This function is meant to be used inside stemPoints. It filters points based on their nearest neighborhood geometries (check fastPointMetrics) and assign them to stem patches if reaching a voxel with enough votes.
stm.eigen.knn(
h_step = 0.5,
max_curvature = 0.1,
max_verticality = 10,
voxel_spacing = 0.025,
max_d = 0.5,
votes_weight = 0.2,
v3d = FALSE
)numeric - height interval to perform point filtering/assignment/classification.
numeric - maximum curvature (from 0 to 1) accepted when filtering a point neighborhood.
numeric - maximum deviation of a point neighborhood's orientation from an absolute vertical axis ( Z = c(0,0,1) ), in degrees (from 0 to 90).
numeric - voxel (or pixel) spacing for counting point normals intersections.
numeric - largest tree diameter expected in the point cloud.
numeric - fraction of votes a point neighborhood needs do reach in order to belong to a stem (applied for every TreeID), in relation to the voxel with most votes with same TreeID.
logical - count votes in 3D voxels (TRUE) or 2D pixels (FALSE).
Point filtering/classification methods that rely on eigen decomposition rely on shape indices calculated for point neighborhoods (knn or voxel). To derive these shape indices, eigen decomposition is performed on the XYZ columns of a point cloud patch. Metrics related to object curvature are calculated upon ratios of the resulting eigen values, and metrics related to object orientation are caltulated from approximate normals obtained from the eigen vectors.
For instance, a point neighborhood that belongs to a perfect flat surface will have all of its variance explained by the first two eigen values, and none explained by the third eigen value. The 'normal' of such surface, i.e. the vector oriented in the direction orthogonal to the surface, is therefore represented by the third eigenvector.
Methods for both tree mapping and stem segmentation use those metrics, so in order
to speed up the workflow one might apply fastPointMetrics to the point
cloud before other methods. The advantages of this approach are that users
can parameterize the point neighborhoods themselves when calculating their metrics.
Those calculations won't be performed again internally in the tree mapping or stem
denoising methods, reducing the overall processing time.
stemPoints methods that filter points based on eigen decomposition
metrics (knn or voxel) provide a rough estimation of stem segments radii
by splitting every stem segment into a local voxel space and counting
the number of times that point normals intersect on every voxel (votes).
Every stem point then has a radius assigned to it, corresponding to the distance
between the point and the voxel with most votes its normal intersects. The average of
all points' radii in a stem segment is the segment's radius. For approximately straight
vertical stem segments, the voting can be done in 2D (pixels).
The point normals of this method are extracted from the eigen vectors calculated by
fastPointMetrics. On top of the point metrics used for stem point filtering, the following
fields are also added to the LAS object:
Votes: number of normal vector intersections crossing the point's normal at its estimated center
VotesWeight: ratio of (votes count) / (highest votes count) per TreeID
Radius: estimated stem segment radius
This method was inspired by the denoising algorithm developed by Olofsson & Holmgren (2016), but it is not an exact reproduction of their work.
Liang, X. et al., 2012. Automatic stem mapping using single-scan terrestrial laser scanning. IEEE Transactions on Geoscience and Remote Sensing, 50(2), pp.661<U+2013>670.
Olofsson, K. & Holmgren, J., 2016. Single Tree Stem Profile Detection Using Terrestrial Laser Scanner Data, Flatness Saliency Features and Curvature Properties. Forests, 7, 207.