[common/search] 2D versions of `pcl::getPointsInBox` and `cropBoxFilter`

[roncapat]

Is your feature request related to a problem? Please describe.

For some operations, e.g. 2D map creation from 3D point cloud, it is useful to search for points in certain XY range, with unconstrained Z.

Context

Robotic mapping, cloud points belonging to the same "column of air" above/below a 2D region (e.g. cell of a 2D grid/map).

Goal

Potential speedups on slow CPUs.

Expected behavior

Ability to set lower/upper bounds only for X and Y, skipping comparisons of Z values in pcl::getPointsInBox and cropBoxFilter (at most 1/3 of the numerical comparisons likely to be skipped).

Current Behavior

User is forced to pass min/max allowed Z value.

Additional Notes
Ideally the same way of reasoning would be applicable to search/indexing methods like Octree (--> only a Quadtree would be needed for such applications).

[larshg]

Maybe you can use the conditional filter and set up your conditions for x and y.

Alternatively you can use the FunctorFilter and add a lambda that meets your filtering needs.
Its seemingly up to 10 times faster, see here.

Its under advanced and experimental, but works quite well. There are one perhaps more PRs for improving it further.

[roncapat]

Thank you!

Here's my experience up until now, when looping over 2D cells (in X and Y direction) to partition a Point Cloud (cumulative times):

1. cropBoxFilter -> ~1000ms
2. getPointsInBox -> ~300ms
3. OctreePointCloudSearch -> ~30ms

as a rought quantification of the progresses I made. @larshg I will surely try your methods (that I didn't know about until now) and if I manage to shave off some additional performance I'll report here.

[mvieth]

@roncapat If you just want to find all points in a 2D box once (one box), then Lars' recommendations are exactly the right filters to use (especially FunctorFilter). However (and I am not sure if I understand your goal correctly), if you want to put all points in a 2D grid (divide the whole 2D space into non-intersecting boxes), then calling the CropBox/ConditionalRemoval/FunctorFilter/getPointsInBox many times in a loop is inefficient. In this case, you want to compute the grid once instead of looping over all points again and again. In 3D, VoxelGrid or an Octree would be fitting for this. For the 2D case, I am not sure if PCL has something usable out-of-the-box. GridMinimum uses a 2D grid, so you could have a look at grid_minimum.hpp