I'm not sure of the "optimal" way, but there are several options for handling somewhat uneven terrain. Unfortunately, most of the approaches I know require more than just a single-layer lidar. Here are a couple in case it gives you ideas:
If you have an IMU, you can filter
the lidar returns by setting z
limits, much like you suggested.
Basically, with the tilt angle from
the IMU, you can calculate the height
of each return point and kick out the
ones that are too far from the 2D
plane. The actual z limits would
depend on your setup and environment.
Some SLAM packages have this
functionality built in.
Calculate
surface normals based on lidar data.
It's easiest with multi-layer lidars,
but the idea is to throw out points
that indicate a plane with a normal
that's mostly in the z direction
(floor or ceiling).
If you're stuck with just a single-layer lidar, you may be able to filter based on further assumptions about your environment/application, but that can be tricky. For example, how do you know (from lidar data alone) that you're looking at the ground and not a big box someone dropped right in front of your robot? One scenario:
- If you know that your robot will never tilt more than a particular amount, and nothing in the environment moves, you may be able to calculate a safe lidar range limit. The lidar will never see ground closer than X m because that would require it to tilt more than is possible. Then, you can just filter out data beyond X m. Again, this is similar to your original suggestion, but it's important to note the inherent assumptions.
As far as computation, these approaches aren't very expensive and shouldn't be a problem--depending on your platform, obviously. Go with the simplest filter that makes sense. If you explain more of your scenario, maybe there's a more clever solution. I should add that if the terrain is too uneven, 2D just won't be enough.
Seems rgb-d or stereo camera might be better suited for the task, in combination with imu. As I asume you have an imu as otherwise the pointcloud approach wouldnt work, you might just make a node that funnels the laserscans. Then subscribe to imu, and when the imu's tilt/pitch exceeds the threshold stop publishing the output scan and start publishing an empty/fake scan for the time being. Inputscan=outputscan if imu within limits of tilt/pitch, kind of idea. You might even want to filter the scan at this point, as to still publish very near points that are really obstacles inside a ditch. The big question is how mapping handles "missed"/empty scans no matter how you achieve them. Might require 3d mapping after all.
If you intend to operate in a 3D environment why use a 2D-Slam package? Remember, you can start with a 3D package and always get like a costmap (2D) but you cannot scale up (reliably). An imu will give you roll and pitch measurements which you can use to transform your data before adding it to the slam package