40 #ifndef PCL_SURFACE_IMPL_MLS_H_
41 #define PCL_SURFACE_IMPL_MLS_H_
43 #include <pcl/point_traits.h>
44 #include <pcl/surface/mls.h>
45 #include <pcl/common/io.h>
46 #include <pcl/common/copy_point.h>
48 #include <pcl/common/eigen.h>
56 template <
typename Po
intInT,
typename Po
intOutT>
void
67 normals_->header = input_->header;
69 normals_->width = normals_->height = 0;
70 normals_->points.clear ();
74 output.
header = input_->header;
78 if (search_radius_ <= 0 || sqr_gauss_param_ <= 0)
80 PCL_ERROR (
"[pcl::%s::process] Invalid search radius (%f) or Gaussian parameter (%f)!\n", getClassName ().c_str (), search_radius_, sqr_gauss_param_);
85 if (upsample_method_ == DISTINCT_CLOUD && !distinct_cloud_)
87 PCL_ERROR (
"[pcl::%s::process] Upsample method was set to DISTINCT_CLOUD, but no distinct cloud was specified.\n", getClassName ().c_str ());
98 if (input_->isOrganized ())
102 setSearchMethod (tree);
106 tree_->setInputCloud (input_);
108 switch (upsample_method_)
111 case (RANDOM_UNIFORM_DENSITY):
113 std::random_device rd;
115 const double tmp = search_radius_ / 2.0;
116 rng_uniform_distribution_.reset (
new std::uniform_real_distribution<> (-tmp, tmp));
120 case (VOXEL_GRID_DILATION):
121 case (DISTINCT_CLOUD):
123 if (!cache_mls_results_)
124 PCL_WARN (
"The cache mls results is forced when using upsampling method VOXEL_GRID_DILATION or DISTINCT_CLOUD.\n");
126 cache_mls_results_ =
true;
133 if (cache_mls_results_)
135 mls_results_.resize (input_->size ());
139 mls_results_.resize (1);
143 performProcessing (output);
145 if (compute_normals_)
147 normals_->height = 1;
148 normals_->width = static_cast<std::uint32_t> (normals_->size ());
150 for (std::size_t i = 0; i < output.
size (); ++i)
163 output.
width = static_cast<std::uint32_t> (output.
size ());
169 template <
typename Po
intInT,
typename Po
intOutT>
void
171 const std::vector<int> &nn_indices,
180 mls_result.
computeMLSSurface<PointInT> (*input_, index, nn_indices, search_radius_, order_);
182 switch (upsample_method_)
187 addProjectedPointNormal (index, proj.
point, proj.
normal, mls_result.
curvature, projected_points, projected_points_normals, corresponding_input_indices);
191 case (SAMPLE_LOCAL_PLANE):
194 for (
float u_disp = -static_cast<float> (upsampling_radius_); u_disp <= upsampling_radius_; u_disp += static_cast<float> (upsampling_step_))
195 for (
float v_disp = -static_cast<float> (upsampling_radius_); v_disp <= upsampling_radius_; v_disp += static_cast<float> (upsampling_step_))
196 if (u_disp * u_disp + v_disp * v_disp < upsampling_radius_ * upsampling_radius_)
199 addProjectedPointNormal (index, proj.
point, proj.
normal, mls_result.
curvature, projected_points, projected_points_normals, corresponding_input_indices);
204 case (RANDOM_UNIFORM_DENSITY):
207 const int num_points_to_add = static_cast<int> (std::floor (desired_num_points_in_radius_ / 2.0 / static_cast<double> (nn_indices.size ())));
210 if (num_points_to_add <= 0)
214 addProjectedPointNormal (index, proj.
point, proj.
normal, mls_result.
curvature, projected_points, projected_points_normals, corresponding_input_indices);
219 for (
int num_added = 0; num_added < num_points_to_add;)
221 const double u = (*rng_uniform_distribution_) (rng_);
222 const double v = (*rng_uniform_distribution_) (rng_);
225 if (u * u + v * v > search_radius_ * search_radius_ / 4)
234 addProjectedPointNormal (index, proj.
point, proj.
normal, mls_result.
curvature, projected_points, projected_points_normals, corresponding_input_indices);
247 template <
typename Po
intInT,
typename Po
intOutT>
void
249 const Eigen::Vector3d &point,
250 const Eigen::Vector3d &normal,
257 aux.x = static_cast<float> (point[0]);
258 aux.y = static_cast<float> (point[1]);
259 aux.z = static_cast<float> (point[2]);
262 copyMissingFields (input_->points[index], aux);
265 corresponding_input_indices.
indices.push_back (index);
267 if (compute_normals_)
270 aux_normal.normal_x = static_cast<float> (normal[0]);
271 aux_normal.normal_y = static_cast<float> (normal[1]);
272 aux_normal.normal_z = static_cast<float> (normal[2]);
274 projected_points_normals.
push_back (aux_normal);
279 template <
typename Po
intInT,
typename Po
intOutT>
void
283 nr_coeff_ = (order_ + 1) * (order_ + 2) / 2;
287 const unsigned int threads = threads_ == 0 ? 1 : threads_;
291 std::vector<PointIndices> corresponding_input_indices (threads);
296 #pragma omp parallel for schedule (dynamic,1000) num_threads (threads)
298 for (
int cp = 0; cp < static_cast<int> (indices_->size ()); ++cp)
302 std::vector<int> nn_indices;
303 std::vector<float> nn_sqr_dists;
306 if (searchForNeighbors ((*indices_)[cp], nn_indices, nn_sqr_dists))
309 if (nn_indices.size () >= 3)
313 const int tn = omp_get_thread_num ();
315 std::size_t pp_size = projected_points[tn].size ();
322 const int index = (*indices_)[cp];
324 std::size_t mls_result_index = 0;
325 if (cache_mls_results_)
326 mls_result_index = index;
329 computeMLSPointNormal (index, nn_indices, projected_points[tn], projected_points_normals[tn], corresponding_input_indices[tn], mls_results_[mls_result_index]);
332 for (std::size_t pp = pp_size; pp < projected_points[tn].
size (); ++pp)
333 copyMissingFields (input_->points[(*indices_)[cp]], projected_points[tn][pp]);
335 computeMLSPointNormal (index, nn_indices, projected_points, projected_points_normals, *corresponding_input_indices_, mls_results_[mls_result_index]);
338 output.
insert (output.
end (), projected_points.
begin (), projected_points.
end ());
339 if (compute_normals_)
340 normals_->insert (normals_->end (), projected_points_normals.
begin (), projected_points_normals.
end ());
348 for (
unsigned int tn = 0; tn < threads; ++tn)
350 output.
insert (output.
end (), projected_points[tn].begin (), projected_points[tn].end ());
351 corresponding_input_indices_->indices.insert (corresponding_input_indices_->indices.end (),
352 corresponding_input_indices[tn].indices.begin (), corresponding_input_indices[tn].indices.end ());
353 if (compute_normals_)
354 normals_->insert (normals_->end (), projected_points_normals[tn].begin (), projected_points_normals[tn].end ());
359 performUpsampling (output);
363 template <
typename Po
intInT,
typename Po
intOutT>
void
367 if (upsample_method_ == DISTINCT_CLOUD)
370 for (std::size_t dp_i = 0; dp_i < distinct_cloud_->size (); ++dp_i)
373 if (!std::isfinite (distinct_cloud_->points[dp_i].x))
378 std::vector<int> nn_indices;
379 std::vector<float> nn_dists;
380 tree_->nearestKSearch (distinct_cloud_->points[dp_i], 1, nn_indices, nn_dists);
381 int input_index = nn_indices.front ();
385 if (mls_results_[input_index].valid ==
false)
388 Eigen::Vector3d add_point = distinct_cloud_->points[dp_i].getVector3fMap ().template cast<double> ();
390 addProjectedPointNormal (input_index, proj.
point, proj.
normal, mls_results_[input_index].curvature, output, *normals_, *corresponding_input_indices_);
396 if (upsample_method_ == VOXEL_GRID_DILATION)
400 MLSVoxelGrid voxel_grid (input_, indices_, voxel_size_);
401 for (
int iteration = 0; iteration < dilation_iteration_num_; ++iteration)
404 for (
typename MLSVoxelGrid::HashMap::iterator m_it = voxel_grid.
voxel_grid_.begin (); m_it != voxel_grid.
voxel_grid_.end (); ++m_it)
415 std::vector<int> nn_indices;
416 std::vector<float> nn_dists;
417 tree_->nearestKSearch (p, 1, nn_indices, nn_dists);
418 int input_index = nn_indices.front ();
422 if (mls_results_[input_index].valid ==
false)
425 Eigen::Vector3d add_point = p.getVector3fMap ().template cast<double> ();
427 addProjectedPointNormal (input_index, proj.
point, proj.
normal, mls_results_[input_index].curvature, output, *normals_, *corresponding_input_indices_);
434 const Eigen::Vector3d &a_mean,
435 const Eigen::Vector3d &a_plane_normal,
436 const Eigen::Vector3d &a_u,
437 const Eigen::Vector3d &a_v,
438 const Eigen::VectorXd &a_c_vec,
439 const int a_num_neighbors,
440 const float a_curvature,
442 query_point (a_query_point), mean (a_mean), plane_normal (a_plane_normal), u_axis (a_u), v_axis (a_v), c_vec (a_c_vec), num_neighbors (a_num_neighbors),
443 curvature (a_curvature), order (a_order), valid (true)
449 Eigen::Vector3d delta = pt - mean;
450 u = delta.dot (u_axis);
451 v = delta.dot (v_axis);
452 w = delta.dot (plane_normal);
458 Eigen::Vector3d delta = pt - mean;
459 u = delta.dot (u_axis);
460 v = delta.dot (v_axis);
471 for (
int ui = 0; ui <= order; ++ui)
474 for (
int vi = 0; vi <= order - ui; ++vi)
476 result += c_vec[j++] * u_pow * v_pow;
491 Eigen::VectorXd u_pow (order + 2), v_pow (order + 2);
494 d.z = d.z_u = d.z_v = d.z_uu = d.z_vv = d.z_uv = 0;
495 u_pow (0) = v_pow (0) = 1;
496 for (
int ui = 0; ui <= order; ++ui)
498 for (
int vi = 0; vi <= order - ui; ++vi)
501 d.z += u_pow (ui) * v_pow (vi) * c_vec[j];
505 d.z_u += c_vec[j] * ui * u_pow (ui - 1) * v_pow (vi);
508 d.z_v += c_vec[j] * vi * u_pow (ui) * v_pow (vi - 1);
510 if (ui >= 1 && vi >= 1)
511 d.z_uv += c_vec[j] * ui * u_pow (ui - 1) * vi * v_pow (vi - 1);
514 d.z_uu += c_vec[j] * ui * (ui - 1) * u_pow (ui - 2) * v_pow (vi);
517 d.z_vv += c_vec[j] * vi * (vi - 1) * u_pow (ui) * v_pow (vi - 2);
520 v_pow (vi + 1) = v_pow (vi) * v;
524 u_pow (ui + 1) = u_pow (ui) * u;
533 Eigen::Vector2f k (1e-5, 1e-5);
539 if (order > 1 && c_vec.size () >= (order + 1) * (order + 2) / 2 && std::isfinite (c_vec[0]))
543 const double Zlen = std::sqrt (Z);
546 const double disc2 = H * H -
K;
547 assert (disc2 >= 0.0);
548 const double disc = std::sqrt (disc2);
552 if (std::abs (k[0]) > std::abs (k[1])) std::swap (k[0], k[1]);
556 PCL_ERROR (
"No Polynomial fit data, unable to calculate the principle curvatures!\n");
570 result.
normal = plane_normal;
571 if (order > 1 && c_vec.size () >= (order + 1) * (order + 2) / 2 && std::isfinite (c_vec[0]))
576 const double dist1 = std::abs (gw - w);
580 double e1 = (gu - u) + d.
z_u * gw - d.
z_u * w;
581 double e2 = (gv - v) + d.
z_v * gw - d.
z_v * w;
589 Eigen::MatrixXd J (2, 2);
595 Eigen::Vector2d err (e1, e2);
596 Eigen::Vector2d update = J.inverse () * err;
600 d = getPolynomialPartialDerivative (gu, gv);
602 dist2 = std::sqrt ((gu - u) * (gu - u) + (gv - v) * (gv - v) + (gw - w) * (gw - w));
604 err_total = std::sqrt (e1 * e1 + e2 * e2);
606 }
while (err_total > 1e-8 && dist2 < dist1);
612 d = getPolynomialPartialDerivative (u, v);
619 result.
normal.normalize ();
622 result.
point = mean + gu * u_axis + gv * v_axis + gw * plane_normal;
633 result.
normal = plane_normal;
634 result.
point = mean + u * u_axis + v * v_axis;
647 result.
normal = plane_normal;
649 if (order > 1 && c_vec.size () >= (order + 1) * (order + 2) / 2 && std::isfinite (c_vec[0]))
654 result.
normal.normalize ();
657 result.
point = mean + u * u_axis + v * v_axis + w * plane_normal;
666 getMLSCoordinates (pt, u, v, w);
669 if (order > 1 && num_neighbors >= required_neighbors && std::isfinite (c_vec[0]) && method != NONE)
671 if (method == ORTHOGONAL)
672 proj = projectPointOrthogonalToPolynomialSurface (u, v, w);
674 proj = projectPointSimpleToPolynomialSurface (u, v);
678 proj = projectPointToMLSPlane (u, v);
688 if (order > 1 && num_neighbors >= required_neighbors && std::isfinite (c_vec[0]) && method != NONE)
690 if (method == ORTHOGONAL)
693 getMLSCoordinates (query_point, u, v, w);
694 proj = projectPointOrthogonalToPolynomialSurface (u, v, w);
699 proj.
point = mean + (c_vec[0] * plane_normal);
702 proj.
normal = plane_normal - c_vec[order + 1] * u_axis - c_vec[1] * v_axis;
708 proj.
normal = plane_normal;
715 template <
typename Po
intT>
void
718 const std::vector<int> &nn_indices,
719 double search_radius,
720 int polynomial_order,
721 std::function<
double(
const double)> weight_func)
724 EIGEN_ALIGN16 Eigen::Matrix3d covariance_matrix;
725 Eigen::Vector4d xyz_centroid;
732 EIGEN_ALIGN16 Eigen::Vector3d::Scalar eigen_value;
733 EIGEN_ALIGN16 Eigen::Vector3d eigen_vector;
734 Eigen::Vector4d model_coefficients (0, 0, 0, 0);
735 pcl::eigen33 (covariance_matrix, eigen_value, eigen_vector);
736 model_coefficients.head<3> ().matrix () = eigen_vector;
737 model_coefficients[3] = -1 * model_coefficients.dot (xyz_centroid);
739 query_point = cloud.
points[index].getVector3fMap ().template cast<double> ();
741 if (!std::isfinite(eigen_vector[0]) || !std::isfinite(eigen_vector[1]) || !std::isfinite(eigen_vector[2]))
752 const double distance = query_point.dot (model_coefficients.head<3> ()) + model_coefficients[3];
753 mean = query_point -
distance * model_coefficients.head<3> ();
755 curvature = covariance_matrix.trace ();
758 curvature = std::abs (eigen_value / curvature);
761 plane_normal = model_coefficients.head<3> ();
764 v_axis = plane_normal.unitOrthogonal ();
765 u_axis = plane_normal.cross (v_axis);
769 num_neighbors = static_cast<int> (nn_indices.size ());
770 order = polynomial_order;
773 const int nr_coeff = (order + 1) * (order + 2) / 2;
775 if (num_neighbors >= nr_coeff)
778 weight_func = [=] (
const double sq_dist) {
return this->computeMLSWeight (sq_dist, search_radius * search_radius); };
781 Eigen::VectorXd weight_vec (num_neighbors);
782 Eigen::MatrixXd P (nr_coeff, num_neighbors);
783 Eigen::VectorXd f_vec (num_neighbors);
784 Eigen::MatrixXd P_weight_Pt (nr_coeff, nr_coeff);
788 std::vector<Eigen::Vector3d, Eigen::aligned_allocator<Eigen::Vector3d> > de_meaned (num_neighbors);
789 for (std::size_t ni = 0; ni < static_cast<std::size_t>(num_neighbors); ++ni)
791 de_meaned[ni][0] = cloud.
points[nn_indices[ni]].x - mean[0];
792 de_meaned[ni][1] = cloud.
points[nn_indices[ni]].y - mean[1];
793 de_meaned[ni][2] = cloud.
points[nn_indices[ni]].z - mean[2];
794 weight_vec (ni) = weight_func (de_meaned[ni].dot (de_meaned[ni]));
799 for (std::size_t ni = 0; ni < static_cast<std::size_t>(num_neighbors); ++ni)
802 const double u_coord = de_meaned[ni].dot(u_axis);
803 const double v_coord = de_meaned[ni].dot(v_axis);
804 f_vec (ni) = de_meaned[ni].dot (plane_normal);
809 for (
int ui = 0; ui <= order; ++ui)
812 for (
int vi = 0; vi <= order - ui; ++vi)
814 P (j++, ni) = u_pow * v_pow;
822 const Eigen::MatrixXd P_weight = P * weight_vec.asDiagonal();
823 P_weight_Pt = P_weight * P.transpose ();
824 c_vec = P_weight * f_vec;
825 P_weight_Pt.llt ().solveInPlace (c_vec);
831 template <
typename Po
intInT,
typename Po
intOutT>
835 voxel_grid_ (), data_size_ (), voxel_size_ (voxel_size)
840 const double max_size = (std::max) ((std::max)(bounding_box_size.x (), bounding_box_size.y ()), bounding_box_size.z ());
843 for (std::size_t i = 0; i < indices->size (); ++i)
844 if (std::isfinite (cloud->points[(*indices)[i]].x))
847 getCellIndex (cloud->points[(*indices)[i]].getVector3fMap (), pos);
849 std::uint64_t index_1d;
857 template <
typename Po
intInT,
typename Po
intOutT>
void
860 HashMap new_voxel_grid = voxel_grid_;
861 for (
typename MLSVoxelGrid::HashMap::iterator m_it = voxel_grid_.begin (); m_it != voxel_grid_.end (); ++m_it)
863 Eigen::Vector3i index;
864 getIndexIn3D (m_it->first, index);
867 for (
int x = -1; x <= 1; ++x)
868 for (
int y = -1; y <= 1; ++y)
869 for (
int z = -1; z <= 1; ++z)
870 if (x != 0 || y != 0 || z != 0)
872 Eigen::Vector3i new_index;
873 new_index = index + Eigen::Vector3i (x, y, z);
875 std::uint64_t index_1d;
876 getIndexIn1D (new_index, index_1d);
878 new_voxel_grid[index_1d] = leaf;
881 voxel_grid_ = new_voxel_grid;
886 template <
typename Po
intInT,
typename Po
intOutT>
void
888 PointOutT &point_out)
const
890 PointOutT temp = point_out;
892 point_out.x = temp.x;
893 point_out.y = temp.y;
894 point_out.z = temp.z;
897 #define PCL_INSTANTIATE_MovingLeastSquares(T,OutT) template class PCL_EXPORTS pcl::MovingLeastSquares<T,OutT>;
898 #define PCL_INSTANTIATE_MovingLeastSquaresOMP(T,OutT) template class PCL_EXPORTS pcl::MovingLeastSquaresOMP<T,OutT>;
900 #endif // PCL_SURFACE_IMPL_MLS_H_