def updateHistory(self):
if (self.init_history is True) and (self.trueX is not None):
self.t0_est = np.array([self.cur_t[0], self.cur_t[1], self.cur_t[2]]) # starting translation
self.t0_gt = np.array([self.trueX, self.trueY, self.trueZ]) # starting translation
self.init_history = False
if (self.t0_est is not None) and (self.t0_gt is not None):
p = [self.cur_t[0]-self.t0_est[0], self.cur_t[1]-self.t0_est[1], self.cur_t[2]-self.t0_est[2]] # the estimated traj starts at 0
self.traj3d_est.append(p)
pg = [self.trueX-self.t0_gt[0], self.trueY-self.t0_gt[1], self.trueZ-self.t0_gt[2]] # the groudtruth traj starts at 0
self.traj3d_gt.append(pg)
self.poses.append(poseRt(self.cur_R, p))
def estimate_pose_ess_mat(self, kpn_ref, kpn_cur):
E, self.mask_match = cv2.findEssentialMat(
kpn_cur,
kpn_ref,
focal=1,
pp=(0., 0.),
method=cv2.RANSAC,
prob=kRansacProb,
threshold=kRansacThresholdNormalized)
_, R, t, mask = cv2.recoverPose(E,
kpn_cur,
kpn_ref,
focal=1,
pp=(0., 0.))
return poseRt(R, t.T) # Rrc,trc (cur with respect to 'ref' frame)
def updateHistory(self):
f_cur = self.map.frames[-1]
self.cur_R = f_cur.pose[:3, :3].T
self.cur_t = np.dot(-self.cur_R, f_cur.pose[:3, 3])
if (self.init_history is True) and (self.trueX is not None):
self.t0_est = np.array(
[self.cur_t[0], self.cur_t[1],
self.cur_t[2]]) # starting translation
self.t0_gt = np.array([self.trueX, self.trueY,
self.trueZ]) # starting translation
if (self.t0_est is not None) and (self.t0_gt is not None):
p = [
self.cur_t[0] - self.t0_est[0], self.cur_t[1] - self.t0_est[1],
self.cur_t[2] - self.t0_est[2]
] # the estimated traj starts at 0
self.traj3d_est.append(p)
self.traj3d_gt.append([
self.trueX - self.t0_gt[0], self.trueY - self.t0_gt[1],
self.trueZ - self.t0_gt[2]
])
self.poses.append(poseRt(self.cur_R, p))
def track(self, img, frame_id, pose=None, verts=None):
# check image size is coherent with camera params
print("img.shape ", img.shape)
print("cam ", self.H, " x ", self.W)
assert img.shape[0:2] == (self.H, self.W)
self.timer_main_track.start()
# build current frame
self.timer_frame.start()
f_cur = Frame(self.map, img, self.K, self.Kinv, self.D, des=verts)
self.timer_frame.refresh()
if self.stage == SLAMStage.NO_IMAGES_YET:
# push first frame in the inizializer
self.intializer.init(f_cur)
self.stage = SLAMStage.NOT_INITIALIZED
return # EXIT (jump to second frame)
if self.stage == SLAMStage.NOT_INITIALIZED:
# try to inizialize
initializer_output, is_ok = self.intializer.initialize(f_cur, img)
if is_ok:
f_ref = self.intializer.f_ref
# add the two initialized frames in the map
self.map.add_frame(
f_ref) # add first frame in map and update its id
self.map.add_frame(
f_cur) # add second frame in map and update its id
# add points in map
new_pts_count, _ = self.map.add_points(
initializer_output.points4d, None, f_cur, f_ref,
initializer_output.idx_cur, initializer_output.idx_ref,
img)
Printer.green("map: initialized %d new points" %
(new_pts_count))
self.stage = SLAMStage.OK
return # EXIT (jump to next frame)
f_ref = self.map.frames[-1] # get previous frame in map
self.map.add_frame(f_cur) # add f_cur to map
# udpdate (velocity) motion model (kinematic without damping)
self.velocity = np.dot(f_ref.pose,
np.linalg.inv(self.map.frames[-2].pose))
predicted_pose = np.dot(self.velocity, f_ref.pose)
if kUseMotionModel is True:
print('using motion model')
# set intial guess for current pose optimization
f_cur.pose = predicted_pose.copy()
#f_cur.pose = f_ref.pose.copy() # get the last pose as an initial guess for optimization
if kUseSearchFrameByProjection:
# search frame by projection: match map points observed in f_ref with keypoints of f_cur
print('search frame by projection...')
idx_ref, idx_cur, num_found_map_pts = search_frame_by_projection(
f_ref, f_cur)
print("# found map points in prev frame: %d " %
num_found_map_pts)
else:
self.timer_match.start()
# find keypoint matches between f_cur and f_ref
idx_cur, idx_ref = match_frames(f_cur, f_ref)
self.num_matched_kps = idx_cur.shape[0]
print('# keypoint matches: ', self.num_matched_kps)
self.timer_match.refresh()
else:
print('estimating pose by fitting essential mat')
self.timer_match.start()
# find keypoint matches between f_cur and f_ref
idx_cur, idx_ref = match_frames(f_cur, f_ref)
self.num_matched_kps = idx_cur.shape[0]
print('# keypoint matches: ', self.num_matched_kps)
self.timer_match.refresh()
# N.B.: please, in order to understand the limitations of fitting an essential mat, read the comments of the method self.estimate_pose_ess_mat()
self.timer_pose_est.start()
# estimate inter frame camera motion by using found keypoint matches
Mrc = self.estimate_pose_ess_mat(f_ref.kpsn[idx_ref],
f_cur.kpsn[idx_cur])
Mcr = np.linalg.inv(poseRt(Mrc[:3, :3], Mrc[:3, 3]))
f_cur.pose = np.dot(Mcr, f_ref.pose)
self.timer_pose_est.refresh()
# remove outliers from keypoint matches by using the mask computed with inter frame pose estimation
mask_index = (self.mask_match.ravel() == 1)
self.num_inliers = sum(mask_index)
print('# inliers: ', self.num_inliers)
idx_ref = idx_ref[mask_index]
idx_cur = idx_cur[mask_index]
# if too many outliers reset estimated pose
if self.num_inliers < kNumMinInliersEssentialMat:
f_cur.pose = f_ref.pose.copy(
) # reset estimated pose to previous frame
Printer.red('Essential mat: not enough inliers!')
# set intial guess for current pose optimization:
# keep the estimated rotation and override translation with ref frame translation (we do not have a proper scale for the translation)
f_cur.pose[:, 3] = f_ref.pose[:, 3].copy()
#f_cur.pose[:,3] = predicted_pose[:,3].copy() # or use motion model for translation
# populate f_cur with map points by propagating map point matches of f_ref:
# we use map points observed in f_ref and keypoint matches between f_ref and f_cur
num_found_map_pts_inter_frame = 0
if not kUseSearchFrameByProjection:
for i, idx in enumerate(idx_ref): # iterate over keypoint matches
if f_ref.points[
idx] is not None: # if we have a map point P for i-th matched keypoint in f_ref
f_ref.points[idx].add_observation(
f_cur, idx_cur[i]
) # then P is automatically matched to the i-th matched keypoint in f_cur
num_found_map_pts_inter_frame += 1
print("# matched map points in prev frame: %d " %
num_found_map_pts_inter_frame)
# f_cur pose optimization 1 (here we use first available information coming from first guess of f_cur pose and map points of f_ref matched keypoints )
self.timer_pose_opt.start()
pose_opt_error, pose_is_ok, self.num_vo_map_points = optimizer_g2o.poseOptimization(
f_cur, verbose=False)
print("pose opt err1: %f, ok: %d" % (pose_opt_error, int(pose_is_ok)))
# discard outliers detected in pose optimization (in current frame)
#f_cur.reset_outlier_map_points()
if pose_is_ok is True:
# discard outliers detected in f_cur pose optimization (in current frame)
f_cur.reset_outlier_map_points()
else:
# if current pose optimization failed, reset f_cur pose to f_ref pose
f_cur.pose = f_ref.pose.copy()
self.timer_pose_opt.pause()
# TODO: add recover in case of f_cur pose optimization failure
# 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 built local map) and {unmatched keypoints of f_cur}
if pose_is_ok is True and not self.map.local_map.is_empty():
self.timer_seach_map.start()
#num_found_map_pts = search_local_frames_by_projection(self.map, f_cur)
num_found_map_pts = search_map_by_projection(
self.map.local_map.points, f_cur) # use the built local map
print("# matched map points in local map: %d " % num_found_map_pts)
self.timer_seach_map.refresh()
# f_cur pose optimization 2 with all the matched map points
self.timer_pose_opt.resume()
pose_opt_error, pose_is_ok, self.num_vo_map_points = optimizer_g2o.poseOptimization(
f_cur, verbose=False)
print("pose opt err2: %f, ok: %d" %
(pose_opt_error, int(pose_is_ok)))
print("# valid matched map points: %d " % self.num_vo_map_points)
# discard outliers detected in pose optimization (in current frame)
if pose_is_ok is True:
f_cur.reset_outlier_map_points()
self.timer_pose_opt.refresh()
if kUseSearchFrameByProjection:
print("search for triangulation with epipolar lines...")
idx_ref, idx_cur, self.num_matched_kps, _ = search_frame_for_triangulation(
f_ref, f_cur, img)
print("# matched keypoints in prev frame: %d " %
self.num_matched_kps)
# if pose is ok, then we try to triangulate the matched keypoints (between f_cur and f_ref) that do not have a corresponding map point
if pose_is_ok is True and len(idx_ref) > 0:
self.timer_triangulation.start()
# TODO: this triangulation should be done from keyframes!
good_pts4d = np.array([
f_cur.points[i] is None for i in idx_cur
]) # matched keypoints of f_cur without a corresponding map point
# do triangulation in global frame
pts4d = triangulate_points(f_cur.pose, f_ref.pose,
f_cur.kpsn[idx_cur],
f_ref.kpsn[idx_ref], good_pts4d)
good_pts4d &= np.abs(pts4d[:, 3]) != 0
#pts4d /= pts4d[:, 3:]
pts4d[good_pts4d] /= pts4d[good_pts4d,
3:] # get homogeneous 3-D coords
new_pts_count, _ = self.map.add_points(pts4d,
good_pts4d,
f_cur,
f_ref,
idx_cur,
idx_ref,
img,
check_parallax=True)
print("# added map points: %d " % (new_pts_count))
self.timer_triangulation.refresh()
# local optimization
self.time_local_opt.start()
err = self.map.locally_optimize(local_window=kLocalWindow)
print("local optimization error: %f" % err)
self.time_local_opt.refresh()
# large window optimization of the map
# TODO: move this in a seperate thread with local mapping
if kUseLargeWindowBA is True and f_cur.id >= parameters.kEveryNumFramesLargeWindowBA and f_cur.id % parameters.kEveryNumFramesLargeWindowBA == 0:
err = self.map.optimize(
local_window=parameters.kLargeWindow) # verbose=True)
Printer.blue("large window optimization error: %f" % err)
print("map: %d points, %d frames" %
(len(self.map.points), len(self.map.frames)))
#self.updateHistory()
self.timer_main_track.refresh()
def optimization(frames,
points,
local_window,
fixed_points=False,
verbose=False,
rounds=40,
use_robust_kernel=False):
if local_window is None:
local_frames = frames
else:
local_frames = frames[-local_window:]
# create g2o optimizer
opt = g2o.SparseOptimizer()
solver = g2o.BlockSolverSE3(g2o.LinearSolverCSparseSE3())
solver = g2o.OptimizationAlgorithmLevenberg(solver)
opt.set_algorithm(solver)
thHuberMono = math.sqrt(5.991)
# chi-square 2 DOFS
graph_frames, graph_points = {}, {}
# add frame vertices to graph
for f in (
local_frames if fixed_points else frames
): # if points are fixed then consider just the local frames, otherwise we need all frames or at least two frames for each point
#print('adding vertex frame ', f.id, ' to graph')
pose = f.pose
se3 = g2o.SE3Quat(pose[0:3, 0:3], pose[0:3, 3])
v_se3 = g2o.VertexSE3Expmap()
v_se3.set_estimate(se3)
v_se3.set_id(f.id * 2) # even ids
v_se3.set_fixed(f.id < 1 or f not in local_frames)
opt.add_vertex(v_se3)
# confirm pose correctness
#est = v_se3.estimate()
#assert np.allclose(pose[0:3, 0:3], est.rotation().matrix())
#assert np.allclose(pose[0:3, 3], est.translation())
graph_frames[f] = v_se3
# add point vertices to graph
for p in points:
if p.is_bad and not fixed_points:
continue
if not any([f in local_frames for f in p.frames]):
continue
#print('adding vertex point ', p.id,' to graph')
v_p = g2o.VertexSBAPointXYZ()
v_p.set_id(p.id * 2 + 1) # odd ids
v_p.set_estimate(p.pt[0:3])
v_p.set_marginalized(True)
v_p.set_fixed(fixed_points)
opt.add_vertex(v_p)
graph_points[p] = v_p
# add edges
for f, idx in zip(p.frames, p.idxs):
if f not in graph_frames:
continue
#print('adding edge between point ', p.id,' and frame ', f.id)
edge = g2o.EdgeSE3ProjectXYZ()
edge.set_vertex(0, v_p)
edge.set_vertex(1, graph_frames[f])
edge.set_measurement(f.kpsu[idx])
invSigma2 = Frame.detector.inv_level_sigmas2[f.octaves[idx]]
edge.set_information(np.eye(2) * invSigma2)
if use_robust_kernel:
edge.set_robust_kernel(g2o.RobustKernelHuber(thHuberMono))
edge.fx = f.fx
edge.fy = f.fy
edge.cx = f.cx
edge.cy = f.cy
opt.add_edge(edge)
if verbose:
opt.set_verbose(True)
opt.initialize_optimization()
opt.optimize(rounds)
# put frames back
for f in graph_frames:
est = graph_frames[f].estimate()
R = est.rotation().matrix()
t = est.translation()
f.pose = poseRt(R, t)
# put points back
if not fixed_points:
for p in graph_points:
p.pt = np.array(graph_points[p].estimate())
return opt.active_chi2()
def localOptimization(frames,
points,
frames_ref=[],
fixed_points=False,
verbose=False,
rounds=10):
# create g2o optimizer
opt = g2o.SparseOptimizer()
solver = g2o.BlockSolverSE3(g2o.LinearSolverCSparseSE3())
solver = g2o.OptimizationAlgorithmLevenberg(solver)
opt.set_algorithm(solver)
#robust_kernel = g2o.RobustKernelHuber(np.sqrt(5.991)) # chi-square 2 DOFs
thHuberMono = math.sqrt(5.991)
# chi-square 2 DOFS
graph_frames, graph_points = {}, {}
all_frames = frames + frames_ref
# add frame vertices to graph
for f in all_frames:
#print('adding vertex frame ', f.id, ' to graph')
pose = f.pose
se3 = g2o.SE3Quat(pose[0:3, 0:3], pose[0:3, 3])
v_se3 = g2o.VertexSE3Expmap()
v_se3.set_estimate(se3)
v_se3.set_id(f.id * 2) # even ids
v_se3.set_fixed(f.id < 1 or f in frames_ref)
opt.add_vertex(v_se3)
graph_frames[f] = v_se3
# confirm pose correctness
#est = v_se3.estimate()
#assert np.allclose(pose[0:3, 0:3], est.rotation().matrix())
#assert np.allclose(pose[0:3, 3], est.translation())
graph_edges = {}
num_point_edges = 0
# add point vertices to graph
for p in points:
assert (p is not None)
if p.is_bad and not fixed_points: # do not consider bad points unless they are fixed
continue
if not any([f in frames for f in p.frames]): # this is redundant now
continue
#print('adding vertex point ', p.id,' to graph')
v_p = g2o.VertexSBAPointXYZ()
v_p.set_id(p.id * 2 + 1) # odd ids
v_p.set_estimate(p.pt[0:3])
v_p.set_marginalized(True)
v_p.set_fixed(fixed_points)
opt.add_vertex(v_p)
graph_points[p] = v_p
# add edges
for f, p_idx in zip(p.frames, p.idxs):
assert (f.points[p_idx] == p)
if f not in graph_frames:
continue
#print('adding edge between point ', p.id,' and frame ', f.id)
edge = g2o.EdgeSE3ProjectXYZ()
edge.set_vertex(0, v_p)
edge.set_vertex(1, graph_frames[f])
edge.set_measurement(f.kpsu[p_idx])
invSigma2 = Frame.detector.inv_level_sigmas2[f.octaves[p_idx]]
edge.set_information(np.eye(2) * invSigma2)
edge.set_robust_kernel(g2o.RobustKernelHuber(thHuberMono))
edge.fx = f.fx
edge.fy = f.fy
edge.cx = f.cx
edge.cy = f.cy
opt.add_edge(edge)
graph_edges[edge] = (p, f, p_idx) # f.points[p_idx] == p
num_point_edges += 1
if verbose:
opt.set_verbose(True)
# initial optimization
opt.initialize_optimization()
opt.optimize(5)
chi2Mono = 5.991 # chi-square 2 DOFs
# check inliers observation
for edge, edge_data in graph_edges.items():
p = edge_data[0]
if p.is_bad is True:
continue
if edge.chi2() > chi2Mono or not edge.is_depth_positive():
edge.set_level(1)
edge.set_robust_kernel(None)
# optimize again without outliers
opt.initialize_optimization()
opt.optimize(rounds)
# clean map observations
num_bad_observations = 0
outliers_data = []
for edge, edge_data in graph_edges.items():
p, f, p_idx = edge_data
if p.is_bad is True:
continue
assert (f.points[p_idx] == p)
if edge.chi2() > chi2Mono or not edge.is_depth_positive():
num_bad_observations += 1
outliers_data.append((p, f, p_idx))
for d in outliers_data:
(p, f, p_idx) = d
assert (f.points[p_idx] == p)
p.remove_observation(f, p_idx)
f.remove_point(p)
# put frames back
for f in graph_frames:
est = graph_frames[f].estimate()
R = est.rotation().matrix()
t = est.translation()
f.pose = poseRt(R, t)
# put points back
if not fixed_points:
for p in graph_points:
p.pt = np.array(graph_points[p].estimate())
return opt.active_chi2(), num_bad_observations / num_point_edges
def poseOptimization(frame, verbose=False, rounds=10):
is_ok = True
# create g2o optimizer
opt = g2o.SparseOptimizer()
solver = g2o.BlockSolverSE3(g2o.LinearSolverCSparseSE3())
solver = g2o.OptimizationAlgorithmLevenberg(solver)
opt.set_algorithm(solver)
#robust_kernel = g2o.RobustKernelHuber(np.sqrt(5.991)) # chi-square 2 DOFs
thHuberMono = math.sqrt(5.991)
# chi-square 2 DOFS
point_edge_pairs = {}
num_point_edges = 0
se3 = g2o.SE3Quat(frame.pose[0:3, 0:3], frame.pose[0:3, 3])
v_se3 = g2o.VertexSE3Expmap()
v_se3.set_estimate(se3)
v_se3.set_id(0)
v_se3.set_fixed(False)
opt.add_vertex(v_se3)
# add point vertices to graph
for idx, p in enumerate(frame.points):
if p is None: # do not use p.is_bad here since a single point observation is ok for pose optimization
continue
frame.outliers[idx] = False
# add edge
#print('adding edge between point ', p.id,' and frame ', frame.id)
edge = g2o.EdgeSE3ProjectXYZOnlyPose()
edge.set_vertex(0, opt.vertex(0))
edge.set_measurement(frame.kpsu[idx])
invSigma2 = Frame.detector.inv_level_sigmas2[frame.octaves[idx]]
edge.set_information(np.eye(2) * invSigma2)
edge.set_robust_kernel(g2o.RobustKernelHuber(thHuberMono))
edge.fx = frame.fx
edge.fy = frame.fy
edge.cx = frame.cx
edge.cy = frame.cy
edge.Xw = p.pt[0:3]
opt.add_edge(edge)
point_edge_pairs[p] = (edge, idx) # one edge per point
num_point_edges += 1
if num_point_edges < 3:
Printer.red('poseOptimization: not enough correspondences!')
is_ok = False
return 0, is_ok, 0
if verbose:
opt.set_verbose(True)
# We perform 4 optimizations, after each optimization we classify observation as inlier/outlier
# At the next optimization, outliers are not included, but at the end they can be classified as inliers again.
chi2Mono = 5.991 # chi-square 2 DOFs
num_bad_points = 0
for it in range(4):
opt.initialize_optimization()
opt.optimize(rounds)
num_bad_points = 0
for p, edge_pair in point_edge_pairs.items():
if frame.outliers[edge_pair[1]] is True:
edge_pair[0].compute_error()
chi2 = edge_pair[0].chi2()
if chi2 > chi2Mono:
frame.outliers[edge_pair[1]] = True
edge_pair[0].set_level(1)
num_bad_points += 1
else:
frame.outliers[edge_pair[1]] = False
edge_pair[0].set_level(0)
if it == 2:
edge_pair[0].set_robust_kernel(None)
if len(opt.edges()) < 10:
Printer.red('poseOptimization: stopped - not enough edges!')
is_ok = False
break
print('pose optimization: initial ', num_point_edges, ' points, found ',
num_bad_points, ' bad points')
if num_point_edges == num_bad_points:
Printer.red(
'poseOptimization: all the initial correspondences are bad!')
is_ok = False
# update pose estimation
if is_ok is True:
est = v_se3.estimate()
R = est.rotation().matrix()
t = est.translation()
frame.pose = poseRt(R, t)
num_valid_points = num_point_edges - num_bad_points
return opt.active_chi2(), is_ok, num_valid_points
def initialize(self, f_cur, img_cur):
# prepare the output
out = InitializerOutput()
is_ok = False
# if too many frames have passed, move the current id_ref forward
if (len(self.frames) - 1) - self.id_ref >= kMaxIdDistBetweenFrames:
self.id_ref = len(self.frames) - 1 # take last frame in the array
self.f_ref = self.frames[self.id_ref]
f_ref = self.f_ref
# append current frame
self.frames.append(f_cur)
# if the current frames do no have enough features exit
if len(f_ref.kps) < kNumMinFeatures or len(
f_cur.kps) < kNumMinFeatures:
Printer.red('Inializer: not enough features!')
return out, is_ok
# find image point matches
idx_cur, idx_ref = match_frames(f_cur, f_ref)
print('├────────')
print('initializing frames ', f_cur.id, ', ', f_ref.id)
Mrc = self.estimatePose(f_ref.kpsn[idx_ref], f_cur.kpsn[idx_cur])
f_cur.pose = np.linalg.inv(poseRt(
Mrc[:3, :3], Mrc[:3, 3])) # [Rcr, tcr] w.r.t. ref frame
# remove outliers
mask_index = [i for i, v in enumerate(self.mask_match) if v > 0]
print('num inliers: ', len(mask_index))
idx_cur_inliers = idx_cur[mask_index]
idx_ref_inliers = idx_ref[mask_index]
# create a temp map for initializing
map = Map()
map.add_frame(f_ref)
map.add_frame(f_cur)
points4d = self.triangulatePoints(f_cur.pose, f_ref.pose,
f_cur.kpsn[idx_cur_inliers],
f_ref.kpsn[idx_ref_inliers])
#pts4d = triangulate(f_cur.pose, f_ref.pose, f_cur.kpsn[idx_cur], f_ref.kpsn[idx_ref])
new_pts_count, mask_points = map.add_points(points4d,
None,
f_cur,
f_ref,
idx_cur_inliers,
idx_ref_inliers,
img_cur,
check_parallax=True)
print("triangulated: %d new points from %d matches" %
(new_pts_count, len(idx_cur)))
err = map.optimize(verbose=False)
print("pose opt err: %f" % err)
#reset points in frames
f_cur.reset_points()
f_ref.reset_points()
num_map_points = len(map.points)
#print("# map points: %d" % num_map_points)
is_ok = num_map_points > kNumMinTriangulatedPoints
out.points4d = points4d[mask_points]
out.f_cur = f_cur
out.idx_cur = idx_cur_inliers[mask_points]
out.f_ref = f_ref
out.idx_ref = idx_ref_inliers[mask_points]
# set median depth to 'desired_median_depth'
desired_median_depth = parameters.kInitializerDesiredMedianDepth
median_depth = f_cur.compute_points_median_depth(out.points4d)
depth_scale = desired_median_depth / median_depth
print('median depth: ', median_depth)
out.points4d = out.points4d * depth_scale # scale points
f_cur.pose[:3,
3] = f_cur.pose[:3,
3] * depth_scale # scale initial baseline
print('├────────')
Printer.green('Inializer: ok!')
return out, is_ok