无组织点云的连接组件的有效实现

Question

我有大约 100 万个点的无组织点云数据。 单个点云的形状是 [npoints, 3]，其中 npoints 个点和 3 个坐标分别为 x、y 和 z。 我想 label 点云中的连接组件，其中 0.05 欧几里德距离内的点被认为是连接的。 我在 Python 中使用深度优先搜索。 我使用了循环和 numpy 来计算欧几里得距离。 完成处理需要数小时。 我正在努力寻找一种更有效的方法来实现这一目标。

由于我的实际目标，我不想使用内置函数。 实际上在我的点云中，每个点都标有一个数字，我想根据距离以及 label 连接这些点。

现在我正在使用 DFS，并使用颜色 RGB 分配连接的组件，因此点云的维度是 [npoints,7]，其中列代表 x、y、z、label、r、g、b。 未连接的点被分配 0 RGB 值。

这是供参考的代码。

import numpy as np
import os
import random

### Read file
def read_txt_file(fn):
  print("******** Reading file:", fn)
  point_list = []
  f = open(fn, "r")
  for x in f:
    point_list.append([float(i) for i in x[:-2].split(' ')])

  # point_list = np.array(point_list, dtype=np.float)
  print("Point cloud read. len:",len(point_list))
  return point_list

data_pts = read_txt_file('example.txt')

random.seed(10)

point_a = np.array(data_pts) #[1:,:]
print("initial point_a.shape",point_a.shape)

# npoints= 100000
# choice = np.random.choice(point_a.shape[0], npoints, replace=True)
# point_a = point_a[choice, :]
# print(f'after random downsampling to {npoints} points, shape: {point_a.shape} ')

point_a = np.hstack((point_a, np.zeros((point_a.shape[0],3))))
print("after adding zero columns for red green blue point_a.shape",point_a.shape)

outer_it = 0

while (np.sum(point_a[:,6] == 0) > 0):
  red, green, blue = random.randint(1, 255), random.randint(1, 255), random.randint(1, 255)
  outer_it+=1
  print(f'******************************************** outer iteration: {outer_it}')
  b0 = np.where(point_a[:,-1] == 0)[0] 
  print("Point indices where blue color is 0:", b0)
  print("Point indices count where blue color is 0:", len(b0)) ### Note if blue color is zero then all others will also be zero.

  ind = np.where(point_a[:,-1] == 0)[0][0] 
  print("First unlabeled index: ", ind)

  # Assign label to that index 
  point_a[ind,4], point_a[ind,5], point_a[ind,6] = red, green, blue

  distance = np.linalg.norm(point_a[:,:3] - point_a[ind,:3], axis = 1)
  print("distance",len(distance) )

  cond = ( distance < 0.05 ) & ( point_a[:,3] == point_a[ind,3] )  & ( point_a[:,-1] == 0)
  # condition where distance < 5cm and label == selected points label and blue color == 0 (last condition is checking if it is unlabeled) 
  indices = np.where(cond)[0]
  print("indices where condition meets",indices)

  filt_distance = distance[indices]
  print("Selected distances: ", filt_distance)
  print("Selected distances count: ",  len(filt_distance)) 


  point_a [indices,4], point_a [indices,5], point_a [indices,6] = red, green, blue
  assigned_cnt = np.sum((point_a[:,4] == red) & (point_a[:,5] == green) & (point_a[:,6] == blue)) 
  print(f'Label is assigned to {assigned_cnt} indices:')

  l = list(indices)

  it = 0
  while len(l) > 0: # while l is not empty
    it+=1
    print(f' Outer iteration: {outer_it} , Iteration: {it} , Length of l: {len(l)}')
    ind = l.pop()
    
    distance = np.linalg.norm(point_a[:,:3] - point_a[ind,:3], axis = 1)
    print("distance",len(distance) )

    cond = ( distance < 0.05 ) & ( point_a[:,3] == point_a[ind,3] )  & ( point_a[:,-1] == 0)
    # condition where distance < 5cm and label == selected points label and blue color == 0 (last condition is checking if it is unlabeled) 
    indices = np.where(cond)[0]
    print("indices where condition meets",indices)

    if len(indices) == 0:
      continue

    filt_distance = distance[indices]
    print("Selected distances: ", filt_distance)
    print("Selected distances count: ",  len(filt_distance)) 

    point_a [indices,4], point_a [indices,5], point_a [indices,6] = red, green, blue
    assigned_cnt = np.sum((point_a[:,4] == red) & (point_a[:,5] == green) & (point_a[:,6] == blue)) 
    print(f'Label is assigned to {assigned_cnt} indices:')

    l.extend(indices)

这是示例输入文件

例子.txt

对此的帮助将不胜感激。 非常感谢。