点云数据常用处理：python实现
目录
- 数据集增强
-
- 仿射变换
-
- 平移变换
- 旋转变换
- 尺度变换
- 仿射变换
- 添加噪声
-
- 高斯噪声
- 随机噪声
- 下采样
-
- 指定体素
- 指定点数
- 数据归一化
-
- 去中心化
- 尺度归一化
数据集增强
仿射变换
平移变换
import numpy as np
import random
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
#平移参数
x_offset=random.uniform(-10, 10)
y_offset=random.uniform(-10, 10)
z_offset=random.uniform(-10, 10)
#变换矩阵
transformation_matrix=np.array([
[1,0,0,x_offset],
[0,1,0,y_offset],
[0,0,1,z_offset],
[0,0,0,1]
])
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
#补充数据为齐次项
ones_data=np.ones(old_xyz.shape[0])
old_xyz=np.insert(old_xyz,3,values=ones_data,axis=1)
#变换数据
new_xyz = np.dot(transformation_matrix,old_xyz.T)
new_array=np.concatenate((new_xyz.T[:,:3],old_array[:,3:]),axis=1)
np.savetxt(new_file,new_array,fmt='%.06f')
旋转变换
import numpy as np
import random
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
#旋转角度
roate_x=random.uniform(-np.pi/10, np.pi/10)
roate_y=random.uniform(-np.pi/10, np.pi/10)
roate_z=random.uniform(-np.pi/10, np.pi/10)
roate_x_matrix=np.array([
[1,0,0,0],
[0,np.cos(roate_x),-np.sin(roate_x),0],
[0,np.sin(roate_x),np.cos(roate_x),0],
[0,0,0,1]
])
roate_y_matrix=np.array([
[np.cos(roate_y),0,np.sin(roate_y),0],
[0,1,0,0],
[-np.sin(roate_y),0,np.cos(roate_y),0],
[0,0,0,1]
])
roate_z_matrix=np.array([
[np.cos(roate_z),-np.sin(roate_z),0,0],
[np.sin(roate_z),np.cos(roate_z),0,0],
[0,0,1,0],
[0,0,0,1]
])
#变换矩阵
transformation_matrix=dot(roate_z_matrix).dot(roate_y_matrix).dot(roate_x_matrix)
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
#补充数据为齐次项
ones_data=np.ones(old_xyz.shape[0])
old_xyz=np.insert(old_xyz,3,values=ones_data,axis=1)
#变换数据
new_xyz = np.dot(transformation_matrix,old_xyz.T)
new_array=np.concatenate((new_xyz.T[:,:3],old_array[:,3:]),axis=1)
np.savetxt(new_file,new_array,fmt='%.06f')
尺度变换
import numpy as np
import random
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
#缩放参数
scale=0.1
#变换矩阵
transformation_matrix=np.array([
[scale,0,0,0],
[0,scale,0,0],
[0,0,scale,0],
[0,0,0,1]
])
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
#补充数据为齐次项
ones_data=np.ones(old_xyz.shape[0])
old_xyz=np.insert(old_xyz,3,values=ones_data,axis=1)
#变换数据
new_xyz = np.dot(transformation_matrix,old_xyz.T)
new_array=np.concatenate((new_xyz.T[:,:3],old_array[:,3:]),axis=1)
np.savetxt(new_file,new_array,fmt='%.06f')
仿射变换
上面三种变换综合可以写成：
import numpy as np
import random
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
#平移参数
x_offset=random.uniform(-10, 10)
y_offset=random.uniform(-10, 10)
z_offset=random.uniform(-10, 10)
#缩放参数
scale=0.1
#旋转角度
roate_x=random.uniform(-np.pi/10, np.pi/10)
roate_y=random.uniform(-np.pi/10, np.pi/10)
roate_z=random.uniform(-np.pi/10, np.pi/10)
roate_x_matrix=np.array([
[1,0,0,0],
[0,np.cos(roate_x),-np.sin(roate_x),0],
[0,np.sin(roate_x),np.cos(roate_x),0],
[0,0,0,1]
])
roate_y_matrix=np.array([
[np.cos(roate_y),0,np.sin(roate_y),0],
[0,1,0,0],
[-np.sin(roate_y),0,np.cos(roate_y),0],
[0,0,0,1]
])
roate_z_matrix=np.array([
[np.cos(roate_z),-np.sin(roate_z),0,0],
[np.sin(roate_z),np.cos(roate_z),0,0],
[0,0,1,0],
[0,0,0,1]
])
#变换矩阵
transformation_matrix=np.array([
[scale,0,0,x_offset],
[0,scale,0,y_offset],
[0,0,scale,z_offset],
[0,0,0,1]
]).dot(roate_z_matrix).dot(roate_y_matrix).dot(roate_x_matrix)
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
#补充数据为齐次项
ones_data=np.ones(old_xyz.shape[0])
old_xyz=np.insert(old_xyz,3,values=ones_data,axis=1)
#变换数据
new_xyz = np.dot(transformation_matrix,old_xyz.T)
new_array=np.concatenate((new_xyz.T[:,:3],old_array[:,3:]),axis=1)
np.savetxt(new_file,new_array,fmt='%.06f')
仿射变换：
PCL 仿射变换，实现点云平移旋转
添加噪声
高斯噪声
import numpy as np
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
def add_noise(point, sigma=0.1, clip=0.1):
point = point.reshape(-1,3)
Row, Col = point.shape
noisy_point = np.clip(sigma * np.random.randn(Row, Col), -1*clip, clip)
noisy_point += point
return noisy_point
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
new_xyz=add_noise(old_xyz)
new_array=np.concatenate((new_xyz,old_array[:,3:]),axis=1)
np.savetxt(new_file,new_array,fmt='%.06f')
随机噪声
import numpy as np
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
def guass_noise(point, sigma=1, clip=1):
point = point.reshape(-1,3)
Row, Col = point.shape
noisy_point = np.clip(sigma * np.random.rand(Row, Col), -1*clip, clip)
noisy_point += point
return noisy_point
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
new_xyz=guass_noise(old_xyz)
new_array=np.concatenate((new_xyz,old_array[:,3:]),axis=1)
np.savetxt(new_file,new_array,fmt='%.06f')
下采样
指定体素
import numpy as np
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
#体素滤波
def voxel_filter(point_cloud, leaf_size):
filtered_points = []
# 计算边界点
x_min, y_min, z_min = np.amin(point_cloud, axis=0) #计算x y z 三个维度的最值
x_max, y_max, z_max = np.amax(point_cloud, axis=0)
# 计算 voxel grid维度
Dx = (x_max - x_min) // leaf_size + 1
Dy = (y_max - y_min) // leaf_size + 1
Dz = (z_max - z_min) // leaf_size + 1
# 计算每个点的voxel索引
h = list() # h为保存索引的列表
for i in range(len(point_cloud)):
hx = (point_cloud[i][0] - x_min) // leaf_size
hy = (point_cloud[i][1] - y_min) // leaf_size
hz = (point_cloud[i][2] - z_min) // leaf_size
h.append(hx + hy * Dx + hz * Dx * Dy)
h = np.array(h)
#筛选点
h_indice = np.argsort(h) #返回h里面的元素按从小到大排序的索引
h_sorted = h[h_indice]
begin = 0
for i in range(len(h_sorted) - 1): # 0~9999
if h_sorted[i] != h_sorted[i + 1]:
point_idx = h_indice[begin: i + 1]
filtered_points.append(np.mean(point_cloud[point_idx], axis=0))
begin = i+1
#把点云格式改成array，并对外返回
filtered_points = np.array(filtered_points, dtype=np.float64)
return filtered_points
#加载文件
old_array=np.loadtxt(old_file)
old_xyz=old_array[:,:3]
#保存文件
new_xyz=voxel_filter(old_xyz, 1)
np.savetxt(new_file,new_xyz,fmt='%.06f')
调用open3d的api：
import open3d as o3d
import numpy as np
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
cloud=o3d.io.read_point_cloud(old_file,format='xyzrgb')
cloud_downsampled=cloud.voxel_down_sample(1)
cloud_downsampled = np.array(cloud_downsampled.points)
np.savetxt(new_file,cloud_downsampled,fmt='%.06f')
调用pcl的api：
import pcl
import numpy as np
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
cloud = pcl.PointCloud()
pcd = np.genfromtxt(old_file, delimiter=" ")
pcd = np.array(pcd, dtype = np.float32)
pcd=pcd[:,:3]
cloud.from_array(pcd)
cloud_down=pcl.ApproximateVoxelGrid()
pcl.ApproximateVoxelGrid.set_InputCloud(cloud_down, cloud)
pcl.ApproximateVoxelGrid.set_leaf_size(cloud_down, 1, 1, 1)
cloud_downsampled=pcl.ApproximateVoxelGrid.filter(cloud_down)
cloud_downsampled = np.array(cloud_downsampled.to_array())
np.savetxt(new_file,cloud_downsampled,fmt='%.06f')
指定点数
import numpy as np
#文件名
old_file=r"rabbit.txt"
new_file=r"rabbit_change.txt"
#加载文件
old_array = np.loadtxt(old_file).astype(np.float32)
points_num =2048
choice = np.random.choice(old_array.shape[0], points_num, replace=True)
#保存文件
new_array = old_array[choice, :]
np.savetxt(new_file,new_array,fmt='%.06f')
数据归一化
去中心化
centroid = np.mean(pc, axis=0)
pc = pc - centroid
尺度归一化
m = np.max(np.sqrt(np.sum(pc**2, axis=1)))
pc = pc / m