-
Notifications
You must be signed in to change notification settings - Fork 4
/
VIO_KITTI_learn.m
400 lines (345 loc) · 12.6 KB
/
VIO_KITTI_learn.m
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
disp('===========================');
clear;
addpath('libviso2');
addpath('datasets/extraction/utils');
addpath('datasets/extraction/utils/devkit');
addpath('utils');
addpath('learning');
dataBaseDir = '/Users/valentinp/Desktop/KITTI/2011_09_26/2011_09_26_drive_0029_sync';
dataCalibDir = '/Users/valentinp/Desktop/KITTI/2011_09_26';
%% Get ground truth and import data
frameRange = 1:300;
%%Image data
leftImageData = loadImageData([dataBaseDir '/image_00'], frameRange);
rightImageData = loadImageData([dataBaseDir '/image_01'], frameRange);
%IMU data
[imuData, imuFrames] = loadImuData(dataBaseDir, leftImageData.timestamps);
%Ground Truth
T_wIMU_GT = getGroundTruth(dataBaseDir, imuFrames);
numFrames = size(leftImageData.rectImages, 3);
skipFrames = 1;
%% Load calibration and parameters
[T_camvelo_struct, P_rect_cam1] = loadCalibration(dataCalibDir);
T_camvelo = T_camvelo_struct{1};
T_veloimu = loadCalibrationRigid(fullfile(dataCalibDir,'calib_imu_to_velo.txt'));
T_camimu = T_camvelo*T_veloimu;
%Add camera ground truth
T_wCam_GT = T_wIMU_GT;
for i = 1:size(T_wIMU_GT, 3)
T_wCam_GT(:,:,i) = T_wIMU_GT(:,:,i)*inv(T_camimu);
end
K= P_rect_cam1(:,1:3);
b_pix = P_rect_cam1(1,4);
cu = K(1,3);
cv = K(2,3);
fu = K(1,1);
fv = K(2,2);
b = -b_pix/fu; %The KITTI calibration supplies the baseline in units of pixels
calibParams.c_u = cu;
calibParams.c_v = cv;
calibParams.f_u = fu;
calibParams.f_v = fv;
calibParams.b = b;
%LIBVISO2 matching
param.f = fu;
param.cu = cu;
param.cv = cv;
param.base = b;
param.nms_n = 5; % non-max-suppression: min. distance between maxima (in pixels)
param.nms_tau = 50; % non-max-suppression: interest point peakiness threshold
param.match_binsize = 50; % matching bin width/height (affects efficiency only)
param.match_radius = 200; % matching radius (du/dv in pixels)
param.match_disp_tolerance = 1; % du tolerance for stereo matches (in pixels)
param.outlier_disp_tolerance = 5; % outlier removal: disparity tolerance (in pixels)
param.outlier_flow_tolerance = 5; % outlier removal: flow tolerance (in pixels)
param.multi_stage = 1; % 0=disabled,1=multistage matching (denser and faster)
param.half_resolution = 1; % 0=disabled,1=match at half resolution, refine at full resolution
param.refinement = 0; % refinement (0=none,1=pixel,2=subpixel)
%% Setup
addpath('settings');
addpath('utils');
R = diag(4*ones(4,1));
optParams.RANSACCostThresh = 0.2;
optParams.maxGNIter = 10;
optParams.lineLambda = 0.75;
optParams.LMlambda = 1e-5;
%% Get ground truth relative steps
T_21_gt = NaN(4,4, length(leftImageData.timestamps)-1);
for j = 2:skipFrames:numFrames
T_21_gt(:,:,j-1) = inv(T_wCam_GT(:,:, j))*T_wCam_GT(:,:, j-1);
end
%% Main loop
% create figure
% figure('Color',[1 1 1]);
% ha1 = axes('Position',[0.05,0.7,0.9,0.25]);
% %axis off;
% ha2 = axes('Position',[0.05,0.05,0.9,0.6]);
% axis equal, grid on, hold on;
repeatIter = 10;
learnedPredSpace.predVectors = [];
learnedPredSpace.weights = [];
p_wcam_hist = NaN(3,length(frameRange), repeatIter);
rng('shuffle');
for repeat_i = 1:repeatIter
usedPredVectors = [];
% init matcher
matcherMex('init',param);
% push back first images
I1prev = uint8(leftImageData.rectImages(:,:,1));
I2prev = uint8(rightImageData.rectImages(:,:,1));
matcherMex('push',I1prev,I2prev);
numFrames = size(leftImageData.rectImages, 3);
k =1;
T_wcam = eye(4);
T_wcam_hist = T_wcam;
p_wcam_hist(:, 1, repeat_i) = zeros(3,1);
% history variables
firstState.C_vi = eye(3);
firstState.r_vi_i = zeros(3,1);
firstState.k = 1;
oldState = firstState;
for frame=2:skipFrames:numFrames
%IMU Data
imuMeasurement.omega = imuData.measOmega(:, frame-1);
imuMeasurement.v = imuData.measVel(:, frame-1);
deltaT = imuData.timestamps(frame) - imuData.timestamps(frame-1);
%Get an estimate through IMU propagation
newState = propagateState(oldState, imuMeasurement, deltaT);
T_21_imu = getTransformation(oldState, newState);
T_21_cam = T_camimu*T_21_imu*inv(T_camimu);
oldState = newState;
% read current images
I1 = uint8(leftImageData.rectImages(:,:,frame));
I2 = uint8(rightImageData.rectImages(:,:,frame));
%Plot image
% axes(ha1); cla;
% imagesc(I1);
% axis off;
matcherMex('push',I1,I2);
% match images
matcherMex('match',2);
p_matched = matcherMex('get_matches',2);
% show matching results
disp(['Number of matched points: ' num2str(length(p_matched))]);
%Triangulate points and prune any at Infinity
[p_f1_1, p_f2_2] = triangulateAllPointsDirect(p_matched, calibParams);
pruneId = isinf(p_f1_1(1,:)) | isinf(p_f1_1(2,:)) | isinf(p_f1_1(3,:)) | isinf(p_f2_2(1,:)) | isinf(p_f2_2(2,:)) | isinf(p_f2_2(3,:));
p_f1_1 = p_f1_1(:, ~pruneId);
p_f2_2 = p_f2_2(:, ~pruneId);
%Select a subset
%selectIdx = randperm(size(p_f1_1,2), 5);
numLm = size(p_f1_1,2);
skip = floor(numLm/repeatIter);
buckets = [1:skip:numLm - mod(numLm, skip); skip:skip:numLm];
selectIdx = buckets(1, repeat_i):buckets(2, repeat_i);
rng(42); randOrder = randperm(numLm,numLm);
selectIdx = randOrder(selectIdx);
p_f1_1 = p_f1_1(:, selectIdx);
p_f2_2 = p_f2_2(:, selectIdx);
p_matched = p_matched(:, selectIdx);
inliers = 1:size(p_f1_1,2);
%Find inliers based on rotation matrix from IMU
%[p_f1_1, p_f2_2, T_21_est, inliers] = findInliersRot(p_f1_1, p_f2_2, T_21_cam(1:3,1:3), optParams);
[predVectors] = computePredVectors( p_matched(1:2,inliers), I1, I1prev, [imuData.measAccel(:, frame); imuData.measOmega(:, frame-1)]);
%usedPredVectors(:,end+1:end+size(predVectors, 2)) = predVectors;
%fprintf('Tracking %d features.', size(p_f1_1,2));
%Calculate initial guess using scalar weights, then use matrix weighted
%non linear optimization
R_1 = repmat(R, [1 1 size(p_f1_1, 2)]);
R_2 = R_1;
T_21_est = scalarWeightedPointCloudAlignment(p_f1_1, p_f2_2,T_21_cam(1:3,1:3));
T_21_opt = matrixWeightedPointCloudAlignment(p_f1_1, p_f2_2, R_1, R_2, T_21_est, calibParams, optParams);
transError = T_21_opt(1:3,4) - T_21_gt(1:3,4,k);
RMSE = sqrt(sum(transError.^2/3))
%Update the prediction space learning
%Prune any vectors with 0 entries
pruneWeightsIdx = predVectors(1,:) == 0 | predVectors(2,:) == 0 | predVectors(3,:) == 0 | predVectors(4,:) == 0 | predVectors(5,:) == 0 | predVectors(6,:) == 0 | predVectors(7,:) == 0;
predVectors = predVectors(:,~pruneWeightsIdx);
learnedPredSpace.predVectors = [learnedPredSpace.predVectors predVectors];
learnedPredSpace.weights = [learnedPredSpace.weights RMSE*ones(1, size(predVectors,2))];
T_wcam = T_wcam*inv(T_21_opt);
T_wcam_hist(:,:,end+1) = T_wcam;
p_wcam_hist(:, frame, repeat_i) = T_wcam(1:3,4);
% update trajectory and plot
% axes(ha2);
% plot(T_wcam(1,4),T_wcam(3,4),'g*');
% hold on;
% grid on;
% drawnow();
I1prev = I1;
I2prev = I2;
k = k + 1;
fprintf('k:%d, repeat_i: %d \n',k,repeat_i);
end
% close matcher
matcherMex('close');
p_vi_i = NaN(3, size(T_wCam_GT,3));
for j = frameRange
T_wcam_gt = inv(T_wCam_GT(:,:,1))*T_wCam_GT(:,:, j);
p_vi_i(:,j) = T_wcam_gt(1:3,4);
end
translation = NaN(3, size(T_wcam_hist, 3));
for i = 1:size(T_wcam_hist, 3)
T_wcam = T_wcam_hist(:, :, i);
translation(:,i) = T_wcam(1:3, 4);
end
%Plot error and variances
transErrVec = zeros(3, length(frameRange));
for i = frameRange
transErrVec(:,i) = translation(:, i) - p_vi_i(:,i);
end
meanRMSE = mean(sqrt(sum(transErrVec.^2,1)/3))
end
learnedPredSpace
%%
%% Plot trajectories
totalDist = 0;
p_wcam_w_gt = NaN(3, size(T_wCam_GT,3));
for j = frameRange
if j > 1
T_12 = inv(T_wCam_GT(:,:,j-1))*T_wCam_GT(:,:, j);
totalDist = totalDist + norm(T_12(1:3,4));
end
T_wcam_gt = inv(T_wCam_GT(:,:,1))*T_wCam_GT(:,:, j);
p_wcam_w_gt(:,j) = T_wcam_gt(1:3,4);
end
figure
for p_i = 1:size(p_wcam_hist,3)
plot(p_wcam_hist(1,:,p_i),p_wcam_hist(3,:,p_i), '-b', 'LineWidth', 1);
hold on;
end
plot(p_wcam_w_gt(1,:),p_wcam_w_gt(3,:), '-r', 'LineWidth', 2);
f = strsplit(dataBaseDir, '/');
f = strsplit(char(f(end)), '.');
fileName = char(f(1));
title(sprintf('Training Runs \n %s', fileName), 'Interpreter', 'none')
xlabel('x [m]')
ylabel('z [m]')
%legend('Training Runs', 'Ground Truth')
grid on;
saveas(gcf,sprintf('plots/%s_training.fig', fileName));
save(sprintf('plots/%s_paths.mat', fileName), 'p_wcam_hist', 'p_wcam_w_gt');
%% Find the best exponent
% searchObject = KDTreeSearcher(learnedPredSpace.predVectors');
% refWeight = mean(learnedPredSpace.weights);
% rng('shuffle');
% gammaList = 10:2:20;
% rmseList = NaN(1, length(gammaList));
% g = 1;
% for gamma = gammaList
%
% % init matcher
% param.nms_n = 10;
% matcherMex('init',param);
% % push back first images
% I1prev = uint8(leftImageData.rectImages(:,:,1));
% I2prev = uint8(rightImageData.rectImages(:,:,1));
% matcherMex('push',I1prev,I2prev);
% numFrames = size(leftImageData.rectImages, 3);
% k = 1;
% T_wcam = eye(4);
% T_wcam_hist = T_wcam;
%
% % history variables
% firstState.C_vi = eye(3);
% firstState.r_vi_i = zeros(3,1);
% firstState.k = 1;
% oldState = firstState;
%
% h = waitbar(0, sprintf('Evaluating gamma=%d ...', gamma));
%
% for frame=2:skipFrames:numFrames
% waitbar(frame / numFrames)
% %IMU Data
% imuMeasurement.omega = imuData.measOmega(:, frame-1);
% imuMeasurement.v = imuData.measVel(:, frame-1);
% deltaT = imuData.timestamps(frame) - imuData.timestamps(frame-1);
%
% %Get an estimate through IMU propagation
% newState = propagateState(oldState, imuMeasurement, deltaT);
% T_21_imu = getTransformation(oldState, newState);
% T_21_cam = T_camimu*T_21_imu*inv(T_camimu);
% oldState = newState;
%
% % read current images
% I1 = uint8(leftImageData.rectImages(:,:,frame));
% I2 = uint8(rightImageData.rectImages(:,:,frame));
%
%
% matcherMex('push',I1,I2);
% % match images
% matcherMex('match',2);
% p_matched = matcherMex('get_matches',2);
% %Triangulate points and prune any at Infinity
% [p_f1_1, p_f2_2] = triangulateAllPointsDirect(p_matched, calibParams);
% pruneId = isinf(p_f1_1(1,:)) | isinf(p_f1_1(2,:)) | isinf(p_f1_1(3,:)) | isinf(p_f2_2(1,:)) | isinf(p_f2_2(2,:)) | isinf(p_f2_2(3,:));
% p_f1_1 = p_f1_1(:, ~pruneId);
% p_f2_2 = p_f2_2(:, ~pruneId);
%
% inliers = 1:size(p_f1_1,2);
% %Calculate weights
% [predVectors] = computePredVectors( p_matched(1:2,inliers), I1, I1prev, [imuData.measAccel(:, frame); imuData.measOmega(:, frame-1)]);
% predWeightList = getPredVectorWeight(predVectors, searchObject, learnedPredSpace.weights, refWeight, gamma);
% R_1 = [];
% thresholdPruneIdx = [];
% r_i = 1;
% for p_i = 1:length(predWeightList)
% predWeight = predWeightList(p_i);
% if predWeight > 100
% thresholdPruneIdx(end+1) = p_i;
% else
% R_1(:,:,r_i) = predWeight*R;
% r_i = r_i + 1;
% end
% end
% inliers(thresholdPruneIdx) = [];
% p_f1_1(:, thresholdPruneIdx) = [];
% p_f2_2(:, thresholdPruneIdx) = [];
% predWeightList(thresholdPruneIdx) = [];
% R_2 = R_1;
% T_21_est = scalarWeightedPointCloudAlignment(p_f1_1, p_f2_2,T_21_cam(1:3,1:3));
%
%
%
% T_21_opt = matrixWeightedPointCloudAlignment(p_f1_1, p_f2_2, R_1, R_2, T_21_est, calibParams, optParams);
%
% T_wcam = T_wcam*inv(T_21_opt);
% T_wcam_hist(:,:,end+1) = T_wcam;
%
% I1prev = I1;
% I2prev = I2;
% k = k + 1;
% end
%
% % close matcher
% matcherMex('close');
%
% p_vi_i = NaN(3, size(T_wCam_GT,3));
% for j = frameRange
% T_wcam_gt = inv(T_wCam_GT(:,:,1))*T_wCam_GT(:,:, j);
% p_vi_i(:,j) = T_wcam_gt(1:3,4);
% end
% translation = NaN(3, size(T_wcam_hist, 3));
% for i = 1:size(T_wcam_hist, 3)
% T_wcam = T_wcam_hist(:, :, i);
% translation(:,i) = T_wcam(1:3, 4);
% end
%
% %Plot error and variances
% transErrVec = zeros(3, length(frameRange));
% for i = frameRange
% transErrVec(:,i) = translation(:, i) - p_vi_i(:,i);
% end
% meanRMSE = mean(sqrt(sum(transErrVec.^2,1)/3));
% fprintf('gamma: %d, ARMSE: %.3f \n',gamma,meanRMSE);
% rmseList(g) = meanRMSE;
% close(h);
% g = g + 1;
% end
%% Export the entire model
%[~,minIdx] = min(rmseList);
learnedPredSpace.gamma = 16%gammaList(minIdx);
f = strsplit(dataBaseDir, '/');
f = strsplit(char(f(end)), '.');
fileName = ['learnedProbeModels/' char(f(1)) '_learnedPredSpaceIter' int2str(repeatIter) 'StepVar.mat'];
save(fileName, 'learnedPredSpace');