How to get transformation fast?

[longchao]

Dear Sir:
   I want to write a real-time slam .But after I tested ICP and G-ICP to get relative transformation between two poses,I found it is too slow for real-time task.Can you give me some instructions?

[matlabbe]

I've seen approaches doing real-time (>30 Hz) odometry using geometric (ICP) and/or visual (features tracking) approaches. There some advantages/disadvantages for both methods, it depends on your application. There is maybe a paper somewhere doing a review of some approaches :P.

[longchao]

THX for your quick reply.Can you give me some examples for using ICP(>30HZ)? I have read the source code of ccny ros package,I found they use SVD instead of ICP for realtime pose calculation in odometry.I think it is not approprite to use SVD in my slam

[matlabbe]

Geometric:
 * ethzasl_icp_mapping, pretty cool video using only ICP in dynamic environments
 * hector_slam

Visual:
 * fovis
 * viso2
 * ptam

[longchao]

Thx for your help,libpointmatcher is really a great lib.I will try it.My point cloud are stored in .pcd format,is it possible to achive this using PCL in a RGBD-SLAM?

[matlabbe]

Don't know if there is a bridge between PCL format and the one used in libpointmatcher, but for basic ICP, you can look at this PCL tutorial. Here an issue with ICP using PCL. Maybe the libpointmatcher library is more robust to this problem. Combining an IMU with ICP could help too! Tutorials for libpointmatcher here.