def locally_optimize(self,
kf_ref,
verbose=False,
rounds=10,
abort_flag=g2o.Flag()):
keyframes, points, ref_keyframes = self.local_map.update(kf_ref)
print('local optimization window: ',
sorted([kf.id for kf in keyframes]))
print(' refs: ',
sorted([kf.id for kf in ref_keyframes]))
print(' #points: ', len(points))
#print(' points: ', sorted([p.id for p in points]))
#err = optimizer_g2o.optimize(frames, points, None, False, verbose, rounds)
err, ratio_bad_observations = optimizer_g2o.local_bundle_adjustment(
keyframes,
points,
ref_keyframes,
False,
verbose,
rounds,
abort_flag=abort_flag,
map_lock=self.update_lock)
Printer.green('local optimization - perc bad observations: %.2f %%' %
(ratio_bad_observations * 100))
return err
def __init__(
self,
num_features=kMinNumFeatureDefault,
num_levels=3, # number of pyramid levels for detector
scale_factor=1.2, # detection scale factor (if it can be set, otherwise it is automatically computed)
detector_type=FeatureDetectorTypes.FAST,
descriptor_type=FeatureDescriptorTypes.NONE,
match_ratio_test=kRatioTest,
tracker_type=FeatureTrackerTypes.LK):
super().__init__(num_features=num_features,
num_levels=num_levels,
scale_factor=scale_factor,
detector_type=detector_type,
descriptor_type=descriptor_type,
tracker_type=tracker_type)
self.feature_manager = feature_manager_factory(
num_features=num_features,
num_levels=num_levels,
scale_factor=scale_factor,
detector_type=detector_type,
descriptor_type=descriptor_type)
#if num_levels < 3:
# Printer.green('LkFeatureTracker: forcing at least 3 levels on LK pyr optic flow')
# num_levels = 3
optic_flow_num_levels = max(kLkPyrOpticFlowNumLevelsMin, num_levels)
Printer.green('LkFeatureTracker: num levels on LK pyr optic flow: ',
optic_flow_num_levels)
# we use LK pyr optic flow for matching
self.lk_params = dict(winSize=(21, 21),
maxLevel=optic_flow_num_levels,
criteria=(cv2.TERM_CRITERIA_EPS
| cv2.TERM_CRITERIA_COUNT, 30, 0.01))
def track(self, img, frame_id, timestamp=None):
Printer.cyan('@tracking')
time_start = time.time()
# check image size is coherent with camera params
print("img.shape: ", img.shape)
print("camera ", self.camera.height, " x ", self.camera.width)
assert img.shape[0:2] == (self.camera.height, self.camera.width)
if timestamp is not None:
print('timestamp: ', timestamp)
self.timer_main_track.start()
# build current frame
self.timer_frame.start()
f_cur = Frame(img, self.camera, timestamp=timestamp)
self.f_cur = f_cur
print("frame: ", f_cur.id)
self.timer_frame.refresh()
# reset indexes of matches
self.idxs_ref = []
self.idxs_cur = []
if self.state == SlamState.NO_IMAGES_YET:
# push first frame in the inizializer
self.intializer.init(f_cur)
self.state = SlamState.NOT_INITIALIZED
return # EXIT (jump to second frame)
if self.state == SlamState.NOT_INITIALIZED:
# try to inizialize
initializer_output, intializer_is_ok = self.intializer.initialize(
f_cur, img)
if intializer_is_ok:
kf_ref = initializer_output.kf_ref
kf_cur = initializer_output.kf_cur
# add the two initialized frames in the map
self.map.add_frame(
kf_ref) # add first frame in map and update its frame id
self.map.add_frame(
kf_cur) # add second frame in map and update its frame id
# add the two initialized frames as keyframes in the map
self.map.add_keyframe(
kf_ref) # add first keyframe in map and update its kid
self.map.add_keyframe(
kf_cur) # add second keyframe in map and update its kid
kf_ref.init_observations()
kf_cur.init_observations()
# add points in map
new_pts_count, _, _ = self.map.add_points(
initializer_output.pts,
None,
kf_cur,
kf_ref,
initializer_output.idxs_cur,
initializer_output.idxs_ref,
img,
do_check=False)
Printer.green("map: initialized %d new points" %
(new_pts_count))
# update covisibility graph connections
kf_ref.update_connections()
kf_cur.update_connections()
# update tracking info
self.f_cur = kf_cur
self.f_cur.kf_ref = kf_ref
self.kf_ref = kf_cur # set reference keyframe
self.kf_last = kf_cur # set last added keyframe
self.map.local_map.update(self.kf_ref)
self.state = SlamState.OK
self.update_tracking_history()
self.motion_model.update_pose(kf_cur.timestamp,
kf_cur.position,
kf_cur.quaternion)
self.motion_model.is_ok = False # after initialization you cannot use motion model for next frame pose prediction (time ids of initialized poses may not be consecutive)
self.intializer.reset()
if kUseDynamicDesDistanceTh:
self.descriptor_distance_sigma = self.dyn_config.update_descriptor_stat(
kf_ref, kf_cur, initializer_output.idxs_ref,
initializer_output.idxs_cur)
return # EXIT (jump to next frame)
# get previous frame in map as reference
f_ref = self.map.get_frame(-1)
#f_ref_2 = self.map.get_frame(-2)
self.f_ref = f_ref
# add current frame f_cur to map
self.map.add_frame(f_cur)
self.f_cur.kf_ref = self.kf_ref
# reset pose state flag
self.pose_is_ok = False
with self.map.update_lock:
# check for map point replacements in previous frame f_ref (some points might have been replaced by local mapping during point fusion)
self.f_ref.check_replaced_map_points()
if kUseDynamicDesDistanceTh:
print('descriptor_distance_sigma: ',
self.descriptor_distance_sigma)
self.local_mapping.descriptor_distance_sigma = self.descriptor_distance_sigma
# udpdate (velocity) old motion model # c1=ref_ref, c2=ref, c3=cur; c=cur, r=ref
#self.velocity = np.dot(f_ref.pose, inv_T(f_ref_2.pose)) # Tc2c1 = Tc2w * Twc1 (predicted Tcr)
#self.predicted_pose = g2o.Isometry3d(np.dot(self.velocity, f_ref.pose)) # Tc3w = Tc2c1 * Tc2w (predicted Tcw)
# set intial guess for current pose optimization
if kUseMotionModel and self.motion_model.is_ok:
print('using motion model for next pose prediction')
# update f_ref pose according to its reference keyframe (the pose of the reference keyframe could be updated by local mapping)
self.f_ref.update_pose(
self.tracking_history.relative_frame_poses[-1] *
self.f_ref.kf_ref.isometry3d)
# predict pose by using motion model
self.predicted_pose, _ = self.motion_model.predict_pose(
timestamp, self.f_ref.position, self.f_ref.orientation)
f_cur.update_pose(self.predicted_pose)
else:
print('setting f_cur.pose <-- f_ref.pose')
# use reference frame pose as initial guess
f_cur.update_pose(f_ref.pose)
# track camera motion from f_ref to f_cur
self.track_previous_frame(f_ref, f_cur)
if not self.pose_is_ok:
# if previous track didn't go well then track the camera motion from kf_ref to f_cur
self.track_keyframe(self.kf_ref, f_cur)
# now, having a better estimate of f_cur pose, we can find more map point matches:
# find matches between {local map points} (points in the local map) and {unmatched keypoints of f_cur}
if self.pose_is_ok:
self.track_local_map(f_cur)
# end block {with self.map.update_lock:}
# TODO: add relocalization
# HACK: since local mapping is not fast enough in python (and tracking is not in real-time) => give local mapping more time to process stuff
self.wait_for_local_mapping(
) # N.B.: this must be outside the `with self.map.update_lock:` block
with self.map.update_lock:
# update slam state
if self.pose_is_ok:
self.state = SlamState.OK
else:
self.state = SlamState.LOST
Printer.red('tracking failure')
# update motion model state
self.motion_model.is_ok = self.pose_is_ok
if self.pose_is_ok: # if tracking was successful
# update motion model
self.motion_model.update_pose(timestamp, f_cur.position,
f_cur.quaternion)
f_cur.clean_vo_map_points()
# do we need a new KeyFrame?
need_new_kf = self.need_new_keyframe(f_cur)
if need_new_kf:
Printer.green('adding new KF with frame id % d: ' %
(f_cur.id))
if kLogKFinfoToFile:
self.kf_info_logger.info(
'adding new KF with frame id % d: ' % (f_cur.id))
kf_new = KeyFrame(f_cur, img)
self.kf_last = kf_new
self.kf_ref = kf_new
f_cur.kf_ref = kf_new
self.local_mapping.push_keyframe(kf_new)
if not kLocalMappingOnSeparateThread:
self.local_mapping.do_local_mapping()
else:
Printer.yellow('NOT KF')
# From ORBSLAM2:
# Clean outliers once keyframe generation has been managed:
# we allow points with high innovation (considered outliers by the Huber Function)
# pass to the new keyframe, so that bundle adjustment will finally decide
# if they are outliers or not. We don't want next frame to estimate its position
# with those points so we discard them in the frame.
f_cur.clean_outlier_map_points()
if self.f_cur.kf_ref is None:
self.f_cur.kf_ref = self.kf_ref
self.update_tracking_history(
) # must stay after having updated slam state (self.state)
Printer.green("map: %d points, %d keyframes" %
(self.map.num_points(), self.map.num_keyframes()))
#self.update_history()
self.timer_main_track.refresh()
duration = time.time() - time_start
print('tracking duration: ', duration)
if display2d is not None:
getchar()
else:
key_cv = cv2.waitKey(
0) & 0xFF # useful when drawing stuff for debugging
if do_step and img_id > 1:
# stop at each frame
if display2d is not None:
getchar()
else:
key_cv = cv2.waitKey(0) & 0xFF
if key == 'd' or (key_cv == ord('d')):
do_step = not do_step
Printer.green('do step: ', do_step)
if key == 'q' or (key_cv == ord('q')):
if display2d is not None:
display2d.quit()
if viewer3D is not None:
viewer3D.quit()
if matched_points_plt is not None:
matched_points_plt.quit()
break
if viewer3D is not None:
is_paused = not viewer3D.is_paused()
slam.quit()
def initialize(self, f_cur, img_cur):
if self.num_failures > kNumOfFailuresAfterWichNumMinTriangulatedPointsIsHalved:
self.num_min_triangulated_points = 0.5 * Parameters.kInitializerNumMinTriangulatedPoints
self.num_failures = 0
Printer.orange(
'Initializer: halved min num triangulated features to ',
self.num_min_triangulated_points)
# prepare the output
out = InitializerOutput()
is_ok = False
#print('num frames: ', len(self.frames))
# if too many frames have passed, move the current idx_f_ref forward
# this is just one very simple policy that can be used
if self.f_ref is not None:
if f_cur.id - self.f_ref.id > kMaxIdDistBetweenIntializingFrames:
self.f_ref = self.frames[-1] # take last frame in the buffer
#self.idx_f_ref = len(self.frames)-1 # take last frame in the buffer
#self.idx_f_ref = self.frames.index(self.f_ref) # since we are using a deque, the code of the previous commented line is not valid anymore
#print('*** idx_f_ref:',self.idx_f_ref)
#self.f_ref = self.frames[self.idx_f_ref]
f_ref = self.f_ref
#print('ref fid: ',self.f_ref.id,', curr fid: ', f_cur.id, ', idxs_ref: ', self.idxs_ref)
# append current frame
self.frames.append(f_cur)
# if the current frames do no have enough features exit
if len(f_ref.kps) < self.num_min_features or len(
f_cur.kps) < self.num_min_features:
Printer.red('Inializer: ko - not enough features!')
self.num_failures += 1
return out, is_ok
# find keypoint matches
idxs_cur, idxs_ref = match_frames(f_cur, f_ref,
kFeatureMatchRatioTestInitializer)
#print('|------------')
#print('deque ids: ', [f.id for f in self.frames])
#print('initializing frames ', f_cur.id, ', ', f_ref.id)
#print("# keypoint matches: ", len(idxs_cur))
Trc = self.estimatePose(f_ref.kpsn[idxs_ref], f_cur.kpsn[idxs_cur])
Tcr = inv_T(Trc) # Tcr w.r.t. ref frame
f_ref.update_pose(np.eye(4))
f_cur.update_pose(Tcr)
# remove outliers from keypoint matches by using the mask computed with inter frame pose estimation
mask_idxs = (self.mask_match.ravel() == 1)
self.num_inliers = sum(mask_idxs)
#print('# keypoint inliers: ', self.num_inliers )
idx_cur_inliers = idxs_cur[mask_idxs]
idx_ref_inliers = idxs_ref[mask_idxs]
# create a temp map for initializing
map = Map()
f_ref.reset_points()
f_cur.reset_points()
#map.add_frame(f_ref)
#map.add_frame(f_cur)
kf_ref = KeyFrame(f_ref)
kf_cur = KeyFrame(f_cur, img_cur)
map.add_keyframe(kf_ref)
map.add_keyframe(kf_cur)
pts3d, mask_pts3d = triangulate_normalized_points(
kf_cur.Tcw, kf_ref.Tcw, kf_cur.kpsn[idx_cur_inliers],
kf_ref.kpsn[idx_ref_inliers])
new_pts_count, mask_points, _ = map.add_points(
pts3d,
mask_pts3d,
kf_cur,
kf_ref,
idx_cur_inliers,
idx_ref_inliers,
img_cur,
do_check=True,
cos_max_parallax=Parameters.kCosMaxParallaxInitializer)
#print("# triangulated points: ", new_pts_count)
if new_pts_count > self.num_min_triangulated_points:
err = map.optimize(verbose=False,
rounds=20,
use_robust_kernel=True)
print("init optimization error^2: %f" % err)
num_map_points = len(map.points)
print("# map points: %d" % num_map_points)
is_ok = num_map_points > self.num_min_triangulated_points
out.pts = pts3d[mask_points]
out.kf_cur = kf_cur
out.idxs_cur = idx_cur_inliers[mask_points]
out.kf_ref = kf_ref
out.idxs_ref = idx_ref_inliers[mask_points]
# set scene median depth to equal desired_median_depth'
desired_median_depth = Parameters.kInitializerDesiredMedianDepth
median_depth = kf_cur.compute_points_median_depth(out.pts)
depth_scale = desired_median_depth / median_depth
print('forcing current median depth ', median_depth, ' to ',
desired_median_depth)
out.pts[:, :3] = out.pts[:, :3] * depth_scale # scale points
tcw = kf_cur.tcw * depth_scale # scale initial baseline
kf_cur.update_translation(tcw)
map.delete()
if is_ok:
Printer.green('Inializer: ok!')
else:
self.num_failures += 1
#Printer.red('Inializer: ko!')
print('|------------')
return out, is_ok
