-
Notifications
You must be signed in to change notification settings - Fork 90
/
registration.py
262 lines (204 loc) · 8.51 KB
/
registration.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
"""
registration.py
---------------
Functions for registering (aligning) point clouds with ICP and feature registration.
"""
from open3d import *
import numpy as np
import cv2
def icp(source,target,voxel_size,max_correspondence_distance_coarse,max_correspondence_distance_fine,
method = "colored-icp"):
"""
Perform pointcloud registration using iterative closest point.
Parameters
----------
source : An open3d.Pointcloud instance
6D pontcloud of a source segment
target : An open3d.Pointcloud instance
6D pointcloud of a target segment
method : string
colored-icp, as in Park, Q.-Y. Zhou, and V. Koltun, Colored Point Cloud
Registration Revisited, ICCV, 2017 (slower)
point-to-plane, a coarse to fine implementation of point-to-plane icp (faster)
max_correspondence_distance_coarse : float
The max correspondence distance used for the course ICP during the process
of coarse to fine registration (if point-to-plane)
max_correspondence_distance_fine : float
The max correspondence distance used for the fine ICP during the process
of coarse to fine registration (if point-to-plane)
Returns
----------
transformation_icp: (4,4) float
The homogeneous rigid transformation that transforms source to the target's
frame
information_icp:
An information matrix returned by open3d.get_information_matrix_from_ \
point_clouds function
"""
assert method in ["point-to-plane","colored-icp"],"point-to-plane or colored-icp"
if method == "point-to-plane":
icp_coarse = registration.registration_icp(source, target,
max_correspondence_distance_coarse, np.identity(4),
registration.TransformationEstimationPointToPlane())
icp_fine = registration.registration_icp(source, target,
max_correspondence_distance_fine, icp_coarse.transformation,
registration.TransformationEstimationPointToPlane())
transformation_icp = icp_fine.transformation
if method == "colored-icp":
result_icp = registration.registration_colored_icp(source,target,voxel_size, np.identity(4),
registration.ICPConvergenceCriteria(relative_fitness = 1e-8,
relative_rmse = 1e-8, max_iteration = 50))
transformation_icp = result_icp.transformation
information_icp = registration.get_information_matrix_from_point_clouds(
source, target, max_correspondence_distance_fine,
transformation_icp)
return transformation_icp, information_icp
def feature_registration(source,target, MIN_MATCH_COUNT = 12):
"""
Obtain the rigid transformation from source to target
first find correspondence of color images by performing fast registration
using SIFT features on color images.
The corresponding depth values of the matching keypoints is then used to
obtain rigid transformation through a ransac process.
Parameters
----------
source : ((n,m) uint8, (n,m) float)
The source color image and the corresponding 3d pointcloud combined in a list
target : ((n,m) uint8, (n,m) float)
The target color image and the corresponding 3d pointcloud combined in a list
MIN_MATCH_COUNT : int
The minimum number of good corresponding feature points for the algorithm to
trust the pairwise registration result with feature matching only
Returns
----------
transform: (4,4) float or None
The homogeneous rigid transformation that transforms source to the target's
frame
if None, registration result using feature matching only cannot be trusted
either due to no enough good matching feature points are found, or the ransac
process does not return a solution
"""
cad_src, depth_src = source
cad_des, depth_des = target
# Initiate SIFT detector
sift = cv2.xfeatures2d.SIFT_create()
# find the keypoints and descripto rs with SIFT
kp1, des1 = sift.detectAndCompute(cad_src,None)
kp2, des2 = sift.detectAndCompute(cad_des,None)
# find good mathces
bf = cv2.BFMatcher()
matches = bf.knnMatch(des1,des2, k=2)
good = []
for m,n in matches:
if m.distance < 0.7*n.distance:
good.append(m)
# if number of good matching feature point is greater than the MIN_MATCH_COUNT
if len(good)>MIN_MATCH_COUNT:
src_pts = np.float32([ kp1[m.queryIdx].pt for m in good ]).reshape(-1,1,2)
dst_pts = np.float32([ kp2[m.trainIdx].pt for m in good ]).reshape(-1,1,2)
M, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC,5.0)
matchesMask = mask.ravel().tolist()
bad_match_index = np.where(np.array(matchesMask) == 0)
src_index=np.vstack(src_pts).squeeze()
src_index = np.delete(src_index, tuple(bad_match_index[0]), axis=0)
src_index[:,[0, 1]] = src_index[:,[1, 0]]
src_index = tuple(src_index.T.astype(np.int32))
src_depths = depth_src[src_index]
dst_index=np.vstack(dst_pts).squeeze()
dst_index = np.delete(dst_index, tuple(bad_match_index[0]), axis=0)
dst_index[:,[0, 1]] = dst_index[:,[1, 0]]
dst_index = tuple(dst_index.T.astype(np.int32))
dst_depths = depth_des[dst_index]
dst_good=[]
src_good=[]
dst_depths=dst_depths[matchesMask>0][0]
src_depths=src_depths[matchesMask>0][0]
for i in xrange(len(dst_depths)):
if np.sum(dst_depths[i])!=0 and np.sum(src_depths[i])!=0:
dst_good.append(dst_depths[i].tolist())
src_good.append(src_depths[i].tolist())
# get rigid transforms between 2 set of feature points through ransac
transform = match_ransac(np.asarray(src_good),np.asarray(dst_good))
return transform
else:
return None
def match_ransac(p, p_prime, tol = 0.01):
"""
A ransac process that estimates the transform between two set of points
p and p_prime.
The transform is returned if the RMSE of the smallest 70% is smaller
than the tol.
Parameters
----------
p : (n,3) float
The source 3d pointcloud as a numpy.ndarray
target : (n,3) float
The target 3d pointcloud as a numpy.ndarray
tol : float
A transform is considered found if the smallest 70% RMSE error between the
transformed p to p_prime is smaller than the tol
Returns
----------
transform: (4,4) float or None
The homogeneous rigid transformation that transforms p to the p_prime's
frame
if None, the ransac does not find a sufficiently good solution
"""
leastError = None
R = None
t= None
# the smallest 70% of the error is used to compute RMSE
k= int(len(p)*0.7)
assert len(p) == len(p_prime)
R_temp,t_temp = rigid_transform_3D(p,p_prime)
R_temp = np.array(R_temp)
t_temp = (np.array(t_temp).T)[0]
transformed = (np.dot(R_temp, p.T).T)+t_temp
error = (transformed - p_prime)**2
error = np.sum(error, axis=1)
error = np.sqrt(error)
RMSE = np.sum(error[np.argpartition(error, k)[:k]])/k
if RMSE < tol:
R = R_temp
t = t_temp
transform = [[R[0][0],R[0][1],R[0][2],t[0]],
[R[1][0],R[1][1],R[1][2],t[1]],
[R[2][0],R[2][1],R[2][2],t[2]],
[0,0,0,1]]
return transform
return None
def rigid_transform_3D(A, B):
"""
Estimate a rigid transform between 2 set of points of equal length
through singular value decomposition(svd), return a rotation and a
transformation matrix
Parameters
----------
A : (n,3) float
The source 3d pointcloud as a numpy.ndarray
B : (n,3) float
The target 3d pointcloud as a numpy.ndarray
Returns
----------
R: (3,3) float
A rigid rotation matrix
t: (3) float
A translation vector
"""
assert len(A) == len(B)
A= np.asmatrix(A)
B= np.asmatrix(B)
N = A.shape[0];
centroid_A = np.mean(A, axis=0)
centroid_B = np.mean(B, axis=0)
AA = A - np.tile(centroid_A, (N, 1))
BB = B - np.tile(centroid_B, (N, 1))
H = AA.T * BB
U, S, Vt = np.linalg.svd(H)
R = Vt.T * U.T
# reflection case
if np.linalg.det(R) < 0:
Vt[2,:] *= -1
R = Vt.T * U.T
t = -R*centroid_A.T + centroid_B.T
return (R, t)