search

introlab / rtabmap_ros

RTAB-Map's ROS package.
http://wiki.ros.org/rtabmap_ros
BSD 3-Clause "New" or "Revised" License
912 stars 549 forks source link

Dense Point Cloud #355

Open caiofis opened 4 years ago

caiofis commented 4 years ago

I am trying to get a more dense point cloud at the output, I already changed all the parameters to make it denser, but with no results. It seems that is have some kind of voxel filtering inside even if I set it to minimal. Is there any way of remove it to get the denser point cloud possible?

cheers, Caio

matlabbe commented 4 years ago

Do you mean cloud_map output? If so, it is indeed downsampled to 5 cm voxel, with parameter Grid/VoxelSize.

Note that producing online more denser clouds may be not optimal. Why you want a dense cloud online? You can export a full resolution cloud offline using exporting tool of rtabmap. Just open the database in rtabmap standalone app, then do "File->Export 3D clouds...", check "Regenerate clouds", and set the density you want.

cheers, Mathieu

caiofis commented 4 years ago

Yeah, I mean cloud_map. I need to perform some 3D segmentation, for now I am using the raw data from Kinect, which is pretty noisy. My idea was to use the RTABMap as a registration tool, maybe using the passthrough filter to publish only a fraction of the dense point cloud. Do you think I should use other tool since I just need a reconstruction with local consistency? I tried to perform ICP with the measurements from Kinect, but my output is kinda of noisy, how do you remove the outliers to merge the point clouds with ICP? I am diving into your code here, but it is quite big. ;)

matlabbe commented 4 years ago

You can decrease Grid/VoxelSize to 1 cm voxel for example to make it a little denser. But as you need only local dense clouds, you may subscribe to odometry and the point cloud of the kinect, then assemble them yourself at the resolution you want. You can also look at the point_cloud_assembler nodelet of rtabmap_ros or the one from http://wiki.ros.org/laser_assembler that do such stuff.

By default, rtabmap does visual registration, not icp with the kinect-like data.