Open3D - Crop Pointcloud avec Polygon Volume

Vous pouvez choisir n'importe quel axe comme axe_orthogonal. Par exemple, si vous choisissez Z, définissez votre polygone avec un ensemble de points avec Z = 0. Ensuite, définissez vos Z min et max, comme si vous extrudiez un volume en utilisant le polygone entre Z min et max. J'espère que cela vous aidera.

Cet article m'a aidé à aller assez loin pour recadrer un nuage de points dans les limites d'un cuboïde. J'ai également constamment rencontré geometry :: PointCloud avec 0 point en utilisant vol.crop_point_cloud (pcd) et je n'ai pas pu le faire fonctionner, mais j'ai trouvé une solution différente.

J'ai fini par faire référence à ce PR # 1218 pour utiliser un open3d.geometry. OrientedBoundingBox volume cuboïde pour rogner le nuage de points. Le code ci-dessous crée un parallélépipède «tuile» de 200 mx 200 m autour d'une position de départ, qui correspond à une pose de départ du véhicule de l'ego dans un nuage de points, et filtre les points qui se trouvent uniquement dans la tuile.

import json
import numpy as np
import open3d as o3d

CUBOID_EXTENT_METERS = 200

METERS_BELOW_START = 5
METERS_ABOVE_START = 30

def main():
  ## Point Cloud
  points = np.array([
    ## These points lie inside the cuboid
    [-2770.94365061042, 722.0595600050154, -20.004812609192445],
    [-2755.94365061042, 710.0595600050154, -20.004812609192445],
    [-2755.94365061042, 710.0595600050154, -15.004812609192445],

    ## These points lie outside the cuboid
    [-2755.94365061042 + CUBOID_EXTENT_METERS, 710.0595600050154, -15.004812609192445],
    [-2755.94365061042, 710.0595600050154 + CUBOID_EXTENT_METERS, -15.004812609192445],
  ]).reshape([-1, 3])

  point_cloud = o3d.geometry.PointCloud()
  point_cloud.points = o3d.utility.Vector3dVector(points)

  ## Start point here corresponds to an ego vehicle position start in a point cloud
  start_position = {'x': -2755.94365061042, 'y': 722.0595600050154, 'z': -20.004812609192445}
  cuboid_points = getCuboidPoints(start_position)

  points = o3d.utility.Vector3dVector(cuboid_points)
  oriented_bounding_box = o3d.geometry.OrientedBoundingBox.create_from_points(points)
  point_cloud_crop = point_cloud.crop(oriented_bounding_box)

  # View original point cloud with the cuboid, all 5 points present
  o3d.visualization.draw_geometries([point_cloud, oriented_bounding_box])

  # View cropped point cloud with the cuboid, only 3 points present
  o3d.visualization.draw_geometries([point_cloud_crop, oriented_bounding_box])

def getCuboidPoints(start_position):
  return np.array([
    # Vertices Polygon1
    [start_position['x'] + (CUBOID_EXTENT_METERS / 2), start_position['y'] + (CUBOID_EXTENT_METERS / 2), start_position['z'] + METERS_ABOVE_START], # face-topright
    [start_position['x'] - (CUBOID_EXTENT_METERS / 2), start_position['y'] + (CUBOID_EXTENT_METERS / 2), start_position['z'] + METERS_ABOVE_START], # face-topleft
    [start_position['x'] - (CUBOID_EXTENT_METERS / 2), start_position['y'] - (CUBOID_EXTENT_METERS / 2), start_position['z'] + METERS_ABOVE_START], # rear-topleft
    [start_position['x'] + (CUBOID_EXTENT_METERS / 2), start_position['y'] - (CUBOID_EXTENT_METERS / 2), start_position['z'] + METERS_ABOVE_START], # rear-topright

    # Vertices Polygon 2
    [start_position['x'] + (CUBOID_EXTENT_METERS / 2), start_position['y'] + (CUBOID_EXTENT_METERS / 2), start_position['z'] - METERS_BELOW_START],
    [start_position['x'] - (CUBOID_EXTENT_METERS / 2), start_position['y'] + (CUBOID_EXTENT_METERS / 2), start_position['z'] - METERS_BELOW_START],
    [start_position['x'] - (CUBOID_EXTENT_METERS / 2), start_position['y'] - (CUBOID_EXTENT_METERS / 2), start_position['z'] - METERS_BELOW_START],
    [start_position['x'] + (CUBOID_EXTENT_METERS / 2), start_position['y'] - (CUBOID_EXTENT_METERS / 2), start_position['z'] - METERS_BELOW_START],
  ]).astype("float64") 

if __name__ == '__main__':
  main()

Voici une version abrégée qui montre comment recadrer un nuage de points en utilisant un np.array de sommets:

"""
corners = [[ 5.31972845 -3.21384387  0.30217625]
 [ 5.34483288 -1.13804348  0.29917539]
 [ 7.69983939 -1.16651864  0.30329364]
 [ 7.67473496 -3.24231903  0.3062945 ]
 [ 5.31845904 -3.21276837  1.03551451]
 [ 5.34356348 -1.13696798  1.03251366]
 [ 7.69856999 -1.16544314  1.03663191]
 [ 7.67346556 -3.24124353  1.03963277]]
"""
corners = np.array(...)

# Convert the corners array to have type float64
bounding_polygon = corners.astype("float64")

# Create a SelectionPolygonVolume
vol = o3d.visualization.SelectionPolygonVolume()

# You need to specify what axis to orient the polygon to.
# I choose the "Y" axis. I made the max value the maximum Y of
# the polygon vertices and the min value the minimum Y of the
# polygon vertices.
vol.orthogonal_axis = "Y"
vol.axis_max = np.max(bounding_polygon[:, 1])
vol.axis_min = np.min(bounding_polygon[:, 1])

# Set all the Y values to 0 (they aren't needed since we specified what they
# should be using just vol.axis_max and vol.axis_min).
bounding_polygon[:, 1] = 0

# Convert the np.array to a Vector3dVector
vol.bounding_polygon = o3d.utility.Vector3dVector(bounding_polygon)

# Crop the point cloud using the Vector3dVector
cropped_pcd = vol.crop_point_cloud(pcd)

# Get a nice looking bounding box to display around the newly cropped point cloud
# (This part is optional and just for display purposes)
bounding_box = cropped_pcd.get_axis_aligned_bounding_box()
bounding_box.color = (1, 0, 0)

# Draw the newly cropped PCD and bounding box
o3d.visualization.draw_geometries([cropped_pcd, bounding_box],
                                  zoom=2,
                                  front=[5, -2, 0.5],
                                  lookat=[7.67473496, -3.24231903,  0.3062945],
                                  up=[1.0, 0.0, 0.0])

Exemple avant:

Exemple après (il s'agit du bloc bleu de points au centre du nuage):