-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
416 lines (323 loc) · 16.1 KB
/
utils.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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
from common.transformations.camera import normalize, get_view_frame_from_calib_frame
from common.transformations.model import medmodel_intrinsics
import common.transformations.orientation as orient
import numpy as np
import math
import os
import cv2
PATH_TO_CACHE = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'cache')
FULL_FRAME_SIZE = (1164, 874) # input frame to openpilot model
W, H = FULL_FRAME_SIZE[0], FULL_FRAME_SIZE[1]
eon_focal_length = FOCAL = 910.0 # focal length of eon devkit image (LeECO Pro 3 smartphone image)
# aka 'K' aka camera_frame_from_view_frame
eon_intrinsics = np.array(
[ # intrinsics image array. For more info: https://docs.opencv.org/4.x/dc/dbb/tutorial_py_calibration.html
[FOCAL, 0., W / 2.],
[0., FOCAL, H / 2.],
[0., 0., 1.]])
ground_from_medmodel_frame = [
[0.00000000e+00, 0.00000000e+00, 1.00000000e+00],
[-1.09890110e-03, 0.00000000e+00, 2.81318681e-01],
[-1.84808520e-20, 9.00738606e-04, -4.28751576e-02]]
# fixed distances
X_IDXs = [
0., 0.1875, 0.75, 1.6875, 3., 4.6875,
6.75, 9.1875, 12., 15.1875, 18.75, 22.6875,
27., 31.6875, 36.75, 42.1875, 48., 54.1875,
60.75, 67.6875, 75., 82.6875, 90.75, 99.1875,
108., 117.1875, 126.75, 136.6875, 147., 157.6875,
168.75, 180.1875, 192.]
_BB_SCALE = W / 640.
_BB_TO_FULL_FRAME = np.asarray([
[_BB_SCALE, 0., 0.],
[0., _BB_SCALE, 0.],
[0., 0., 1.]])
def printf(*args, **kwargs):
print(flush=True, *args, **kwargs)
def sigmoid(x):
return 1 / (1 + np.exp(-x))
def extract_preds(outputs, best_plan_only=True):
"""
This function extracts lane lines, road edges and estimated plan outputs from the model output tensor.
:param outputs: Output tensor from the openpilot supercombo model
:param best_plan_only: Default True, will only output driving plan with highest confidence otherwise will output
all plans
:return: Tuple ((lanelines, lanelines_probs), (road_edges, road_edges_probs), plan).
"""
# N is batch_size
plan_start_idx = 0
plan_end_idx = 4955
lanes_start_idx = plan_end_idx
lanes_end_idx = lanes_start_idx + 528
lane_lines_prob_start_idx = lanes_end_idx
lane_lines_prob_end_idx = lane_lines_prob_start_idx + 8
road_start_idx = lane_lines_prob_end_idx
road_end_idx = road_start_idx + 264
# plan
plan = outputs[:, plan_start_idx:plan_end_idx] # (N, 4955)
plans = plan.reshape((-1, 5, 991)) # (N, 5, 991)
plan_probs = plans[:, :, -1] # (N, 5)
plans = plans[:, :, :-1].reshape(-1, 5, 2, 33, 15) # (N, 5, 2, 33, 15)
best_plan_idx = np.argmax(plan_probs, axis=1)[0] # (N,)
best_plan = plans[:, best_plan_idx, ...] # (N, 2, 33, 15)
# lane lines
lane_lines = outputs[:, lanes_start_idx:lanes_end_idx] # (N, 528)
lane_lines_deflat = lane_lines.reshape((-1, 2, 264)) # (N, 2, 264)
lane_lines_means = lane_lines_deflat[:, 0, :] # (N, 264)
lane_lines_means = lane_lines_means.reshape(-1, 4, 33, 2) # (N, 4, 33, 2)
outer_left_lane = lane_lines_means[:, 0, :, :] # (N, 33, 2)
inner_left_lane = lane_lines_means[:, 1, :, :] # (N, 33, 2)
inner_right_lane = lane_lines_means[:, 2, :, :] # (N, 33, 2)
outer_right_lane = lane_lines_means[:, 3, :, :] # (N, 33, 2)
# lane lines probs
lane_lines_probs = outputs[:, lane_lines_prob_start_idx:lane_lines_prob_end_idx] # (N, 8)
lane_lines_probs = lane_lines_probs.reshape((-1, 4, 2)) # (N, 4, 2)
lane_lines_probs = sigmoid(lane_lines_probs[:, :, 1]) # (N, 4), 0th is deprecated
outer_left_prob = lane_lines_probs[:, 0] # (N,)
inner_left_prob = lane_lines_probs[:, 1] # (N,)
inner_right_prob = lane_lines_probs[:, 2] # (N,)
outer_right_prob = lane_lines_probs[:, 3] # (N,)
# road edges
road_edges = outputs[:, road_start_idx:road_end_idx]
road_edges_deflat = road_edges.reshape((-1, 2, 132)) # (N, 2, 132)
road_edge_means = road_edges_deflat[:, 0, :].reshape(-1, 2, 33, 2) # (N, 2, 33, 2)
road_edge_stds = road_edges_deflat[:, 1, :].reshape(-1, 2, 33, 2) # (N, 2, 33, 2)
left_edge = road_edge_means[:, 0, :, :] # (N, 33, 2)
right_edge = road_edge_means[:, 1, :, :]
left_edge_std = road_edge_stds[:, 0, :, :] # (N, 33, 2)
right_edge_std = road_edge_stds[:, 1, :, :]
batch_size = best_plan.shape[0]
result_batch = []
for i in range(batch_size):
lanelines = [outer_left_lane[i], inner_left_lane[i], inner_right_lane[i], outer_right_lane[i]]
lanelines_probs = [outer_left_prob[i], inner_left_prob[i], inner_right_prob[i], outer_right_prob[i]]
road_edges = [left_edge[i], right_edge[i]]
road_edges_probs = [left_edge_std[i], right_edge_std[i]]
if best_plan_only:
plan = best_plan[i]
else:
plan = (plans[i], plan_probs[i])
result_batch.append(((lanelines, lanelines_probs), (road_edges, road_edges_probs), plan))
return result_batch
def get_transform_matrix(base_img,
augment_trans=np.array([0, 0, 0]),
augment_eulers=np.array([0, 0, 0]),
from_intr=eon_intrinsics,
to_intr=eon_intrinsics,
output_size=None,
h=1.22):
size = base_img.shape[:2]
if not output_size:
output_size = size[::-1]
cy = from_intr[1, 2]
quadrangle = np.array([[0, cy + 20],
[size[1] - 1, cy + 20],
[0, size[0] - 1],
[size[1] - 1, size[0] - 1]], dtype=np.float32)
quadrangle_norm = np.hstack((normalize(quadrangle, intrinsics=from_intr), np.ones((4, 1))))
quadrangle_world = np.column_stack((h * quadrangle_norm[:, 0] / quadrangle_norm[:, 1],
h * np.ones(4),
h / quadrangle_norm[:, 1]))
rot = orient.rot_from_euler(augment_eulers)
to_extrinsics = np.hstack((rot.T, -augment_trans[:, None]))
to_KE = to_intr.dot(to_extrinsics)
warped_quadrangle_full = np.einsum('jk,ik->ij', to_KE, np.hstack((quadrangle_world, np.ones((4, 1)))))
warped_quadrangle = np.column_stack((warped_quadrangle_full[:, 0] / warped_quadrangle_full[:, 2],
warped_quadrangle_full[:, 1] / warped_quadrangle_full[:, 2])).astype(
np.float32)
M = cv2.getPerspectiveTransform(quadrangle, warped_quadrangle.astype(np.float32))
return M
def transform_img(base_img,
augment_trans=np.array([0, 0, 0]),
augment_eulers=np.array([0, 0, 0]),
from_intr=eon_intrinsics,
to_intr=eon_intrinsics,
output_size=None,
pretransform=None,
top_hacks=False,
yuv=False,
alpha=1.0,
beta=0,
blur=0):
# import cv2 # pylint: disable=import-error
cv2.setNumThreads(1)
if yuv:
base_img = cv2.cvtColor(base_img, cv2.COLOR_YUV2RGB_I420)
size = base_img.shape[:2]
if not output_size:
output_size = size[::-1]
cy = from_intr[1, 2]
def get_M(h=1.22):
quadrangle = np.array([[0, cy + 20],
[size[1] - 1, cy + 20],
[0, size[0] - 1],
[size[1] - 1, size[0] - 1]], dtype=np.float32)
quadrangle_norm = np.hstack((normalize(quadrangle, intrinsics=from_intr), np.ones((4, 1))))
quadrangle_world = np.column_stack((h * quadrangle_norm[:, 0] / quadrangle_norm[:, 1],
h * np.ones(4),
h / quadrangle_norm[:, 1]))
rot = orient.rot_from_euler(augment_eulers)
to_extrinsics = np.hstack((rot.T, -augment_trans[:, None]))
to_KE = to_intr.dot(to_extrinsics)
warped_quadrangle_full = np.einsum('jk,ik->ij', to_KE, np.hstack((quadrangle_world, np.ones((4, 1)))))
warped_quadrangle = np.column_stack((warped_quadrangle_full[:, 0] / warped_quadrangle_full[:, 2],
warped_quadrangle_full[:, 1] / warped_quadrangle_full[:, 2])).astype(
np.float32)
M = cv2.getPerspectiveTransform(quadrangle, warped_quadrangle.astype(np.float32))
return M
M = get_M()
if pretransform is not None:
M = M.dot(pretransform)
augmented_rgb = cv2.warpPerspective(base_img, M, output_size, borderMode=cv2.BORDER_REPLICATE)
if top_hacks:
cyy = int(math.ceil(to_intr[1, 2]))
M = get_M(1000)
if pretransform is not None:
M = M.dot(pretransform)
augmented_rgb[:cyy] = cv2.warpPerspective(base_img, M, (output_size[0], cyy), borderMode=cv2.BORDER_REPLICATE)
# brightness and contrast augment
# augmented_rgb = np.clip((float(alpha)*augmented_rgb + beta), 0, 255).astype(np.uint8)
# print('after clip:', augmented_rgb.shape, augmented_rgb.dtype)
# gaussian blur
if blur > 0:
augmented_rgb = cv2.GaussianBlur(augmented_rgb, (blur * 2 + 1, blur * 2 + 1), cv2.BORDER_DEFAULT)
if yuv:
augmented_img = cv2.cvtColor(augmented_rgb, cv2.COLOR_RGB2YUV_I420)
else:
augmented_img = augmented_rgb
return augmented_img
def reshape_yuv(frames):
H = (frames.shape[1] * 2) // 3
W = frames.shape[2]
in_img1 = np.zeros((frames.shape[0], 6, H // 2, W // 2), dtype=np.uint8)
in_img1[:, 0] = frames[:, 0:H:2, 0::2]
in_img1[:, 1] = frames[:, 1:H:2, 0::2]
in_img1[:, 2] = frames[:, 0:H:2, 1::2]
in_img1[:, 3] = frames[:, 1:H:2, 1::2]
in_img1[:, 4] = frames[:, H:H + H // 4].reshape((-1, H // 2, W // 2))
in_img1[:, 5] = frames[:, H + H // 4:H + H // 2].reshape((-1, H // 2, W // 2))
return in_img1
# color space conversion methods
def bgr_to_yuv(img_bgr):
"""
Model accepts images in the YUV color space. This function converts a BGR (blue, green, red) to YUV420
See: https://github.com/commaai/openpilot/tree/master/selfdrive/modeld/models
:param img_bgr: BGR Image (numpy array)
:return: image in YUV420 color space
"""
img_yuv = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2YUV_I420)
assert img_yuv.shape == ((874 * 3 // 2, 1164))
return img_yuv
def bgr_to_rgb(bgr):
return cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
def yuv_to_rgb(yuv):
return cv2.cvtColor(yuv, cv2.COLOR_YUV2RGB_I420)
def rgb_to_yuv(rgb):
return cv2.cvtColor(rgb, cv2.COLOR_RGB2YUV_I420)
def transform_frames(frames):
"""
This function transforms input such that they are acceptable by the model.
See: https://github.com/commaai/openpilot/tree/master/selfdrive/modeld/models
:param frames: A list with two consecutive frames captured at 20 Hz.
:return: Transformed images to feed into openpilot model input frame tensor.
"""
imgs_med_model = np.zeros((len(frames), 384, 512), dtype=np.uint8)
for i, img in enumerate(frames):
imgs_med_model[i] = transform_img(img,
from_intr=eon_intrinsics,
to_intr=medmodel_intrinsics,
yuv=True,
output_size=(512, 256))
reshaped = reshape_yuv(imgs_med_model)
return reshaped
class Calibration:
""" The Calibration class provides functionality to map points between model space (meters) to image space and vice
versa (pixels).
"""
def __init__(self, rpy, intrinsic=eon_intrinsics, plot_img_width=640, plot_img_height=480):
self.intrinsic = intrinsic
self.extrinsics_matrix = get_view_frame_from_calib_frame(rpy[0], rpy[1], rpy[2], 1.22)[:, :3]
self.plot_img_width = plot_img_width
self.plot_img_height = plot_img_height
self.zoom = 1
self.CALIB_BB_TO_FULL = np.asarray([
[self.zoom, 0., 0.],
[0., self.zoom, 0.],
[0., 0., 1.]])
def car_space_to_ff(self, x, y, z):
car_space_projective = np.column_stack((x, y, z)).T
ep = self.extrinsics_matrix.dot(car_space_projective)
kep = self.intrinsic.dot(ep)
# TODO: fix numerical instability (add 1e-16)
# UPD: this turned out to slow things down a lot. How do we do it then?
return (kep[:-1, :] / kep[-1, :]).T
def car_space_to_bb(self, x, y, z):
pts = self.car_space_to_ff(x, y, z)
return pts / self.zoom
def project_path(path, calibration, z_off):
'''Projects paths from calibration space (model input/output) to image space.'''
x = path[:, 0]
y = path[:, 1]
z = path[:, 2] + z_off
pts = calibration.car_space_to_bb(x, y, z)
# filter out invalid points
pts[pts < 0] = np.nan
valid = np.isfinite(pts).all(axis=1)
pts = pts[valid].astype(int)
return pts
def create_image_canvas(img_rgb, zoom_matrix, plot_img_height, plot_img_width):
'''Transform with a correct warp/zoom transformation.'''
img_plot = np.zeros((plot_img_height, plot_img_width, 3), dtype='uint8')
cv2.warpAffine(img_rgb, zoom_matrix[:2], (img_plot.shape[1], img_plot.shape[0]), dst=img_plot,
flags=cv2.WARP_INVERSE_MAP)
return img_plot
def draw_path(lane_lines, road_edges, path_plan, img_plot, calibration, lane_line_color_list, width=1, height=1.22,
fill_color=(128, 0, 255), line_color=(0, 255, 0)):
'''Draw model predictions on an image.'''
overlay = img_plot.copy()
alpha = 0.4
fixed_distances = np.array(X_IDXs)[:, np.newaxis]
# lane_lines are sequentially parsed ::--> means--> std's
if lane_lines is not None:
(oll, ill, irl, orl), (oll_prob, ill_prob, irl_prob, orl_prob) = lane_lines
calib_pts_oll = np.hstack((fixed_distances, oll)) # (33, 3)
calib_pts_ill = np.hstack((fixed_distances, ill)) # (33, 3)
calib_pts_irl = np.hstack((fixed_distances, irl)) # (33, 3)
calib_pts_orl = np.hstack((fixed_distances, orl)) # (33, 3)
img_pts_oll = project_path(calib_pts_oll, calibration, z_off=0).reshape(-1, 1, 2)
img_pts_ill = project_path(calib_pts_ill, calibration, z_off=0).reshape(-1, 1, 2)
img_pts_irl = project_path(calib_pts_irl, calibration, z_off=0).reshape(-1, 1, 2)
img_pts_orl = project_path(calib_pts_orl, calibration, z_off=0).reshape(-1, 1, 2)
lane_lines_with_probs = [(img_pts_oll, oll_prob), (img_pts_ill, ill_prob), (img_pts_irl, irl_prob),
(img_pts_orl, orl_prob)]
line_overlay = overlay.copy()
# plot lanelines
for i, (line_pts, prob) in enumerate(lane_lines_with_probs):
cv2.polylines(line_overlay, [line_pts], False, lane_line_color_list[i], thickness=2)
img_plot = line_overlay
# road edges
if road_edges is not None:
(left_road_edge, right_road_edge), _ = road_edges
calib_pts_ledg = np.hstack((fixed_distances, left_road_edge))
calib_pts_redg = np.hstack((fixed_distances, right_road_edge))
img_pts_ledg = project_path(calib_pts_ledg, calibration, z_off=0).reshape(-1, 1, 2)
img_pts_redg = project_path(calib_pts_redg, calibration, z_off=0).reshape(-1, 1, 2)
# plot road_edges
cv2.polylines(overlay, [img_pts_ledg], False, (255, 128, 0), thickness=2)
cv2.polylines(overlay, [img_pts_redg], False, (255, 234, 0), thickness=2)
# path plan
if path_plan is not None:
path_plan_l = path_plan - np.array([0, 1, 0])
path_plan_r = path_plan + np.array([0, 1, 0])
img_pts_l = project_path(path_plan_l, calibration, z_off=height)
img_pts_r = project_path(path_plan_r, calibration, z_off=height)
for i in range(1, len(img_pts_l)):
if i >= len(img_pts_r): break
u1, v1, u2, v2 = np.append(img_pts_l[i - 1], img_pts_r[i - 1])
u3, v3, u4, v4 = np.append(img_pts_l[i], img_pts_r[i])
pts = np.array([[u1, v1], [u2, v2], [u4, v4], [u3, v3]], np.int32).reshape((-1, 1, 2))
cv2.fillPoly(overlay, [pts], fill_color)
cv2.polylines(overlay, [pts], True, line_color)
# drawing the plots on original iamge
img_plot = cv2.addWeighted(overlay, alpha, img_plot, 1 - alpha, 0)
return img_plot