1
$\begingroup$

I am currently trying to register a pointcloud in time to find my change in position and heading at each timestep (High speed application). So this is essentially an implementation of SLAM. I am currently using ICP with an SVD rotation solver to try to find rotation and translation. This solution works with simulated pointclouds.

The issue is that reobservation of previous points is non-deterministic for the type of scanner I am using. So this makes neighbor matching between frames difficult.

Is there any prepossessing I can do to get better matches in the neighbor finding step? or are there other methods for pointcloud registration that are more robust to noise than SVD based ICP?

$\endgroup$
1
$\begingroup$

I'm not sure what sort of ICP did you used but a voxel based plane-to-point ICP is robust and works well even in an unstructured environment.

This or his previous paper describes the method. http://ieeexplore.ieee.org/document/6220900/

Or this one might be helpful. http://static.adrian-haarbach.de/mscthesis_adrian.pdf

$\endgroup$
  • $\begingroup$ Sorry I was not specific enough. I would like to accomplish this with lidar vision only without any IMU/gyro. Updated question. I have it working with an IMU but want to eliminate the IMU $\endgroup$ – Ian Campbell Moore Jan 6 '18 at 7:53
  • $\begingroup$ Actually, the voxel based plane-to-point ICP does not depend on odometry/IMU/Gyro. They are just an additional constraints which are just good to have for an initial guess of the trajectory. The point is to use PCA and voxelization for a point to plane ICP. $\endgroup$ – C.O Park Jan 8 '18 at 6:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.