38 #ifndef PCL_VOXEL_GRID_COVARIANCE_IMPL_H_
39 #define PCL_VOXEL_GRID_COVARIANCE_IMPL_H_
42 #include <pcl/filters/boost.h>
43 #include <pcl/filters/voxel_grid_covariance.h>
44 #include <Eigen/Dense>
45 #include <Eigen/Cholesky>
48 template<
typename Po
intT>
void
51 voxel_centroids_leaf_indices_.clear ();
56 PCL_WARN (
"[pcl::%s::applyFilter] No input dataset given!\n", getClassName ().c_str ());
67 Eigen::Vector4f min_p, max_p;
69 if (!filter_field_name_.empty ())
70 getMinMax3D<PointT> (input_, filter_field_name_,
static_cast<float> (filter_limit_min_),
static_cast<float> (filter_limit_max_), min_p, max_p, filter_limit_negative_);
72 getMinMax3D<PointT> (*input_, min_p, max_p);
75 int64_t dx =
static_cast<int64_t
>((max_p[0] - min_p[0]) * inverse_leaf_size_[0])+1;
76 int64_t dy =
static_cast<int64_t
>((max_p[1] - min_p[1]) * inverse_leaf_size_[1])+1;
77 int64_t dz =
static_cast<int64_t
>((max_p[2] - min_p[2]) * inverse_leaf_size_[2])+1;
79 if((dx*dy*dz) > std::numeric_limits<int32_t>::max())
81 PCL_WARN(
"[pcl::%s::applyFilter] Leaf size is too small for the input dataset. Integer indices would overflow.", getClassName().c_str());
87 min_b_[0] =
static_cast<int> (floor (min_p[0] * inverse_leaf_size_[0]));
88 max_b_[0] =
static_cast<int> (floor (max_p[0] * inverse_leaf_size_[0]));
89 min_b_[1] =
static_cast<int> (floor (min_p[1] * inverse_leaf_size_[1]));
90 max_b_[1] =
static_cast<int> (floor (max_p[1] * inverse_leaf_size_[1]));
91 min_b_[2] =
static_cast<int> (floor (min_p[2] * inverse_leaf_size_[2]));
92 max_b_[2] =
static_cast<int> (floor (max_p[2] * inverse_leaf_size_[2]));
95 div_b_ = max_b_ - min_b_ + Eigen::Vector4i::Ones ();
102 divb_mul_ = Eigen::Vector4i (1, div_b_[0], div_b_[0] * div_b_[1], 0);
104 int centroid_size = 4;
106 if (downsample_all_data_)
107 centroid_size = boost::mpl::size<FieldList>::value;
110 std::vector<pcl::PCLPointField> fields;
113 if (rgba_index == -1)
117 rgba_index = fields[rgba_index].offset;
122 if (!filter_field_name_.empty ())
125 std::vector<pcl::PCLPointField> fields;
127 if (distance_idx == -1)
128 PCL_WARN (
"[pcl::%s::applyFilter] Invalid filter field name. Index is %d.\n", getClassName ().c_str (), distance_idx);
131 for (
size_t cp = 0; cp < input_->points.size (); ++cp)
133 if (!input_->is_dense)
135 if (!pcl_isfinite (input_->points[cp].x) ||
136 !pcl_isfinite (input_->points[cp].y) ||
137 !pcl_isfinite (input_->points[cp].z))
141 const uint8_t* pt_data =
reinterpret_cast<const uint8_t*
> (&input_->points[cp]);
142 float distance_value = 0;
143 memcpy (&distance_value, pt_data + fields[distance_idx].offset,
sizeof (
float));
145 if (filter_limit_negative_)
148 if ((distance_value < filter_limit_max_) && (distance_value > filter_limit_min_))
154 if ((distance_value > filter_limit_max_) || (distance_value < filter_limit_min_))
158 int ijk0 =
static_cast<int> (floor (input_->points[cp].x * inverse_leaf_size_[0]) -
static_cast<float> (min_b_[0]));
159 int ijk1 =
static_cast<int> (floor (input_->points[cp].y * inverse_leaf_size_[1]) -
static_cast<float> (min_b_[1]));
160 int ijk2 =
static_cast<int> (floor (input_->points[cp].z * inverse_leaf_size_[2]) -
static_cast<float> (min_b_[2]));
163 int idx = ijk0 * divb_mul_[0] + ijk1 * divb_mul_[1] + ijk2 * divb_mul_[2];
165 Leaf& leaf = leaves_[idx];
166 if (leaf.nr_points == 0)
168 leaf.centroid.resize (centroid_size);
169 leaf.centroid.setZero ();
172 Eigen::Vector3d pt3d (input_->points[cp].x, input_->points[cp].y, input_->points[cp].z);
176 leaf.cov_ += pt3d * pt3d.transpose ();
179 if (!downsample_all_data_)
181 Eigen::Vector4f pt (input_->points[cp].x, input_->points[cp].y, input_->points[cp].z, 0);
182 leaf.centroid.template head<4> () += pt;
187 Eigen::VectorXf centroid = Eigen::VectorXf::Zero (centroid_size);
193 const pcl::RGB& rgb = *
reinterpret_cast<const RGB*
> (
reinterpret_cast<const char*
> (&input_->points[cp]) + rgba_index);
194 centroid[centroid_size - 4] = rgb.a;
195 centroid[centroid_size - 3] = rgb.r;
196 centroid[centroid_size - 2] = rgb.g;
197 centroid[centroid_size - 1] = rgb.b;
199 leaf.centroid += centroid;
208 for (
size_t cp = 0; cp < input_->points.size (); ++cp)
210 if (!input_->is_dense)
212 if (!pcl_isfinite (input_->points[cp].x) ||
213 !pcl_isfinite (input_->points[cp].y) ||
214 !pcl_isfinite (input_->points[cp].z))
217 int ijk0 =
static_cast<int> (floor (input_->points[cp].x * inverse_leaf_size_[0]) -
static_cast<float> (min_b_[0]));
218 int ijk1 =
static_cast<int> (floor (input_->points[cp].y * inverse_leaf_size_[1]) -
static_cast<float> (min_b_[1]));
219 int ijk2 =
static_cast<int> (floor (input_->points[cp].z * inverse_leaf_size_[2]) -
static_cast<float> (min_b_[2]));
222 int idx = ijk0 * divb_mul_[0] + ijk1 * divb_mul_[1] + ijk2 * divb_mul_[2];
225 Leaf& leaf = leaves_[idx];
226 if (leaf.nr_points == 0)
228 leaf.centroid.resize (centroid_size);
229 leaf.centroid.setZero ();
232 Eigen::Vector3d pt3d (input_->points[cp].x, input_->points[cp].y, input_->points[cp].z);
236 leaf.cov_ += pt3d * pt3d.transpose ();
239 if (!downsample_all_data_)
241 Eigen::Vector4f pt (input_->points[cp].x, input_->points[cp].y, input_->points[cp].z, 0);
242 leaf.centroid.template head<4> () += pt;
247 Eigen::VectorXf centroid = Eigen::VectorXf::Zero (centroid_size);
253 const pcl::RGB& rgb = *
reinterpret_cast<const RGB*
> (
reinterpret_cast<const char*
> (&input_->points[cp]) + rgba_index);
254 centroid[centroid_size - 4] = rgb.a;
255 centroid[centroid_size - 3] = rgb.r;
256 centroid[centroid_size - 2] = rgb.g;
257 centroid[centroid_size - 1] = rgb.b;
259 leaf.centroid += centroid;
266 output.
points.reserve (leaves_.size ());
268 voxel_centroids_leaf_indices_.reserve (leaves_.size ());
270 if (save_leaf_layout_)
271 leaf_layout_.resize (div_b_[0] * div_b_[1] * div_b_[2], -1);
274 Eigen::SelfAdjointEigenSolver<Eigen::Matrix3d> eigensolver;
275 Eigen::Matrix3d eigen_val;
276 Eigen::Vector3d pt_sum;
279 double min_covar_eigvalue;
281 for (
typename std::map<size_t, Leaf>::iterator it = leaves_.begin (); it != leaves_.end (); ++it)
285 Leaf& leaf = it->second;
288 leaf.centroid /=
static_cast<float> (leaf.nr_points);
292 leaf.mean_ /= leaf.nr_points;
296 if (leaf.nr_points >= min_points_per_voxel_)
298 if (save_leaf_layout_)
299 leaf_layout_[it->first] = cp++;
304 if (!downsample_all_data_)
306 output.
points.back ().x = leaf.centroid[0];
307 output.
points.back ().y = leaf.centroid[1];
308 output.
points.back ().z = leaf.centroid[2];
316 pcl::RGB& rgb = *
reinterpret_cast<RGB*
> (
reinterpret_cast<char*
> (&output.
points.back ()) + rgba_index);
317 rgb.a = leaf.centroid[centroid_size - 4];
318 rgb.r = leaf.centroid[centroid_size - 3];
319 rgb.g = leaf.centroid[centroid_size - 2];
320 rgb.b = leaf.centroid[centroid_size - 1];
326 voxel_centroids_leaf_indices_.push_back (
static_cast<int> (it->first));
329 leaf.cov_ = (leaf.cov_ - 2 * (pt_sum * leaf.mean_.transpose ())) / leaf.nr_points + leaf.mean_ * leaf.mean_.transpose ();
330 leaf.cov_ *= (leaf.nr_points - 1.0) / leaf.nr_points;
333 eigensolver.compute (leaf.cov_);
334 eigen_val = eigensolver.eigenvalues ().asDiagonal ();
335 leaf.evecs_ = eigensolver.eigenvectors ();
337 if (eigen_val (0, 0) < 0 || eigen_val (1, 1) < 0 || eigen_val (2, 2) <= 0)
345 min_covar_eigvalue = min_covar_eigvalue_mult_ * eigen_val (2, 2);
346 if (eigen_val (0, 0) < min_covar_eigvalue)
348 eigen_val (0, 0) = min_covar_eigvalue;
350 if (eigen_val (1, 1) < min_covar_eigvalue)
352 eigen_val (1, 1) = min_covar_eigvalue;
355 leaf.cov_ = leaf.evecs_ * eigen_val * leaf.evecs_.inverse ();
357 leaf.evals_ = eigen_val.diagonal ();
359 leaf.icov_ = leaf.cov_.inverse ();
360 if (leaf.icov_.maxCoeff () == std::numeric_limits<float>::infinity ( )
361 || leaf.icov_.minCoeff () == -std::numeric_limits<float>::infinity ( ) )
369 output.
width =
static_cast<uint32_t
> (output.
points.size ());
373 template<
typename Po
intT>
int
380 Eigen::Vector4i ijk (
static_cast<int> (floor (reference_point.x / leaf_size_[0])),
381 static_cast<int> (floor (reference_point.y / leaf_size_[1])),
382 static_cast<int> (floor (reference_point.z / leaf_size_[2])), 0);
383 Eigen::Array4i diff2min = min_b_ - ijk;
384 Eigen::Array4i diff2max = max_b_ - ijk;
385 neighbors.reserve (relative_coordinates.cols ());
389 for (
int ni = 0; ni < relative_coordinates.cols (); ni++)
391 Eigen::Vector4i displacement = (Eigen::Vector4i () << relative_coordinates.col (ni), 0).finished ();
393 if ((diff2min <= displacement.array ()).all () && (diff2max >= displacement.array ()).all ())
395 typename std::map<size_t, Leaf>::iterator leaf_iter = leaves_.find (((ijk + displacement - min_b_).dot (divb_mul_)));
396 if (leaf_iter != leaves_.end () && leaf_iter->second.nr_points >= min_points_per_voxel_)
398 LeafConstPtr leaf = &(leaf_iter->second);
399 neighbors.push_back (leaf);
404 return (
static_cast<int> (neighbors.size ()));
408 template<
typename Po
intT>
void
413 int pnt_per_cell = 1000;
415 boost::normal_distribution<> nd (0.0, leaf_size_.head (3).norm ());
416 boost::variate_generator<boost::mt19937&, boost::normal_distribution<> > var_nor (rng, nd);
418 Eigen::LLT<Eigen::Matrix3d> llt_of_cov;
419 Eigen::Matrix3d cholesky_decomp;
420 Eigen::Vector3d cell_mean;
421 Eigen::Vector3d rand_point;
422 Eigen::Vector3d dist_point;
425 for (
typename std::map<size_t, Leaf>::iterator it = leaves_.begin (); it != leaves_.end (); ++it)
427 Leaf& leaf = it->second;
429 if (leaf.nr_points >= min_points_per_voxel_)
431 cell_mean = leaf.mean_;
432 llt_of_cov.compute (leaf.cov_);
433 cholesky_decomp = llt_of_cov.matrixL ();
436 for (
int i = 0; i < pnt_per_cell; i++)
438 rand_point = Eigen::Vector3d (var_nor (), var_nor (), var_nor ());
439 dist_point = cell_mean + cholesky_decomp * rand_point;
440 cell_cloud.
push_back (
PointXYZ (
static_cast<float> (dist_point (0)),
static_cast<float> (dist_point (1)),
static_cast<float> (dist_point (2))));
446 #define PCL_INSTANTIATE_VoxelGridCovariance(T) template class PCL_EXPORTS pcl::VoxelGridCovariance<T>;
448 #endif // PCL_VOXEL_GRID_COVARIANCE_IMPL_H_