Point Cloud Plane Segmentation

This library can be used to extract planar patches from unorganized point clouds. It is a stripped-down version of @abnerrjo's Robust Statistics-based Plane Detection (rspd) work published as

AMC Araújo and MM Oliveira, “A robust statistics approach for plane detection in unorganized point clouds,” Pattern Recognition, 2020.

The original implementation and other details can be found on GitHub and their project page.

Look and Feel

When executed on a raw point cloud with 50k points, rspd finds 20 planar patches in approximately 120 ms (i9-7920X, 64GB).

Build and Run

This library and an example app can be built with the usual

$ mkdir build && cd build
$ cmake .. && make

Using the provided test data, you can run rspd on a point cloud with

$ cd build
$ ./main ../data/blue.pcd

Open3D Dependency

The master branch depends on Open3D. If you did not install Open3D system-wide, you will need to provide the CMAKE_PREFIX_PATH instead of the cmake .. command above

$ cmake -DCMAKE_PREFIX_PATH=/path/to/Open3D/build/install ..

In this case, Open3D was built from source after running the cmake command

# Open3D was built from source using
$ cmake -DCMAKE_INSTALL_PREFIX=$(pwd)/install ..

Parameters

Parameter	Default	Description
`mMinNormalDiff`	60 degrees	A patch's associated point normals are tested against the patch's estimated normal. If the lower bound of the spread of similarity scores (roughly: 3σ below median) is lower than this, the patch is considered not robust and is thrown out.
`mMaxDist`	75 degrees	A patch's associated points are scored on distance to plane. Smaller values encourage tighter distribution of points around the plane. Also encourages truly planar patches. See Fig 4 of paper.
`mOutlierRatio`	0.75	Maximum allowable ratio of outlier points in a given patch's associated point set.
`minNumPoints`	30	Do not subdivide an octree node with less than this number of points
`nrNeighbors` (normal estimation)	75	Number of neighbors used to estimate the normals of each point.
`nrNeighbors` (knn search)	75	Number of neighbors used when attempting to grow and merge patches. More neighbors will result in (potentially) fewer output patches at the expense of more runtime.
minimum length of longest patch edge	0.01 * longest dimension of point cloud	In the original code, this parameter is hardcoded. This value is used to detect false positives and may be a valuable knob to tune. Alternatively, area could be used to cull too small planes.

Differences with Original Implementation

This implementation only provides the rspd functionality for planar patch detection. Additionally, it removes the Qt dependency and uses modern CMake for the build system. The algorithm and default parameters are the same as in the original.

Although the master branch rspd library depends on Open3D, the rspd-original branch does not (the rspd library, though the app used to test / visualize does).

kinggreat24 / pointcloud-plane-segmentation Goto Github PK