37 #include <pcl/recognition/hv/hv_go.h>
39 #include <pcl/common/time.h>
40 #include <pcl/point_types.h>
42 template<
typename Po
intT,
typename NormalT>
45 unsigned int min_pts_per_cluster,
unsigned int max_pts_per_cluster = (std::numeric_limits<int>::max) ())
50 PCL_ERROR(
"[pcl::extractEuclideanClusters] Tree built for a different point cloud dataset\n");
55 PCL_ERROR(
"[pcl::extractEuclideanClusters] Number of points in the input point cloud different than normals!\n");
60 std::vector<bool> processed (cloud.
points.size (),
false);
62 std::vector<int> nn_indices;
63 std::vector<float> nn_distances;
65 int size =
static_cast<int> (cloud.
points.size ());
66 for (
int i = 0; i < size; ++i)
71 std::vector<unsigned int> seed_queue;
73 seed_queue.push_back (i);
77 while (sq_idx < static_cast<int> (seed_queue.size ()))
80 if (normals.
points[seed_queue[sq_idx]].curvature > curvature_threshold)
87 if (!tree->
radiusSearch (seed_queue[sq_idx], tolerance, nn_indices, nn_distances))
93 for (
size_t j = 1; j < nn_indices.size (); ++j)
95 if (processed[nn_indices[j]])
98 if (normals.
points[nn_indices[j]].curvature > curvature_threshold)
106 double dot_p = normals.
points[seed_queue[sq_idx]].normal[0] * normals.
points[nn_indices[j]].normal[0]
107 + normals.
points[seed_queue[sq_idx]].normal[1] * normals.
points[nn_indices[j]].normal[1]
108 + normals.
points[seed_queue[sq_idx]].normal[2] * normals.
points[nn_indices[j]].normal[2];
110 if (fabs (acos (dot_p)) < eps_angle)
112 processed[nn_indices[j]] =
true;
113 seed_queue.push_back (nn_indices[j]);
121 if (seed_queue.size () >= min_pts_per_cluster && seed_queue.size () <= max_pts_per_cluster)
124 r.
indices.resize (seed_queue.size ());
125 for (
size_t j = 0; j < seed_queue.size (); ++j)
132 clusters.push_back (r);
137 template<
typename ModelT,
typename SceneT>
145 updateExplainedVector (recognition_models_[changed]->explained_, recognition_models_[changed]->explained_distances_, explained_by_RM_,
146 explained_by_RM_distance_weighted, 1.f);
147 updateUnexplainedVector (recognition_models_[changed]->unexplained_in_neighborhood, recognition_models_[changed]->unexplained_in_neighborhood_weights,
148 unexplained_by_RM_neighboorhods, recognition_models_[changed]->explained_, explained_by_RM_, 1.f);
149 updateCMDuplicity(recognition_models_[changed]->complete_cloud_occupancy_indices_, complete_cloud_occupancy_by_RM_, 1.f);
153 updateExplainedVector (recognition_models_[changed]->explained_, recognition_models_[changed]->explained_distances_, explained_by_RM_,
154 explained_by_RM_distance_weighted, -1.f);
155 updateUnexplainedVector (recognition_models_[changed]->unexplained_in_neighborhood, recognition_models_[changed]->unexplained_in_neighborhood_weights,
156 unexplained_by_RM_neighboorhods, recognition_models_[changed]->explained_, explained_by_RM_, -1.f);
157 updateCMDuplicity(recognition_models_[changed]->complete_cloud_occupancy_indices_, complete_cloud_occupancy_by_RM_, -1.f);
161 int duplicity = getDuplicity ();
162 float good_info = getExplainedValue ();
164 float unexplained_info = getPreviousUnexplainedValue ();
165 float bad_info =
static_cast<float> (getPreviousBadInfo ())
166 + (recognition_models_[changed]->outliers_weight_ * static_cast<float> (recognition_models_[changed]->bad_information_)) * sign;
168 setPreviousBadInfo (bad_info);
170 int n_active_hyp = 0;
171 for(
size_t i=0; i < active.size(); i++) {
176 float duplicity_cm =
static_cast<float> (getDuplicityCM ()) * w_occupied_multiple_cm_;
177 return static_cast<mets::gol_type
> ((good_info - bad_info -
static_cast<float> (duplicity) - unexplained_info - duplicity_cm - static_cast<float> (n_active_hyp)) * -1.f);
181 template<
typename ModelT,
typename SceneT>
185 recognition_models_.clear ();
186 unexplained_by_RM_neighboorhods.clear ();
187 explained_by_RM_distance_weighted.clear ();
188 explained_by_RM_.clear ();
191 complete_cloud_occupancy_by_RM_.clear ();
194 mask_.resize (complete_models_.size ());
195 for (
size_t i = 0; i < complete_models_.size (); i++)
198 indices_.resize (complete_models_.size ());
200 NormalEstimator_ n3d;
204 normals_tree->setInputCloud (scene_cloud_downsampled_);
206 n3d.setRadiusSearch (radius_normals_);
207 n3d.setSearchMethod (normals_tree);
208 n3d.setInputCloud (scene_cloud_downsampled_);
209 n3d.compute (*scene_normals_);
213 for (
size_t i = 0; i < scene_normals_->points.size (); ++i)
215 if (!pcl_isfinite (scene_normals_->points[i].normal_x) || !pcl_isfinite (scene_normals_->points[i].normal_y)
216 || !pcl_isfinite (scene_normals_->points[i].normal_z))
219 scene_normals_->points[j] = scene_normals_->points[i];
220 scene_cloud_downsampled_->points[j] = scene_cloud_downsampled_->points[i];
225 scene_normals_->points.resize (j);
226 scene_normals_->width = j;
227 scene_normals_->height = 1;
229 scene_cloud_downsampled_->points.resize (j);
230 scene_cloud_downsampled_->width = j;
231 scene_cloud_downsampled_->height = 1;
233 explained_by_RM_.resize (scene_cloud_downsampled_->points.size (), 0);
234 explained_by_RM_distance_weighted.resize (scene_cloud_downsampled_->points.size (), 0.f);
235 unexplained_by_RM_neighboorhods.resize (scene_cloud_downsampled_->points.size (), 0.f);
242 scene_downsampled_tree_->setInputCloud (scene_cloud_downsampled_);
244 std::vector<pcl::PointIndices> clusters;
245 double eps_angle_threshold = 0.2;
247 float curvature_threshold = 0.045f;
249 extractEuclideanClustersSmooth<SceneT, pcl::Normal> (*scene_cloud_downsampled_, *scene_normals_, inliers_threshold_ * 2.f, scene_downsampled_tree_,
250 clusters, eps_angle_threshold, curvature_threshold, min_points);
253 clusters_cloud_->points.resize (scene_cloud_downsampled_->points.size ());
254 clusters_cloud_->width = scene_cloud_downsampled_->width;
255 clusters_cloud_->height = 1;
257 for (
size_t i = 0; i < scene_cloud_downsampled_->points.size (); i++)
260 p.getVector3fMap () = scene_cloud_downsampled_->points[i].getVector3fMap ();
262 clusters_cloud_->points[i] = p;
265 float intens_incr = 100.f /
static_cast<float> (clusters.size ());
266 float intens = intens_incr;
267 for (
size_t i = 0; i < clusters.size (); i++)
269 for (
size_t j = 0; j < clusters[i].indices.size (); j++)
271 clusters_cloud_->points[clusters[i].indices[j]].
intensity = intens;
274 intens += intens_incr;
281 recognition_models_.resize (complete_models_.size ());
283 for (
int i = 0; i < static_cast<int> (complete_models_.size ()); i++)
286 recognition_models_[valid].reset (
new RecognitionModel ());
287 if(addModel (visible_models_[i], complete_models_[i], recognition_models_[valid])) {
293 recognition_models_.resize(valid);
294 indices_.resize(valid);
298 ModelT min_pt_all, max_pt_all;
299 min_pt_all.x = min_pt_all.y = min_pt_all.z = std::numeric_limits<float>::max ();
300 max_pt_all.x = max_pt_all.y = max_pt_all.z = (std::numeric_limits<float>::max () - 0.001f) * -1;
302 for (
size_t i = 0; i < recognition_models_.size (); i++)
304 ModelT min_pt, max_pt;
306 if (min_pt.x < min_pt_all.x)
307 min_pt_all.x = min_pt.x;
309 if (min_pt.y < min_pt_all.y)
310 min_pt_all.y = min_pt.y;
312 if (min_pt.z < min_pt_all.z)
313 min_pt_all.z = min_pt.z;
315 if (max_pt.x > max_pt_all.x)
316 max_pt_all.x = max_pt.x;
318 if (max_pt.y > max_pt_all.y)
319 max_pt_all.y = max_pt.y;
321 if (max_pt.z > max_pt_all.z)
322 max_pt_all.z = max_pt.z;
325 int size_x, size_y, size_z;
326 size_x =
static_cast<int> (std::ceil (std::abs (max_pt_all.x - min_pt_all.x) / res_occupancy_grid_)) + 1;
327 size_y =
static_cast<int> (std::ceil (std::abs (max_pt_all.y - min_pt_all.y) / res_occupancy_grid_)) + 1;
328 size_z =
static_cast<int> (std::ceil (std::abs (max_pt_all.z - min_pt_all.z) / res_occupancy_grid_)) + 1;
330 complete_cloud_occupancy_by_RM_.resize (size_x * size_y * size_z, 0);
332 for (
size_t i = 0; i < recognition_models_.size (); i++)
335 std::map<int, bool> banned;
336 std::map<int, bool>::iterator banned_it;
338 for (
size_t j = 0; j < complete_models_[indices_[i]]->points.size (); j++)
340 int pos_x, pos_y, pos_z;
341 pos_x =
static_cast<int> (std::floor ((complete_models_[indices_[i]]->points[j].x - min_pt_all.x) / res_occupancy_grid_));
342 pos_y =
static_cast<int> (std::floor ((complete_models_[indices_[i]]->points[j].y - min_pt_all.y) / res_occupancy_grid_));
343 pos_z =
static_cast<int> (std::floor ((complete_models_[indices_[i]]->points[j].z - min_pt_all.z) / res_occupancy_grid_));
345 int idx = pos_z * size_x * size_y + pos_y * size_x + pos_x;
346 banned_it = banned.find (idx);
347 if (banned_it == banned.end ())
349 complete_cloud_occupancy_by_RM_[idx]++;
350 recognition_models_[i]->complete_cloud_occupancy_indices_.push_back (idx);
358 #pragma omp parallel for schedule(dynamic, 4) num_threads(omp_get_num_procs())
359 for (
int j = 0; j < static_cast<int> (recognition_models_.size ()); j++)
360 computeClutterCue (recognition_models_[j]);
366 for (
size_t i = 0; i < recognition_models_.size (); i++)
367 cc_[0].push_back (static_cast<int> (i));
371 template<
typename ModelT,
typename SceneT>
376 std::vector < boost::shared_ptr<RecognitionModel> > recognition_models_copy;
377 recognition_models_copy = recognition_models_;
379 recognition_models_.clear ();
381 for (
size_t j = 0; j < cc_indices.size (); j++)
383 recognition_models_.push_back (recognition_models_copy[cc_indices[j]]);
386 for (
size_t j = 0; j < recognition_models_.size (); j++)
388 boost::shared_ptr < RecognitionModel > recog_model = recognition_models_[j];
389 for (
size_t i = 0; i < recog_model->explained_.size (); i++)
391 explained_by_RM_[recog_model->explained_[i]]++;
392 explained_by_RM_distance_weighted[recog_model->explained_[i]] += recog_model->explained_distances_[i];
397 for (
size_t i = 0; i < recog_model->unexplained_in_neighborhood.size (); i++)
399 unexplained_by_RM_neighboorhods[recog_model->unexplained_in_neighborhood[i]] += recog_model->unexplained_in_neighborhood_weights[i];
404 int occupied_multiple = 0;
405 for(
size_t i=0; i < complete_cloud_occupancy_by_RM_.size(); i++) {
406 if(complete_cloud_occupancy_by_RM_[i] > 1) {
407 occupied_multiple+=complete_cloud_occupancy_by_RM_[i];
411 setPreviousDuplicityCM(occupied_multiple);
416 float good_information_ = getTotalExplainedInformation (explained_by_RM_, explained_by_RM_distance_weighted, &duplicity);
417 float bad_information_ = 0;
418 float unexplained_in_neighboorhod = getUnexplainedInformationInNeighborhood (unexplained_by_RM_neighboorhods, explained_by_RM_);
420 for (
size_t i = 0; i < initial_solution.size (); i++)
422 if (initial_solution[i])
423 bad_information_ += recognition_models_[i]->outliers_weight_ *
static_cast<float> (recognition_models_[i]->bad_information_);
426 setPreviousExplainedValue (good_information_);
427 setPreviousDuplicity (duplicity);
428 setPreviousBadInfo (bad_information_);
429 setPreviousUnexplainedValue (unexplained_in_neighboorhod);
432 model.cost_ =
static_cast<mets::gol_type
> ((good_information_ - bad_information_
433 -
static_cast<float> (duplicity)
434 - static_cast<float> (occupied_multiple) * w_occupied_multiple_cm_
435 -
static_cast<float> (recognition_models_.size ())
436 - unexplained_in_neighboorhod) * -1.f);
438 model.setSolution (initial_solution);
439 model.setOptimizer (
this);
440 SAModel best (model);
442 move_manager neigh (static_cast<int> (cc_indices.size ()));
444 mets::best_ever_solution best_recorder (best);
445 mets::noimprove_termination_criteria noimprove (max_iterations_);
446 mets::linear_cooling linear_cooling;
447 mets::simulated_annealing<move_manager> sa (model, best_recorder, neigh, noimprove, linear_cooling, initial_temp_, 1e-7, 2);
448 sa.setApplyAndEvaluate(
true);
455 best_seen_ =
static_cast<const SAModel&
> (best_recorder.best_seen ());
456 for (
size_t i = 0; i < best_seen_.solution_.size (); i++)
458 initial_solution[i] = best_seen_.solution_[i];
461 recognition_models_ = recognition_models_copy;
466 template<
typename ModelT,
typename SceneT>
472 for (
int c = 0; c < n_cc_; c++)
476 std::vector<bool> subsolution (cc_[c].size (),
true);
477 SAOptimize (cc_[c], subsolution);
478 for (
size_t i = 0; i < subsolution.size (); i++)
480 mask_[indices_[cc_[c][i]]] = (subsolution[i]);
485 template<
typename ModelT,
typename SceneT>
493 float size_model = resolution_;
496 voxel_grid.
setLeafSize (size_model, size_model, size_model);
497 voxel_grid.
filter (*(recog_model->cloud_));
501 voxel_grid2.
setLeafSize (size_model, size_model, size_model);
502 voxel_grid2.
filter (*(recog_model->complete_cloud_));
507 for (
size_t i = 0; i < recog_model->cloud_->points.size (); ++i)
509 if (!pcl_isfinite (recog_model->cloud_->points[i].x) || !pcl_isfinite (recog_model->cloud_->points[i].y)
510 || !pcl_isfinite (recog_model->cloud_->points[i].z))
513 recog_model->cloud_->points[j] = recog_model->cloud_->points[i];
517 recog_model->cloud_->points.resize (j);
518 recog_model->cloud_->width = j;
519 recog_model->cloud_->height = 1;
522 if (recog_model->cloud_->points.size () <= 0)
524 PCL_WARN(
"The model cloud has no points..\n");
531 NormalEstimator_ n3d;
534 n3d.setRadiusSearch (radius_normals_);
535 n3d.setSearchMethod (normals_tree);
536 n3d.setInputCloud ((recog_model->cloud_));
537 n3d.compute (*(recog_model->normals_));
541 for (
size_t i = 0; i < recog_model->normals_->points.size (); ++i)
543 if (!pcl_isfinite (recog_model->normals_->points[i].normal_x) || !pcl_isfinite (recog_model->normals_->points[i].normal_y)
544 || !pcl_isfinite (recog_model->normals_->points[i].normal_z))
547 recog_model->normals_->points[j] = recog_model->normals_->points[i];
548 recog_model->cloud_->points[j] = recog_model->cloud_->points[i];
552 recog_model->normals_->points.resize (j);
553 recog_model->normals_->width = j;
554 recog_model->normals_->height = 1;
556 recog_model->cloud_->points.resize (j);
557 recog_model->cloud_->width = j;
558 recog_model->cloud_->height = 1;
560 std::vector<int> explained_indices;
561 std::vector<float> outliers_weight;
562 std::vector<float> explained_indices_distances;
563 std::vector<float> unexplained_indices_weights;
565 std::vector<int> nn_indices;
566 std::vector<float> nn_distances;
568 std::map<int, boost::shared_ptr<std::vector<std::pair<int, float> > > > model_explains_scene_points;
569 std::map<int, boost::shared_ptr<std::vector<std::pair<int, float> > > >::iterator it;
571 outliers_weight.resize (recog_model->cloud_->points.size ());
572 recog_model->outlier_indices_.resize (recog_model->cloud_->points.size ());
575 for (
size_t i = 0; i < recog_model->cloud_->points.size (); i++)
577 if (!scene_downsampled_tree_->radiusSearch (recog_model->cloud_->points[i], inliers_threshold_, nn_indices, nn_distances, std::numeric_limits<int>::max ()))
580 outliers_weight[o] = regularizer_;
581 recog_model->outlier_indices_[o] =
static_cast<int> (i);
585 for (
size_t k = 0; k < nn_distances.size (); k++)
587 std::pair<int, float> pair = std::make_pair (i, nn_distances[k]);
588 it = model_explains_scene_points.find (nn_indices[k]);
589 if (it == model_explains_scene_points.end ())
591 boost::shared_ptr < std::vector<std::pair<int, float> > > vec (
new std::vector<std::pair<int, float> > ());
592 vec->push_back (pair);
593 model_explains_scene_points[nn_indices[k]] = vec;
596 it->second->push_back (pair);
602 outliers_weight.resize (o);
603 recog_model->outlier_indices_.resize (o);
605 recog_model->outliers_weight_ = (std::accumulate (outliers_weight.begin (), outliers_weight.end (), 0.f) / static_cast<float> (outliers_weight.size ()));
606 if (outliers_weight.size () == 0)
607 recog_model->outliers_weight_ = 1.f;
614 for (it = model_explains_scene_points.begin (); it != model_explains_scene_points.end (); it++, p++)
617 float min_d = std::numeric_limits<float>::min ();
618 for (
size_t i = 0; i < it->second->size (); i++)
620 if (it->second->at (i).second > min_d)
622 min_d = it->second->at (i).second;
627 float d = it->second->at (closest).second;
628 float d_weight = -(d * d / (inliers_threshold_)) + 1;
632 Eigen::Vector3f scene_p_normal = scene_normals_->points[it->first].getNormalVector3fMap ();
633 Eigen::Vector3f model_p_normal = recog_model->normals_->points[it->second->at (closest).first].getNormalVector3fMap ();
634 float dotp = scene_p_normal.dot (model_p_normal) * 1.f;
639 explained_indices.push_back (it->first);
640 explained_indices_distances.push_back (d_weight * dotp);
644 recog_model->bad_information_ =
static_cast<int> (recog_model->outlier_indices_.size ());
645 recog_model->explained_ = explained_indices;
646 recog_model->explained_distances_ = explained_indices_distances;
651 template<
typename ModelT,
typename SceneT>
657 float rn_sqr = radius_neighborhood_GO_ * radius_neighborhood_GO_;
658 std::vector<int> nn_indices;
659 std::vector<float> nn_distances;
661 std::vector < std::pair<int, int> > neighborhood_indices;
662 for (
int i = 0; i < static_cast<int> (recog_model->explained_.size ()); i++)
664 if (scene_downsampled_tree_->radiusSearch (scene_cloud_downsampled_->points[recog_model->explained_[i]], radius_neighborhood_GO_, nn_indices,
665 nn_distances, std::numeric_limits<int>::max ()))
667 for (
size_t k = 0; k < nn_distances.size (); k++)
669 if (nn_indices[k] != i)
670 neighborhood_indices.push_back (std::make_pair (nn_indices[k], i));
676 std::sort (neighborhood_indices.begin (), neighborhood_indices.end (),
677 boost::bind (&std::pair<int, int>::first, _1) < boost::bind (&std::pair<int, int>::first, _2));
680 neighborhood_indices.erase (
681 std::unique (neighborhood_indices.begin (), neighborhood_indices.end (),
682 boost::bind (&std::pair<int, int>::first, _1) == boost::bind (&std::pair<int, int>::first, _2)), neighborhood_indices.end ());
685 std::vector<int> exp_idces (recog_model->explained_);
686 std::sort (exp_idces.begin (), exp_idces.end ());
688 recog_model->unexplained_in_neighborhood.resize (neighborhood_indices.size ());
689 recog_model->unexplained_in_neighborhood_weights.resize (neighborhood_indices.size ());
693 for (
size_t i = 0; i < neighborhood_indices.size (); i++)
695 if ((j < exp_idces.size ()) && (neighborhood_indices[i].first == exp_idces[j]))
703 recog_model->unexplained_in_neighborhood[p] = neighborhood_indices[i].first;
705 if (clusters_cloud_->points[recog_model->explained_[neighborhood_indices[i].second]].intensity != 0.f
706 && (clusters_cloud_->points[recog_model->explained_[neighborhood_indices[i].second]].intensity
707 == clusters_cloud_->points[neighborhood_indices[i].first].intensity))
710 recog_model->unexplained_in_neighborhood_weights[p] = clutter_regularizer_;
716 float d =
static_cast<float> (pow (
717 (scene_cloud_downsampled_->points[recog_model->explained_[neighborhood_indices[i].second]].getVector3fMap ()
718 - scene_cloud_downsampled_->points[neighborhood_indices[i].first].getVector3fMap ()).norm (), 2));
719 float d_weight = -(d / rn_sqr) + 1;
722 Eigen::Vector3f scene_p_normal = scene_normals_->points[neighborhood_indices[i].first].getNormalVector3fMap ();
723 Eigen::Vector3f model_p_normal = scene_normals_->points[recog_model->explained_[neighborhood_indices[i].second]].getNormalVector3fMap ();
724 float dotp = scene_p_normal.dot (model_p_normal);
729 recog_model->unexplained_in_neighborhood_weights[p] = d_weight * dotp;
735 recog_model->unexplained_in_neighborhood_weights.resize (p);
736 recog_model->unexplained_in_neighborhood.resize (p);
740 #define PCL_INSTANTIATE_GoHV(T1,T2) template class PCL_EXPORTS pcl::GlobalHypothesesVerification<T1,T2>;
boost::shared_ptr< const PointCloud< PointT > > ConstPtr
boost::shared_ptr< std::vector< int > > IndicesPtr
virtual int radiusSearch(const PointT &point, double radius, std::vector< int > &k_indices, std::vector< float > &k_sqr_distances, unsigned int max_nn=0) const =0
Search for all the nearest neighbors of the query point in a given radius.
Class to measure the time spent in a scope.
std::vector< int > indices
void filter(PointCloud &output)
Calls the filtering method and returns the filtered dataset in output.
VoxelGrid assembles a local 3D grid over a given PointCloud, and downsamples + filters the data...
NormalEstimation estimates local surface properties (surface normals and curvatures)at each 3D point...
pcl::PCLHeader header
The point cloud header.
std::vector< PointT, Eigen::aligned_allocator< PointT > > points
The point data.
boost::shared_ptr< pcl::search::Search< PointT > > Ptr
boost::shared_ptr< KdTree< PointT > > Ptr
void getMinMax3D(const pcl::PointCloud< PointT > &cloud, PointT &min_pt, PointT &max_pt)
Get the minimum and maximum values on each of the 3 (x-y-z) dimensions in a given pointcloud...
virtual PointCloudConstPtr getInputCloud() const
Get a pointer to the input point cloud dataset.
virtual void setInputCloud(const PointCloudConstPtr &cloud)
Provide a pointer to the input dataset.
virtual void setInputCloud(const PointCloudConstPtr &cloud)
Provide a pointer to the input dataset.
A hypothesis verification method proposed in "A Global Hypotheses Verification Method for 3D Object R...
void setLeafSize(const Eigen::Vector4f &leaf_size)
Set the voxel grid leaf size.