1 | #!/usr/bin/env python |
||
2 | |||
3 | 1 | """ |
|
4 | Registration algorithms and utility functions |
||
5 | """ |
||
6 | |||
7 | 1 | from __future__ import print_function |
|
8 | 1 | import numpy as np |
|
9 | 1 | from .. import BoundingBox, force_srs, extract_mask, clone |
|
10 | 1 | from .stickscale import get_stick_scale |
|
11 | 1 | from pcl.registration import gicp |
|
12 | |||
13 | 1 | from patty.utils import ( |
|
14 | log, |
||
15 | save, |
||
16 | downsample_voxel, |
||
17 | ) |
||
18 | |||
19 | 1 | from patty.segmentation import ( |
|
20 | boundary_of_center_object, |
||
21 | boundary_of_drivemap, |
||
22 | boundary_of_lowest_points, |
||
23 | ) |
||
24 | |||
25 | 1 | from sklearn.decomposition import PCA |
|
26 | |||
27 | |||
28 | 1 | def align_footprints(loose_pc, fixed_pc, |
|
29 | allow_scaling=True, |
||
30 | allow_rotation=True, |
||
31 | allow_translation=True): |
||
32 | ''' |
||
33 | Align a pointcloud 'loose_pc' by placing it on top of |
||
34 | 'fixed_pc' as good as poosible. Done by aligning the |
||
35 | principle axis of both pointclouds. |
||
36 | |||
37 | NOTE: Both pointclouds are assumed to be the footprint (or projection) |
||
38 | on the xy plane, with basically zero extent along the z-axis. |
||
39 | |||
40 | (allow_rotation=True) |
||
41 | The pointcloud boundary is alinged with the footprint |
||
42 | by rotating its pricipal axis in the (x,y) plane. |
||
43 | |||
44 | (allow_translation=True) |
||
45 | Then, it is translated so the centers of mass coincide. |
||
46 | |||
47 | (allow_scaling=True) |
||
48 | Finally, the pointcloud is scaled to have the same extent. |
||
49 | |||
50 | Arguments: |
||
51 | loose_pc : pcl.PointCloud |
||
52 | fixed_pc : pcl.PointCloud |
||
53 | |||
54 | allow_scaling : Bolean |
||
55 | allow_rotation : Bolean |
||
56 | allow_translation : Bolean |
||
57 | |||
58 | Returns: |
||
59 | rot_matrix, rot_center, scale, translation : np.array() |
||
60 | |||
61 | ''' |
||
62 | |||
63 | rot_center = loose_pc.center() |
||
64 | |||
65 | if allow_rotation: |
||
66 | log(" - Finding rotation") |
||
67 | rot_matrix = find_rotation_xy(loose_pc, fixed_pc) |
||
68 | loose_pc.rotate(rot_matrix, origin=rot_center) |
||
69 | else: |
||
70 | log(" - Skipping rotation") |
||
71 | rot_matrix = np.eye(3) |
||
72 | |||
73 | if allow_scaling: |
||
74 | fixed_bb = BoundingBox(fixed_pc) # used 2x below |
||
75 | loose_bb = BoundingBox(loose_pc) |
||
76 | scale = fixed_bb.size[0:2] / loose_bb.size[0:2] |
||
77 | |||
78 | # take the average scale factor for the x and y dimensions |
||
79 | scale = np.mean(scale) |
||
80 | loose_pc.scale(scale, origin=rot_center) |
||
81 | log(" - Scale: %s" % scale) |
||
82 | else: |
||
83 | log(" - Skipping scale") |
||
84 | scale = 1.0 |
||
85 | |||
86 | if allow_translation: |
||
87 | translation = fixed_pc.center() - rot_center |
||
88 | loose_pc.translate(translation) |
||
89 | log(" - Translation: %s" % translation) |
||
90 | else: |
||
91 | log(" - Skipping translation") |
||
92 | translation = np.array([0.0, 0.0, 0.0]) |
||
93 | |||
94 | return rot_matrix, rot_center, scale, translation |
||
95 | |||
96 | |||
97 | 1 | def estimate_pancake_up(pointcloud): |
|
98 | ''' |
||
99 | Assuming a pancake like pointcloud, the up direction is the third PCA. |
||
100 | ''' |
||
101 | pca = PCA(n_components=3) |
||
102 | |||
103 | points = np.asarray(pointcloud) |
||
104 | pca.fit(points[:, 0:3]) |
||
105 | |||
106 | return pca.components_[2] |
||
107 | |||
108 | |||
109 | 1 | def _find_rotation_xy_helper(pointcloud): |
|
110 | 1 | pca = PCA(n_components=2) |
|
111 | |||
112 | 1 | points = np.asarray(pointcloud) |
|
113 | 1 | pca.fit(points[:, 0:2]) |
|
114 | |||
115 | 1 | rotxy = np.array(pca.components_) |
|
116 | |||
117 | # make sure the rotation is a proper rotation, ie det = +1 |
||
118 | 1 | if np.linalg.det(rotxy) < 0: |
|
119 | 1 | rotxy[:, 1] *= -1.0 |
|
120 | |||
121 | # create a 3D rotation around the z-axis |
||
122 | 1 | rotation = np.eye(3) |
|
123 | 1 | rotation[0:2, 0:2] = rotxy |
|
124 | |||
125 | 1 | return rotation |
|
126 | |||
127 | |||
128 | 1 | def find_rotation_xy(pc, ref): |
|
129 | '''Find the transformation that rotates the principal axis of the |
||
130 | pointcloud onto those of the reference. |
||
131 | Keep the z-axis pointing upwards. |
||
132 | |||
133 | Arguments: |
||
134 | pc: pcl.PointCloud |
||
135 | |||
136 | ref: pcl.PointCloud |
||
137 | |||
138 | Returns: |
||
139 | numpy array of shape [3,3], can be used to rotate pointclouds |
||
140 | with pc.rotate() |
||
141 | ''' |
||
142 | |||
143 | 1 | pc_transform = _find_rotation_xy_helper(pc) |
|
144 | 1 | ref_transform = _find_rotation_xy_helper(ref) |
|
145 | |||
146 | 1 | return np.dot(np.linalg.inv(ref_transform), pc_transform) |
|
147 | |||
148 | |||
149 | 1 | def rotate_upwards(pc, up): |
|
150 | ''' |
||
151 | Rotate the pointcloud in-place around its center, such that the |
||
152 | 'up' vector points along [0,0,1] |
||
153 | |||
154 | Arguments: |
||
155 | pc : pcl.PointCloud |
||
156 | up : np.array([3]) |
||
157 | |||
158 | Returns: |
||
159 | pc : pcl.PointCloud the input pointcloud, for convenience. |
||
160 | |||
161 | ''' |
||
162 | |||
163 | newz = np.array(up) |
||
164 | |||
165 | # Right-handed coordiante system: |
||
166 | # np.cross(x,y) = z |
||
167 | # np.cross(y,z) = x |
||
168 | # np.cross(z,x) = y |
||
169 | |||
170 | # normalize |
||
171 | newz /= (np.dot(newz, newz)) ** 0.5 |
||
172 | |||
173 | # find two orthogonal vectors to represent x and y, |
||
174 | # randomly choose a vector, and take cross product. If we're unlucky, |
||
175 | # this ones is parallel to z, so cross pruduct is zero. |
||
176 | # In that case, try another one |
||
177 | try: |
||
178 | newx = np.cross(np.array([0, 1, 0]), newz) |
||
179 | newx /= (np.dot(newx, newx)) ** 0.5 |
||
180 | |||
181 | newy = np.cross(newz, newx) |
||
182 | newy /= (np.dot(newy, newy)) ** 0.5 |
||
183 | except: |
||
184 | print("Alternative") |
||
185 | newy = np.cross(newz, np.array([1, 0, 0])) |
||
186 | newy /= (np.dot(newy, newy)) ** 0.5 |
||
187 | |||
188 | newx = np.cross(newy, newz) |
||
189 | newx /= (np.dot(newx, newx)) ** 0.5 |
||
190 | |||
191 | rotation = np.zeros([3, 3]) |
||
192 | rotation[0, 0] = newx[0] |
||
193 | rotation[1, 0] = newx[1] |
||
194 | rotation[2, 0] = newx[2] |
||
195 | |||
196 | rotation[0, 1] = newy[0] |
||
197 | rotation[1, 1] = newy[1] |
||
198 | rotation[2, 1] = newy[2] |
||
199 | |||
200 | rotation[0, 2] = newz[0] |
||
201 | rotation[1, 2] = newz[1] |
||
202 | rotation[2, 2] = newz[2] |
||
203 | |||
204 | rotation = np.linalg.inv(rotation) |
||
205 | pc.rotate(rotation, origin=pc.center()) |
||
206 | |||
207 | return pc |
||
208 | |||
209 | |||
210 | 1 | def initial_registration(pointcloud, up, drivemap, |
|
211 | initial_scale=None, trust_up=True): |
||
212 | """ |
||
213 | Initial registration adds an spatial reference system to the pointcloud, |
||
214 | and place the pointlcoud on top of the drivemap. The pointcloud is rotated |
||
215 | so that the up vector points along [0,0,1], and scaled such that it has the |
||
216 | right order of magnitude in size. |
||
217 | |||
218 | Arguments: |
||
219 | pointcloud : pcl.PointCloud |
||
220 | The high-res object to register. |
||
221 | |||
222 | up: np.array([3]) |
||
223 | Up direction for the pointcloud. |
||
224 | If None, assume the object is pancake shaped, and chose the |
||
225 | upvector such that it is perpendicullar to the pancake. |
||
226 | |||
227 | drivemap : pcl.PointCloud |
||
228 | A small part of the low-res drivemap on which to register. |
||
229 | |||
230 | initial_scale : float |
||
231 | if given, scale pointcloud using this value; estimate scale factor |
||
232 | from bounding boxes. |
||
233 | |||
234 | trust_up : Boolean, default to True |
||
235 | True: Assume the up vector is exact. |
||
236 | False: Calculate 'up' as if it was None, but orient it such that |
||
237 | np.dot( up, pancake_up ) > 0 |
||
238 | |||
239 | NOTE: Modifies the input pointcloud in-place, and leaves |
||
240 | it in a undefined state. |
||
241 | |||
242 | """ |
||
243 | log("Starting initial registration") |
||
244 | |||
245 | ##### |
||
246 | # set scale and offset of pointcloud, drivemap, and footprint |
||
247 | # as the pointcloud is unregisterd, the coordinate system is undefined, |
||
248 | # and we lose nothing if we just copy it |
||
249 | |||
250 | if(hasattr(pointcloud, "offset")): |
||
251 | log(" - Dropping initial offset, was: %s" % pointcloud.offset) |
||
252 | else: |
||
253 | log(" - No initial offset") |
||
254 | force_srs(pointcloud, same_as=drivemap) |
||
255 | log(" - New offset forced to: %s" % pointcloud.offset) |
||
256 | |||
257 | if up is not None: |
||
258 | log(" - Rotating the pointcloud so up points along [0,0,1]") |
||
259 | |||
260 | if trust_up: |
||
261 | rotate_upwards(pointcloud, up) |
||
262 | log(" - Using trusted up: %s" % up) |
||
263 | else: |
||
264 | pancake_up = estimate_pancake_up(pointcloud) |
||
265 | if np.dot(up, pancake_up) < 0.0: |
||
266 | pancake_up *= -1.0 |
||
267 | log(" - Using estimated up: %s" % pancake_up) |
||
268 | rotate_upwards(pointcloud, pancake_up) |
||
269 | |||
270 | else: |
||
271 | log(" - No upvector, skipping") |
||
272 | |||
273 | if initial_scale is None: |
||
274 | bbDrivemap = BoundingBox(points=np.asarray(drivemap)) |
||
0 ignored issues
–
show
|
|||
275 | bbObject = BoundingBox(points=np.asarray(pointcloud)) |
||
0 ignored issues
–
show
The name
bbObject does not conform to the variable naming conventions ([a-z_][a-z0-9_]{1,30}$ ).
This check looks for invalid names for a range of different identifiers. You can set regular expressions to which the identifiers must conform if the defaults do not match your requirements. If your project includes a Pylint configuration file, the settings contained in that file take precedence. To find out more about Pylint, please refer to their site. ![]() |
|||
276 | scale = bbDrivemap.size[0:2] / bbObject.size[0:2] # ignore z-direction |
||
277 | |||
278 | # take the average scale factor for x and y dimensions |
||
279 | scale = np.mean(scale) |
||
280 | else: |
||
281 | # use user provided scale |
||
282 | scale = initial_scale |
||
283 | |||
284 | log(" - Applying rough estimation of scale factor", scale) |
||
285 | pointcloud.scale(scale) # dont care about origin of scaling |
||
286 | |||
287 | |||
288 | 1 | def coarse_registration(pointcloud, drivemap, footprint, downsample=None): |
|
289 | """ |
||
290 | Improve the initial registration. |
||
291 | Find the proper scale by looking for the red meter sticks, and calculate |
||
292 | and align the pointcloud's footprint. |
||
293 | |||
294 | Arguments: |
||
295 | pointcloud: pcl.PointCloud |
||
296 | The high-res object to register. |
||
297 | |||
298 | drivemap: pcl.PointCloud |
||
299 | A small part of the low-res drivemap on which to register |
||
300 | |||
301 | footprint: pcl.PointCloud |
||
302 | Pointlcloud containing the objects footprint |
||
303 | |||
304 | downsample: float, default=None, no resampling |
||
305 | Downsample the high-res pointcloud before footprint |
||
306 | calculation. |
||
307 | """ |
||
308 | log("Starting coarse registration") |
||
309 | |||
310 | ### |
||
311 | # find redstick scale, and use it if possible |
||
312 | log(" - Redstick scaling") |
||
313 | |||
314 | allow_scaling = True |
||
315 | |||
316 | scale, confidence = get_stick_scale(pointcloud) |
||
317 | log(" - Red stick scale=%s confidence=%s" % (scale, confidence)) |
||
318 | |||
319 | if (confidence > 0.5): |
||
320 | log(" - Applying red stick scale") |
||
321 | pointcloud.scale(1.0 / scale) # dont care about origin of scaling |
||
322 | allow_scaling = False |
||
323 | else: |
||
324 | log(" - Not applying red stick scale, confidence too low") |
||
325 | |||
326 | ##### |
||
327 | # find all the points in the drivemap along the footprint |
||
328 | # use bottom two meters of drivemap (not trees) |
||
329 | |||
330 | dm_boundary = boundary_of_drivemap(drivemap, footprint) |
||
331 | dm_bb = BoundingBox(dm_boundary) |
||
332 | |||
333 | # set footprint height from minimum value, |
||
334 | # as trees, or high object make the pc.center() too high |
||
335 | fixed_boundary = clone(footprint) |
||
336 | fp_array = np.asarray(fixed_boundary) |
||
337 | fp_array[:, 2] = dm_bb.min[2] |
||
338 | |||
339 | save(fixed_boundary, "fixed_bound.las") |
||
340 | |||
341 | ##### |
||
342 | # find all the boundary points of the pointcloud |
||
343 | |||
344 | loose_boundary = boundary_of_center_object(pointcloud, downsample) |
||
345 | if loose_boundary is None: |
||
346 | log(" - boundary estimation failed, using lowest 30 percent of points") |
||
347 | loose_boundary = boundary_of_lowest_points(pointcloud, |
||
348 | height_fraction=0.3) |
||
349 | |||
350 | #### |
||
351 | # match the pointcloud boundary with the footprint boundary |
||
352 | |||
353 | log(" - Aligning footprints:") |
||
354 | rot_matrix, rot_center, scale, translation = align_footprints( |
||
355 | loose_boundary, fixed_boundary, |
||
356 | allow_scaling=allow_scaling, |
||
357 | allow_rotation=True, |
||
358 | allow_translation=True) |
||
359 | |||
360 | save(loose_boundary, "aligned_bound.las") |
||
361 | |||
362 | #### |
||
363 | # Apply to the main pointcloud |
||
364 | |||
365 | pointcloud.rotate(rot_matrix, origin=rot_center) |
||
366 | pointcloud.scale(scale, origin=rot_center) |
||
367 | pointcloud.translate(translation) |
||
368 | rot_center += translation |
||
369 | |||
370 | return rot_center |
||
371 | |||
372 | |||
373 | 1 | def _fine_registration_helper(pointcloud, drivemap, voxelsize=0.05, attempt=0): |
|
374 | """ |
||
375 | Perform ICP on pointcloud with drivemap, and return convergence indicator. |
||
376 | Reject large translatoins. |
||
377 | |||
378 | Returns: |
||
379 | transf : np.array([4,4]) |
||
380 | transform |
||
381 | success : Boolean |
||
382 | if icp was successful |
||
383 | fitness : float |
||
384 | sort of sum of square differences, ie. smaller is better |
||
385 | """ |
||
386 | #### |
||
387 | # Downsample to speed up |
||
388 | # use voxel filter to keep evenly distributed spatial extent |
||
389 | |||
390 | log(" - Downsampling with voxel filter: %s" % voxelsize) |
||
391 | pc = downsample_voxel(pointcloud, voxelsize) |
||
392 | |||
393 | #### |
||
394 | # Clip to drivemap to prevent outliers confusing the ICP algorithm |
||
395 | |||
396 | log(" - Clipping to drivemap") |
||
397 | bb = BoundingBox(drivemap) |
||
398 | z = bb.center[2] |
||
0 ignored issues
–
show
The name
z does not conform to the variable naming conventions ([a-z_][a-z0-9_]{1,30}$ ).
This check looks for invalid names for a range of different identifiers. You can set regular expressions to which the identifiers must conform if the defaults do not match your requirements. If your project includes a Pylint configuration file, the settings contained in that file take precedence. To find out more about Pylint, please refer to their site. ![]() |
|||
399 | extracted = extract_mask(pc, [bb.contains([point[0], point[1], z]) |
||
400 | for point in pc]) |
||
401 | |||
402 | log(" - Remaining points: %s" % len(extracted)) |
||
403 | |||
404 | #### |
||
405 | # GICP |
||
406 | |||
407 | converged, transf, estimate, fitness = gicp(extracted, drivemap) |
||
408 | |||
409 | #### |
||
410 | # Dont accept large translations |
||
411 | |||
412 | translation = transf[0:3, 3] |
||
413 | if np.dot(translation, translation) > 5 ** 2: |
||
414 | log(" - Translation too large, considering it a failure.") |
||
415 | converged = False |
||
416 | fitness = 1e30 |
||
417 | else: |
||
418 | log(" - Success, fitness: ", converged, fitness) |
||
419 | |||
420 | force_srs(estimate, same_as=pointcloud) |
||
421 | save(estimate, "attempt%s.las" % attempt) |
||
422 | |||
423 | return transf, converged, fitness |
||
424 | |||
425 | |||
426 | 1 | def fine_registration(pointcloud, drivemap, center, voxelsize=0.05): |
|
427 | """ |
||
428 | Final registration step using ICP. |
||
429 | |||
430 | Find the local optimal postion of the pointcloud on the drivemap; due to |
||
431 | our coarse_registration algorithm, we have to try two orientations: |
||
432 | original, and rotated by 180 degrees around the z-axis. |
||
433 | |||
434 | Arguments: |
||
435 | pointcloud: pcl.PointCloud |
||
436 | The high-res object to register. |
||
437 | |||
438 | drivemap: pcl.PointCloud |
||
439 | A small part of the low-res drivemap on which to register |
||
440 | |||
441 | center: np.array([3]) |
||
442 | Vector giving the centerpoint of the pointcloud, used to do |
||
443 | the 180 degree rotations. |
||
444 | |||
445 | voxelsize: float default : 0.05 |
||
446 | Size in [m] of the voxel grid used for downsampling |
||
447 | """ |
||
448 | log("Starting fine registration") |
||
449 | |||
450 | # for rotation around z-axis |
||
451 | rot = np.array([[0, -1, 0], [1, 0, 0], [0, 0, 1]]) |
||
452 | |||
453 | #### |
||
454 | # do a ICP step for 4 orientations |
||
455 | |||
456 | transf = {} |
||
457 | success = {} |
||
458 | fitness = {} |
||
459 | for i in range(4): |
||
460 | log(" - attempt: %s" % i) |
||
461 | transf[i], success[i], fitness[i] = _fine_registration_helper( |
||
462 | pointcloud, |
||
463 | drivemap, |
||
464 | attempt=i, |
||
465 | voxelsize=voxelsize) |
||
466 | pointcloud.rotate(rot, origin=center) |
||
467 | |||
468 | #### |
||
469 | # pick best |
||
470 | |||
471 | best, value = min(fitness.iteritems(), key=lambda x: x[1]) |
||
472 | if success[best]: |
||
473 | log(" - Best attempt: %s" % best) |
||
474 | if best > 0: # np.array()**0 does weird things |
||
475 | pointcloud.rotate(rot**best, origin=center) |
||
476 | pointcloud.transform(transf[best]) |
||
477 | return |
||
478 | |||
479 | # ICP failed: |
||
480 | # return the pointcloud with just footprints aligned |
||
481 | # no use to undo a rotation, as any orientationi is equally likely. |
||
482 | log(" - Unable to do fine registration") |
||
483 | |||
484 | return |
||
485 |
This check looks for invalid names for a range of different identifiers.
You can set regular expressions to which the identifiers must conform if the defaults do not match your requirements.
If your project includes a Pylint configuration file, the settings contained in that file take precedence.
To find out more about Pylint, please refer to their site.