open3d-点云及其操作

open3d提供了一个专门用于点云的数据结构 PointCloud。

class PointCloud(Geometry3D):
    color   # 颜色
    normals # 法向量
    points  # 点云
    def __init__(self, *args, **kwargs):
        """
        __init__(*args, **kwargs)
        Overloaded function.
        1. __init__(self: open3d.cpu.pybind.geometry.PointCloud) -> None
        Default constructor
        2. __init__(self: open3d.cpu.pybind.geometry.PointCloud, arg0: open3d.cpu.pybind.geometry.PointCloud) -> None
        Copy constructor
        3. __init__(self: open3d.cpu.pybind.geometry.PointCloud, points: open3d.cpu.pybind.utility.Vector3dVector) -> None
        Create a PointCloud from points
        """
    # dbscan聚类
    def cluster_dbscan(self, eps, min_points, print_progress=False):
    # 计算凸包
    def compute_convex_hull(self):
    # 计算马氏距离。 返回每个点的马氏距离
    def compute_mahalanobis_distance(self):
    # 计算均值与协方差矩阵
    def compute_mean_and_covariance(self): 
    # 计算点云每个点到其最近点的距离
    def compute_nearest_neighbor_distance(self):
    # 计算当前点云每个点到目标点云的最近距离
    def compute_point_cloud_distance(self, target):
    def create_from_depth_image(self, depth, intrinsic, extrinsic, *args, **kwargs):
    def create_from_rgbd_image(self, image, intrinsic, extrinsic, *args, **kwargs):
    # 裁剪。 输入一个aabb框或obb框
    def crop(self, *args, **kwargs):
    # 计算顶点法向量
    def estimate_normals(self, search_param=None, *args, **kwargs):
    # 是否有color
    def has_colors(self):
    # 是否有法向量
    def has_normals(self):
    # 是否有点云点
    def has_points(self):    
    # 隐藏点去除。 
    def hidden_point_removal(self, camera_location, radius):
    # 归一化法向量。 法向量长度为1
    def normalize_normals(self):
    # 法向量方向一致
    def orient_normals_consistent_tangent_plane(self, k):
    # 法向量方向一致。 指定相机位置
    def orient_normals_towards_camera_location(self, camera_location=None, *args, **kwargs):
    # 法向量方向一致。 指定参考方向
    def orient_normals_to_align_with_direction(self, orientation_reference=None, *args, **kwargs):
    # 上色。 颜色rgb，范围0~1
    def paint_uniform_color(self, color):
    # 随机下采样。 指定下采样率
    def random_down_sample(self, sampling_ratio):
    # 删除non 和 inf 值的点
    def remove_non_finite_points(self, remove_nan=True, remove_infinite=True):   
    # 删除指定半径内少于指定点数的点
    def remove_radius_outlier(self, nb_points, radius):
    # 删除相邻点中距离大于平均距离的点
    def remove_statistical_outlier(self, nb_neighbors, std_ratio):
    # 平面分割
    def segment_plane(self, distance_threshold, ransac_n, num_iterations):
    # 按照下标筛选点云
    def select_by_index(self, indices, invert=False):
    # 下采样。 每隔多少个点取一个
    def uniform_down_sample(self, every_k_points):
    # 体素下采样。 指定体素尺寸
    def voxel_down_sample(self, voxel_size):
    # 体素下采样并记录原数据。 指定体素尺寸
    def voxel_down_sample_and_trace(self, voxel_size, min_bound, max_bound, approximate_class=False): 
    

import numpy as np
import open3d as o3d
from open3d.web_visualizer import draw
from open3d.visualization import draw_geometries

pcd = o3d.io.read_point_cloud('datas/fragment.ply')
draw(pcd)

1.体素下采样

open3d.geometry.PointCloud 提供了 voxel_down_sample(self, voxel_size) 方法，来进行体素下采样操作。

    def voxel_down_sample(self, voxel_size):
        """
        voxel_down_sample(self, voxel_size)
        对输入点云进行体素下采样，如果法线和颜色存在，则法线和颜色取均值。

        Args:
            voxel_size (float): 体素尺寸

        Returns:
            open3d.geometry.PointCloud
        """

downsample = pcd.voxel_down_sample(voxel_size=0.05)
draw(downsample)

2. 顶点法向量估计

open3d.geometry.PointCloud 提供了 estimate_normals(self, voxel_size) 方法，来计算顶点法向量。

    def estimate_normals(self, search_param=None, *args, **kwargs): 
        """
        estimate_normals(self, search_param=KDTreeSearchParamKNN with knn = 30, fast_normal_computation=True)        
        Args:
            search_param (open3d.geometry.KDTreeSearchParam, optional): 用于领域搜索的KDTree搜索参数。 默认值为：KDTreeSearchParamKNN with knn = 30
            fast_normal_computation (bool, optional, default=True): 如果为true，通过协方差矩阵计算特征向量，速度更快，但数值不稳定。如果为False,则使用迭代方式。

        Returns:
            None   无返回值，法向量直接存储于 PointCloud.normals 
        """

search_param 参数有:

• class KDTreeSearchParamHybrid(KDTreeSearchParam):
  def init(self, radius, max_nn): # 搜索半径、最大近邻点数
• class KDTreeSearchParamKNN(KDTreeSearchParam):
  def init(self, knn=30): # 近邻点数
• class KDTreeSearchParamRadius(KDTreeSearchParam):
  def init(self, radius): # 搜索半径

downsample.estimate_normals(search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1, max_nn=30))

# 此处使用 draw_geometries绘制点云以及法线。
draw_geometries([downsample], point_show_normal=True)

3. 裁剪点云

裁剪点云，首先需要确定裁剪区域

通过o3d.visualization.read_selection_polygon_volume()函数，读取一个多边形区域。

然后通过多边形裁剪点云。

def read_selection_polygon_volume(filename): 
    """
    read_selection_polygon_volume(filename)
    Function to read SelectionPolygonVolume from file

    Args:
        filename (str): The file path.

    Returns:
        open3d.visualization.SelectionPolygonVolume
    """
    pass

open3d.visualization.SelectionPolygonVolume 含有两个方法：

• crop_point_cloud(input)
• crop_triangle_mesh(input)

# 读取多边形
vol = o3d.visualization.read_selection_polygon_volume('datas/cropped.json')
chair = vol.crop_point_cloud(pcd)
draw(chair)

4. 点云上色

open3d.geometry.PointCloud 提供了 paint_uniform_color(self, color)方法，来为点云进行上色。

def paint_uniform_color(self, color): 
    """
    paint_uniform_color(self, color)
    Assigns each point in the PointCloud the same color.

    Args:
        color (numpy.ndarray[float64[3, 1]]):RGB颜色，值在0~1范围内

    Returns:
        open3d.geometry.PointCloud
    """
    pass

chair.paint_uniform_color([1,0,0])  # 红色
draw(chair)

5. 点云距离与筛选

open3d.geometry.PointCloud 提供了 compute_point_cloud_distance(self, target)方法，计算当前点云中每个点到目标点云中点的最近距离。

def compute_point_cloud_distance(self, target):
    """        
    Args:
        target (open3d.geometry.PointCloud): 目标点云

    Returns:
        open3d.utility.DoubleVector
    """

open3d.geometry.PointCloud 提供了 select_by_index(self, indices, invert=False)方法，通过下标来筛选点云。

def select_by_index(self, indices, invert=False):
    """
    select_by_index(self, indices, invert=False)        
    Args:
        indices (List[int]): 下标
        invert (bool, optional, default=False): 反选

    Returns:
        open3d.geometry.PointCloud
    """

dists = pcd.compute_point_cloud_distance(chair)  # 计算整体点云中，每个点到椅子点云中最近点的距离。
dists = np.array(dists)
ind = np.where(dists > 0.01)[0]  # 获取距离大于0.01的点的下标
pcd_without_chair = pcd.select_by_index(ind)  # 通过下标筛选点云中点
draw(pcd_without_chair)

6. 边界框

o3d.geometry.Geometry3D 提供了 get_axis_aligned_bounding_box() 方法，来获取aabb包围盒（轴对齐包围盒）

def get_axis_aligned_bounding_box(self):
    """
    get_axis_aligned_bounding_box(self)        
    Returns:
        open3d.geometry.AxisAlignedBoundingBox
    """

o3d.geometry.Geometry3D 提供了 get_oriented_bounding_box() 方法，来获取obb包围盒（有向包围盒）

def get_oriented_bounding_box(self):
    """
    Returns:
        open3d.geometry.OrientedBoundingBox
    """

aabb = chair.get_axis_aligned_bounding_box()
print(aabb)
draw([chair, aabb])

obb = chair.get_oriented_bounding_box()
print(obb)
draw([chair, obb])

7.凸包

o3d.geometry.Geometry3D 提供了 compute_convex_hull() 方法，来获取点云的凸包。

def compute_convex_hull(self):
    """
    Returns:
        Tuple[open3d.geometry.TriangleMesh, List[int]] 返回两个值，第一个以三角形网格返回凸包，第二个返回凸包的顶点下标。
    """

hull, ind = chair.compute_convex_hull()

hull.paint_uniform_color([1,0,0])
draw([hull, chair])

chair.paint_uniform_color([0.5,0.5,0.5])
points = chair.select_by_index(ind)  # 红色点为凸包顶点
points.paint_uniform_color([1,0,0])
draw([chair, points])

8. dbscan聚类

open3d.geometry.PointCloud 提供了 cluster_dbscan(self, eps, min_points, print_progress=False) 方法，实现dbscan密度聚类。

def cluster_dbscan(self, eps, min_points, print_progress=False):
    """
    cluster_dbscan(self, eps, min_points, print_progress=False)        
    Args:
        eps (float):密度
        min_points (int): 形成簇的最小点数
        print_progress (bool, optional, default=False): 
    Returns:
        open3d.utility.IntVector label值，int
    """

from matplotlib import pyplot as plt
labels = pcd.cluster_dbscan(eps=0.02, min_points=10, print_progress=True)
labels = np.asarray(labels)
max_label = labels.max()
colors = plt.get_cmap("tab20")(labels / (max_label if max_label > 0 else 1))
colors[labels < 0] = 0
pcd.colors = o3d.utility.Vector3dVector(colors[:, :3])

draw(pcd)

9. 平面分割

open3d.geometry.PointCloud 提供了 **segment_plane(self, distance_threshold, ransac_n, num_iterations)** 方法，通过RANSAC从点云中分割平面。

def segment_plane(self, distance_threshold, ransac_n, num_iterations):
    """        
    Args:
        distance_threshold (float): 点到面的最大距离
        ransac_n (int): 随机采样估计平面的点数
        num_iterations (int): 迭代次数

    Returns:
        Tuple[numpy.ndarray[float64[4, 1]], List[int]]
        """

pcd = o3d.io.read_point_cloud('datas/fragment.ply')
plane_model, ind = pcd.segment_plane(distance_threshold=0.01, ransac_n=3, num_iterations=1000)

plane = pcd.select_by_index(ind)
plane.paint_uniform_color([1,0,0])
without_plane = pcd.select_by_index(ind, True)
draw([plane, without_plane])

10. 隐藏点去除

open3d.geometry.PointCloud 提供了 hidden_point_removal(self, camera_location, radius) 方法。

def hidden_point_removal(self, camera_location, radius):
    """        
    Args:
        camera_location (numpy.ndarray[float64[3, 1]]): All points not visible from that location will be reomved
        radius (float): The radius of the sperical projection

    Returns:
        Tuple[open3d.geometry.TriangleMesh, List[int]]
    """

diameter = np.linalg.norm(np.asarray(pcd.get_max_bound()) - np.asarray(pcd.get_min_bound()))
camera = [0, 0, diameter]
radius = diameter * 100
_, pt_map = pcd.hidden_point_removal(camera, radius)

pcd = pcd.select_by_index(pt_map)
draw(pcd)