forked from Ronnypetson/tfoe_vo
-
Notifications
You must be signed in to change notification settings - Fork 0
/
kitti_init.py
300 lines (258 loc) · 10.9 KB
/
kitti_init.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
import os
import sys
import time
import numpy as np
import cv2
import torch
from liegroups import SE3
#from matplotlib import pyplot as plt
from opt_ba import OptSingle
import pykitti
from versions import files_to_hash, save_state
from utils import norm_t, plot_trajs, ba_graph
from utils import save_poses, pt_cloud, plot_pt_cloud
class KpT0_BA:
''' Iterator class for returning key-points and pose initialization
'''
def __init__(self, h, w, basedir, seq, fast_th=25):
super().__init__()
self.size = (h, w)
self.kitti = pykitti.odometry(basedir, seq)
self.gt_odom = self.kitti.poses
self.seq_len = len(self.gt_odom)
self.camera_matrix = np.array([[718.8560, 0.0, 607.1928],
[0.0, 718.8560, 185.2157],
[0.0, 0.0, 1.0]])
self.c_ = np.linalg.inv(self.camera_matrix)
self.feature_detector = cv2.FastFeatureDetector_create(threshold=fast_th,
nonmaxSuppression=True)
self.lk_params = dict(winSize=(21, 21),
criteria=(cv2.TERM_CRITERIA_EPS |
cv2.TERM_CRITERIA_COUNT, 30, 0.03))
self.bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
self._kpts = {} # keypoint cache, i -> kp
self._kptsij = {} # [i,j] -> kpij
self._flow = {} # [i,j] -> f
self._T0 = {} # i -> T
self._ep0 = {} # i -> ep
self._Tgt = {} # i -> Tgt
self._gTgt = {} # i -> gTgt
self._vids = {} # [i,j] -> vid
self._avids = {} # [i,j] -> avid
self._moving = {} # i -> moving
self._Tij0 = {} # [i,j] -> Tij
def init_frame(self, i):
if i not in self._T0:
i_ = min(i+1, self.seq_len-1)
im0 = self.kitti.get_cam0(i)
im0 = np.array(im0)
im1 = self.kitti.get_cam0(i_)
im1 = np.array(im1)
kp0 = self.feature_detector.detect(im0, None)
kp0 = np.array([[x.pt] for x in kp0], dtype=np.float32)
try:
p1, st, err = cv2.calcOpticalFlowPyrLK(im0, im1, kp0,
None, **self.lk_params)
E, mask = cv2.findEssentialMat(p1, kp0, self.camera_matrix,
cv2.RANSAC, 0.999, 0.1, None)
vids = [j for j in range(len(mask)) if mask[j] == 1.0]
avids = vids
_, R, t, mask = cv2.recoverPose(E, p1, kp0, self.camera_matrix, mask=mask)
T0 = np.eye(4)
T0[:3, :3] = R
T0[:3, 3:] = t
if len(mask) > 3:
vids_ = [j for j in range(len(mask)) if mask[j] == 1.0]
if len(vids_) > 3:
vids = vids_
if np.mean(np.abs(p1[vids] - kp0[vids])) < 1e-2:
T0 = np.eye(4)
moving = False
else:
moving = True
ep2 = T0[:3, 3:]
ep2 = ep2 / (ep2[-1] + 1e-8)
ep2 = self.camera_matrix @ ep2
ep0 = ep2[:2, 0]
inert = np.linalg.inv(self.gt_odom[i])
Tgt = inert @ self.gt_odom[i_]
#norm_gt = np.linalg.norm(Tgt[:3, 3:])
#T0 = norm_t(T0.copy(), norm_gt)
self._kpts[i] = kp0
self._flow[(i, i_)] = p1 - kp0
self._T0[i] = T0
self._ep0[i] = ep0
self._Tgt[i] = Tgt
if i > 0:
self._gTgt[i] = T0 @ self._gTgt[i-1] #self.gt_odom[i_]
else:
self._gTgt[i] = T0
self._vids[(i, i_)] = vids
self._avids[(i, i_)] = avids
self._moving[i] = moving
except Exception as e:
print(e)
raise e
def init_BA(self, i, j):
msg = f'Check frame indexes: {i} {j}'
assert 0 <= i < self.seq_len, msg
assert 0 <= j < self.seq_len, msg
assert i != j, msg
self.init_frame(i)
self.init_frame(j)
kp0 = self._kpts[i]
if (i, j) not in self._flow:
im_i = self.kitti.get_cam0(i)
im_i = np.array(im_i)
im_j = self.kitti.get_cam0(j)
im_j = np.array(im_j)
#kp0 = self.feature_detector.detect(im_i, None)
#kp0 = np.array([[x.pt] for x in kp0], dtype=np.float32)
kp1, st, err = cv2.calcOpticalFlowPyrLK(im_i, im_j, kp0,
None, **self.lk_params)
E, mask = cv2.findEssentialMat(kp1, kp0, self.camera_matrix,
cv2.RANSAC, 0.999, 0.1, None)
vids = [k for k in range(len(mask)) if mask[k] == 1.0]
avids = vids
_, R, t, mask = cv2.recoverPose(E, kp1, kp0, self.camera_matrix, mask=mask)
#T0 = np.eye(4)
#T0[:3, :3] = R
#T0[:3, 3:] = t
if len(mask) > 3:
vids_ = [k for k in range(len(mask)) if mask[k] == 1.0]
if len(vids_) > 3:
vids = vids_
#self._kptsij[(i, j)] = kp0
self._flow[(i, j)] = kp1 - kp0
self._vids[(i, j)] = vids
self._avids[(i, j)] = avids
#self._Tij0[(i, j)] = T0
return kp0, self._flow[(i, j)]
def main():
seq_id = sys.argv[1]
run_date = time.asctime().replace(' ', '_')
state_fns = ['kitti_init.py', 'opt_ba.py', 'utils.py', 'reproj.py']
run_hash = files_to_hash(state_fns)
run_dir = f'odom/{run_hash}/{run_date}/'
save_state('odom/', state_fns)
if not os.path.isdir(run_dir):
os.makedirs(run_dir)
bdir = '/home/ronnypetson/Downloads/kitti_seq/dataset/'
h, w = 376, 1241
kp = KpT0_BA(h, w, bdir, seq_id, fast_th=15)
c = kp.camera_matrix
c_ = np.linalg.inv(c)
failure_eps = 1e-7
poses = []
poses_gt = []
poses_ = []
show_cloud = False
pose0 = np.eye(4)
W_poses = []
cloud_all = np.zeros((3, 1))
gT = np.zeros((kp.seq_len+1, 6))
ge = np.zeros((kp.seq_len+1, 2))
ge[0] = np.array([607.1928, 185.2157]) / 1e3
baw = 3
try:
for i in range(kp.seq_len):
i_ = min(i+1, kp.seq_len-1)
kp.init_frame(i)
Tgt = kp._Tgt[i]
T = kp._T0[i]
g = ba_graph(i, i+(baw-1))
p = {}
f = {}
for ij in g:
kp.init_BA(ij[0], ij[1])
p[ij] = kp._kpts[ij[0]]
f[ij] = kp._flow[ij]
normT = np.linalg.norm(Tgt[:3, 3])
poses.append(norm_t(T.copy(), normT))
poses_gt.append(Tgt)
x = {}
x_ = {}
for ij in g:
x[ij] = p[ij][kp._vids[ij], 0, :].transpose(1, 0) #
z = np.ones((1, x[ij].shape[-1]))
x[ij] = np.concatenate([x[ij], z], axis=0)
x_[ij] = (p[ij] + f[ij])[kp._vids[ij], 0, :].transpose(1, 0) #
x_[ij] = np.concatenate([x_[ij], z], axis=0)
opt = OptSingle(x, x_, c, g) # kp.E
T0 = SE3.from_matrix(T, normalize=True)
#T0 = T0.inv().log()
T0 = T0.log()
#T00 = SE3.from_matrix(kp._T0[i+1], normalize=True).log()
#T00 = SE3.from_matrix(kp._gTgt[i+1], normalize=True).log()
for j in range(baw):
gT[i + j + 1] = SE3.from_matrix(kp._gTgt[i + j], normalize=True).log()
#gT[i+1] = SE3.from_matrix(kp._gTgt[i], normalize=True).log() # i -> i-1
#gT[i+2] = SE3.from_matrix(kp._gTgt[i+1], normalize=True).log() ###
#gT[i+3] = SE3.from_matrix(kp._gTgt[i+2], normalize=True).log()
#gT[i+4] = SE3.from_matrix(kp._gTgt[i+3], normalize=True).log()
#T0 = np.zeros(6)
#foe0 = kp._ep0[i] / 1e3
for j in range(baw):
ge[i+j+1] = kp._ep0[i+j] / 1e3
#ge[i+2] = kp._ep0[i+1] / 1e3 ###
#ge[i+3] = kp._ep0[i+2] / 1e3
#ge[i+4] = kp._ep0[i+3] / 1e3
#foe0 = np.array([607.1928, 185.2157]) / 1e3
Tfoe = opt.optimize(gT, ge, freeze=False)
#Tfoe = np.zeros((ge.shape[0], 8))
print(opt.min_obj)
if opt.min_obj > failure_eps and False:
print('Initialization failure.')
x = p[kp._avids[i], 0, :].transpose(1, 0) # kp.avids
z = np.ones((1, x.shape[-1]))
x = np.concatenate([x, z], axis=0)
x_ = p + f
x_ = x_[kp._avids[i], 0, :].transpose(1, 0) # kp.avids
x_ = np.concatenate([x_, z], axis=0)
opt = OptSingle(x, x_, c, None)
T0 = np.zeros(6)
#foe0 = np.array([w/2.0, h/2.0])
foe0 = np.array([607.1928, 185.2157]) / 1e3
Tfoe = opt.optimize(T0, foe0, freeze=False)
print(f'New x0 status: {opt.min_obj <= failure_eps}')
Tfoe = Tfoe.reshape(-1, 8)
T_ = Tfoe[i+1, :6]
foe = Tfoe[i+1, 6:]
#print(gT[:i+5])
#print(Tfoe[:i+5, :6])
#input()
#print('diff', np.linalg.norm(gT[:i + 5]-Tfoe[:i + 5, :6]))
gT[i+1] = T_.copy()
#T_ = gT[i]
print(foe)
T_ = SE3.exp(T_).as_matrix() # .inv()
if i > 0:
T_ = T_ @ np.linalg.inv(SE3.exp(gT[i]).as_matrix())
#T_ = np.linalg.inv(SE3.exp(gT[i]).as_matrix()) @ T_
#T_ = norm_t(T_, normT)
poses_.append(norm_t(T_.copy(), normT))
#poses_.append(T_.copy())
pose0 = pose0 @ T_
W_poses.append(pose0)
if i % 10 == 9:
P = [poses_gt, poses, poses_]
#P = [poses_gt, poses, W_poses]
plot_trajs(P, f'{run_dir}/{seq_id}.svg', glb=False)
if show_cloud:
scale = normT / (np.linalg.norm(T_[:3, 3])+1e-8)
p = torch.from_numpy(p[kp._vids[i], 0]).double()
p_ = p + torch.from_numpy(f[kp._vids[i], 0]).double()
T_acc = torch.from_numpy(pose0).double()
foe = torch.from_numpy(foe).double().unsqueeze(-1)
c_tc = torch.from_numpy(c).double()
cloud = pt_cloud(p, p_, T_acc, foe, scale, c_tc, T_)
cloud_all = np.concatenate([cloud_all, cloud], axis=1) # [:,:20]
if i % 10 == 9 and show_cloud:
plot_pt_cloud(np.array(cloud_all), f'{run_dir}/{seq_id}_pt_cloud.svg')
i += 1
save_poses(W_poses, f'{run_dir}/KITTI_{seq_id}.txt')
except KeyboardInterrupt as e:
save_poses(W_poses, f'{run_dir}/KITTI_{seq_id}.txt')
raise e
if __name__ == '__main__':
main()