Question

在对地面进行分类后，我试图分割LiDAR点云。我正在使用numpy创建点云（pc）的“图像”，并遍历numpy数组。我想加快循环速度，或者避免一起循环。我将使用图像分割技术，但是首先我需要运行此代码来创建“图像”，这是需要一段时间的部分。有没有办法提高这个循环的速度或避免它的速度？


import numpy as np
from math import ceil, floor


'''In this case:
pc = point cloud (X,Y,Z values)'''

# point cloud is in the numpy array, pc
minx,maxx,miny,maxy = floor(np.min(pc[:,0]-1)),ceil(np.max(pc[:,0]+1)),floor(np.min(pc[:,1]-1)),ceil(np.max(pc[:,1]+1))# x,y bounding box

# grid x and y direction (resolution: 0.2 meters)
gridx = np.linspace(minx,maxx,int((maxx - minx+0.2)*5),endpoint=True) 
gridy = np.linspace(miny,maxy,int((maxy - miny +0.2)*5),endpoint=True)

#shape of the new image with 0.2 meter resolution.
imgx,imgy = int((maxx-minx+0.2)*5),int((maxy - miny +0.2)*5)

# this is what will be created at the end.  It will be a binary image.
img = np.zeros((imgx,imgy))

#loop through array to generate image (this is the part that takes a while)
for x,i in enumerate(gridx):
    for y,j in enumerate(gridy):

# Test if there any points in this "grid"
        input_point = pc[np.where(((pc[:,0]>i) & (pc[:,0]<i+1))& ((pc[:,1]>j) & (pc[:,1]<j+1)))]
# if there are points, give pixel value 1.
        if input_point.shape[0]!=0:
            img[x,y]=1

print('Image made')

谢谢。

Answer 1

这里是矢量化版本，可在随机测试集上产生相同的输出：

import numpy as np
from math import ceil, floor
import time

width = 0.2

t = [time.time()]

pc = np.random.uniform(-10, 10, (100, 3))

# point cloud is in the numpy array, pc
minx,maxx,miny,maxy = floor(np.min(pc[:,0]-1)),ceil(np.max(pc[:,0]+1)),floor(np.min(pc[:,1]-1)),ceil(np.max(pc[:,1]+1))# x,y bounding box

# grid x and y direction (resolution: 0.2 meters)
gridx = np.linspace(minx,maxx,int((maxx - minx+0.2)*5),endpoint=True) 
gridy = np.linspace(miny,maxy,int((maxy - miny +0.2)*5),endpoint=True)

#shape of the new image with 0.2 meter resolution.
imgx,imgy = int((maxx-minx+0.2)*5),int((maxy - miny +0.2)*5)

print('Shared ops done')
t.append(time.time())

# this is what will be created at the end.  It will be a binary image.
img = np.zeros((imgx,imgy))

#loop through array to generate image (this is the part that takes a while)
for x,i in enumerate(gridx):
    for y,j in enumerate(gridy):

# Test if there any points in this "grid"
        input_point = pc[np.where(((pc[:,0]>i) & (pc[:,0]<i+width))& ((pc[:,1]>j) & (pc[:,1]<j+width)))]
# if there are points, give pixel value 1.
        if input_point.shape[0]!=0:
            img[x,y]=1

t.append(time.time())            
print('Image made')

if width == 0.2:
    img2 = np.zeros((imgx, imgy), 'u1')
    x2, y2 = (((pc[:, :2] - (minx, miny)) * (5, 5))).astype(int).T
    img2[x2, y2] = 1

elif width == 1:
    img2 = np.zeros((imgx+4, imgy+4), 'u1')
    x2, y2 = (((pc[:, :2] - (minx, miny)) * (5, 5))).astype(int).T

    np.lib.stride_tricks.as_strided(img2, (imgx, imgy, 5, 5), 2 * img2.strides)[x2, y2] = 1
    img2 = img2[4:, 4:]

t.append(time.time())
print('Image remade')

print('took', np.diff(t), 'secs respectively')
assert((img2==img).all())
print('results equal')

您的代码产生5x5像素。那是故意的吗？要重现它，我必须有点棘手。

更新：添加了代替普通像素的版本。

样品运行：

Shared ops done
Image made
Image remade
took [2.29120255e-04 1.54510736e-01 1.44481659e-04] secs respectively
results equal

提高通过numpy数组循环的速度

1 个答案: