, but when I run the rosbag nothing appears to happen.  I don't think I have a properly configured launch file, so I was curious how you would suggest to configure a launch file for this?

[feta]

If it's any help here's the launch file that I used, it's been modified from a file intended for Velodyne Puck mapping. This is the mapping package that I used (I ran the parking lot rosbag recorded via Livox Horizon through "mapping_horizon.launch"). Note that this isn't perfect and ICP will probably break at some point during mapping, there are probably some parameters that could be tweaked to prevent this.

[Mikor]

Hello tteresi7, check this configuration it is using only a topic containing sensor_msgs/Pointcloud2.

[tteresi7]

Mikor wrote

Hello tteresi7, check this configuration it is using only a topic containing sensor_msgs/Pointcloud2.

I attempted this and I get this repeatedly.

"[ERROR] (2021-03-10 02:06:06.472) RegistrationIcp.cpp:1473::computeTransformationImpl() RegistrationIcp cannot do registration with a null guess.
[ WARN] (2021-03-10 02:06:06.472) OdometryF2M.cpp:557::computeTransform() Registration failed: "RegistrationIcp cannot do registration with a null guess."
[ INFO] [1615341966.474929925]: Odom: ratio=0.000000, std dev=0.000000m|0.000000rad, update time=0.002614s"

And the visualization just refreshes to show the current point cloud.  I'm guessing this config file is also looking for a transformation or odom as well?

[Mikor]

It seems like the odometry drifts too much, if it is done only by lidar it can not be recovered.

[matlabbe]

I can say that I also tried Livox point clouds, and I don't find a way to make it work with rtabmap/libpoint matcher. The butterfly pattern from Livox Lidar doesn't create a lot of overlap between consecutive scans, which make it very challenging for the ICP approach used in RTAB-Map. IF the Livox mapper could be used as odometry with local assembled point clouds, they could be fed to rtabmap with a camera to find loop closures.