GTSAM for a LiDAR odometry front-end?

[Charles Hamesse]

Hi all,

I'm looking to implement various multi-sensor SLAM systems (LiDAR, visual, inertial) and GTSAM seems like an excellent choice for the optimization part. As I browse through existing solutions, I remark that most current LiDAR(-inertial) systems use an error-state iterative Kalman filter (ESIKF) for odometry. I can imagine that the obvious solution if I want to create a LiDAR-inertial-visual system would be to take the output of the ESIKF and put that as a GTSAM pose observation factor, then include other factors (IMU preintegration and some visual factors like feature reprojection).

But then I have a couple of questions:
1. Is there a particular reason for which there doesn't seem to exist a GTSAM-based LiDAR odometry system? Most LiDAR odometry systems nowadays also use some sort of feature-based approach (see e.g. VoxelMap), so the scale of the problem would not be much different from the visual odometry systems.
2. If you know such a system, could you mention it here? I believe using some LiDAR point-to-plane registration factors or similar directly instead of pose factors as proposed earlier would make the global system more elegant. Also, many systems use GTSAM in the back-end for pose graph optimization, so it's already there - why not use it also in the front-end? Would there be any reason not to use GTSAM for a LiDAR odometry front-end, i.e. adapting FAST-LIO, VoxelMap, etc to use iSAM2 instead of the ESIKF? This could make the formulation more elegant and allow to implement extensions more easily with other sensors, but am I missing something?

Thank you very much.

Kind regards,

Charles