Point Cloud Library (PCL)  1.9.1-dev
ppf_registration.hpp
1 /*
2  * Software License Agreement (BSD License)
3  *
4  * Point Cloud Library (PCL) - www.pointclouds.org
5  * Copyright (c) 2011, Alexandru-Eugen Ichim
6  * Willow Garage, Inc
7  * Copyright (c) 2012-, Open Perception, Inc.
8  *
9  * All rights reserved.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  * * Redistributions of source code must retain the above copyright
16  * notice, this list of conditions and the following disclaimer.
17  * * Redistributions in binary form must reproduce the above
18  * copyright notice, this list of conditions and the following
19  * disclaimer in the documentation and/or other materials provided
20  * with the distribution.
21  * * Neither the name of the copyright holder(s) nor the names of its
22  * contributors may be used to endorse or promote products derived
23  * from this software without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
29  * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
31  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
33  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
35  * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36  * POSSIBILITY OF SUCH DAMAGE.
37  *
38  * $Id$
39  *
40  */
41 
42 
43 #ifndef PCL_REGISTRATION_IMPL_PPF_REGISTRATION_H_
44 #define PCL_REGISTRATION_IMPL_PPF_REGISTRATION_H_
45 
46 #include <pcl/registration/ppf_registration.h>
47 #include <pcl/features/ppf.h>
48 #include <pcl/common/transforms.h>
49 
50 #include <pcl/features/pfh.h>
51 //////////////////////////////////////////////////////////////////////////////////////////////
52 template <typename PointSource, typename PointTarget> void
54 {
56 
57  scene_search_tree_ = typename pcl::KdTreeFLANN<PointTarget>::Ptr (new pcl::KdTreeFLANN<PointTarget>);
58  scene_search_tree_->setInputCloud (target_);
59 }
60 
61 //////////////////////////////////////////////////////////////////////////////////////////////
62 template <typename PointSource, typename PointTarget> void
64 {
65  if (!search_method_)
66  {
67  PCL_ERROR("[pcl::PPFRegistration::computeTransformation] Search method not set - skipping computeTransformation!\n");
68  return;
69  }
70 
71  if (guess != Eigen::Matrix4f::Identity ())
72  {
73  PCL_ERROR("[pcl::PPFRegistration::computeTransformation] setting initial transform (guess) not implemented!\n");
74  }
75 
76  PoseWithVotesList voted_poses;
77  std::vector <std::vector <unsigned int> > accumulator_array;
78  accumulator_array.resize (input_->points.size ());
79 
80  size_t aux_size = static_cast<size_t> (floor (2 * M_PI / search_method_->getAngleDiscretizationStep ()));
81  for (size_t i = 0; i < input_->points.size (); ++i)
82  {
83  std::vector<unsigned int> aux (aux_size);
84  accumulator_array[i] = aux;
85  }
86  PCL_INFO ("Accumulator array size: %u x %u.\n", accumulator_array.size (), accumulator_array.back ().size ());
87 
88  // Consider every <scene_reference_point_sampling_rate>-th point as the reference point => fix s_r
89  float f1, f2, f3, f4;
90  for (size_t scene_reference_index = 0; scene_reference_index < target_->points.size (); scene_reference_index += scene_reference_point_sampling_rate_)
91  {
92  Eigen::Vector3f scene_reference_point = target_->points[scene_reference_index].getVector3fMap (),
93  scene_reference_normal = target_->points[scene_reference_index].getNormalVector3fMap ();
94 
95  float rotation_angle_sg = acosf (scene_reference_normal.dot (Eigen::Vector3f::UnitX ()));
96  bool parallel_to_x_sg = (scene_reference_normal.y() == 0.0f && scene_reference_normal.z() == 0.0f);
97  Eigen::Vector3f rotation_axis_sg = (parallel_to_x_sg)?(Eigen::Vector3f::UnitY ()):(scene_reference_normal.cross (Eigen::Vector3f::UnitX ()). normalized());
98  Eigen::AngleAxisf rotation_sg (rotation_angle_sg, rotation_axis_sg);
99  Eigen::Affine3f transform_sg (Eigen::Translation3f ( rotation_sg * ((-1) * scene_reference_point)) * rotation_sg);
100 
101  // For every other point in the scene => now have pair (s_r, s_i) fixed
102  std::vector<int> indices;
103  std::vector<float> distances;
104  scene_search_tree_->radiusSearch (target_->points[scene_reference_index],
105  search_method_->getModelDiameter () /2,
106  indices,
107  distances);
108  for(size_t i = 0; i < indices.size (); ++i)
109 // for(size_t i = 0; i < target_->points.size (); ++i)
110  {
111  //size_t scene_point_index = i;
112  size_t scene_point_index = indices[i];
113  if (scene_reference_index != scene_point_index)
114  {
115  if (/*pcl::computePPFPairFeature*/pcl::computePairFeatures (target_->points[scene_reference_index].getVector4fMap (),
116  target_->points[scene_reference_index].getNormalVector4fMap (),
117  target_->points[scene_point_index].getVector4fMap (),
118  target_->points[scene_point_index].getNormalVector4fMap (),
119  f1, f2, f3, f4))
120  {
121  std::vector<std::pair<size_t, size_t> > nearest_indices;
122  search_method_->nearestNeighborSearch (f1, f2, f3, f4, nearest_indices);
123 
124  // Compute alpha_s angle
125  Eigen::Vector3f scene_point = target_->points[scene_point_index].getVector3fMap ();
126 
127  Eigen::Vector3f scene_point_transformed = transform_sg * scene_point;
128  float alpha_s = atan2f ( -scene_point_transformed(2), scene_point_transformed(1));
129  if (sin (alpha_s) * scene_point_transformed(2) < 0.0f)
130  alpha_s *= (-1);
131  alpha_s *= (-1);
132 
133  // Go through point pairs in the model with the same discretized feature
134  for (std::vector<std::pair<size_t, size_t> >::iterator v_it = nearest_indices.begin (); v_it != nearest_indices.end (); ++ v_it)
135  {
136  size_t model_reference_index = v_it->first,
137  model_point_index = v_it->second;
138  // Calculate angle alpha = alpha_m - alpha_s
139  float alpha = search_method_->alpha_m_[model_reference_index][model_point_index] - alpha_s;
140  unsigned int alpha_discretized = static_cast<unsigned int> (floor (alpha) + floor (M_PI / search_method_->getAngleDiscretizationStep ()));
141  accumulator_array[model_reference_index][alpha_discretized] ++;
142  }
143  }
144  else PCL_ERROR ("[pcl::PPFRegistration::computeTransformation] Computing pair feature vector between points %u and %u went wrong.\n", scene_reference_index, scene_point_index);
145  }
146  }
147 
148  size_t max_votes_i = 0, max_votes_j = 0;
149  unsigned int max_votes = 0;
150 
151  for (size_t i = 0; i < accumulator_array.size (); ++i)
152  for (size_t j = 0; j < accumulator_array.back ().size (); ++j)
153  {
154  if (accumulator_array[i][j] > max_votes)
155  {
156  max_votes = accumulator_array[i][j];
157  max_votes_i = i;
158  max_votes_j = j;
159  }
160  // Reset accumulator_array for the next set of iterations with a new scene reference point
161  accumulator_array[i][j] = 0;
162  }
163 
164  Eigen::Vector3f model_reference_point = input_->points[max_votes_i].getVector3fMap (),
165  model_reference_normal = input_->points[max_votes_i].getNormalVector3fMap ();
166  float rotation_angle_mg = acosf (model_reference_normal.dot (Eigen::Vector3f::UnitX ()));
167  bool parallel_to_x_mg = (model_reference_normal.y() == 0.0f && model_reference_normal.z() == 0.0f);
168  Eigen::Vector3f rotation_axis_mg = (parallel_to_x_mg)?(Eigen::Vector3f::UnitY ()):(model_reference_normal.cross (Eigen::Vector3f::UnitX ()). normalized());
169  Eigen::AngleAxisf rotation_mg (rotation_angle_mg, rotation_axis_mg);
170  Eigen::Affine3f transform_mg (Eigen::Translation3f ( rotation_mg * ((-1) * model_reference_point)) * rotation_mg);
171  Eigen::Affine3f max_transform =
172  transform_sg.inverse () *
173  Eigen::AngleAxisf ((static_cast<float> (max_votes_j) - floorf (static_cast<float> (M_PI) / search_method_->getAngleDiscretizationStep ())) * search_method_->getAngleDiscretizationStep (), Eigen::Vector3f::UnitX ()) *
174  transform_mg;
175 
176  voted_poses.push_back (PoseWithVotes (max_transform, max_votes));
177  }
178  PCL_DEBUG ("Done with the Hough Transform ...\n");
179 
180  // Cluster poses for filtering out outliers and obtaining more precise results
181  PoseWithVotesList results;
182  clusterPoses (voted_poses, results);
183 
184  pcl::transformPointCloud (*input_, output, results.front ().pose);
185 
186  transformation_ = final_transformation_ = results.front ().pose.matrix ();
187  converged_ = true;
188 }
189 
190 
191 //////////////////////////////////////////////////////////////////////////////////////////////
192 template <typename PointSource, typename PointTarget> void
195 {
196  PCL_INFO ("Clustering poses ...\n");
197  // Start off by sorting the poses by the number of votes
198  sort(poses.begin (), poses.end (), poseWithVotesCompareFunction);
199 
200  std::vector<PoseWithVotesList> clusters;
201  std::vector<std::pair<size_t, unsigned int> > cluster_votes;
202  for (size_t poses_i = 0; poses_i < poses.size(); ++ poses_i)
203  {
204  bool found_cluster = false;
205  for (size_t clusters_i = 0; clusters_i < clusters.size(); ++ clusters_i)
206  {
207  if (posesWithinErrorBounds (poses[poses_i].pose, clusters[clusters_i].front ().pose))
208  {
209  found_cluster = true;
210  clusters[clusters_i].push_back (poses[poses_i]);
211  cluster_votes[clusters_i].second += poses[poses_i].votes;
212  break;
213  }
214  }
215 
216  if (found_cluster == false)
217  {
218  // Create a new cluster with the current pose
219  PoseWithVotesList new_cluster;
220  new_cluster.push_back (poses[poses_i]);
221  clusters.push_back (new_cluster);
222  cluster_votes.push_back (std::pair<size_t, unsigned int> (clusters.size () - 1, poses[poses_i].votes));
223  }
224  }
225 
226  // Sort clusters by total number of votes
227  std::sort (cluster_votes.begin (), cluster_votes.end (), clusterVotesCompareFunction);
228  // Compute pose average and put them in result vector
229  /// @todo some kind of threshold for determining whether a cluster has enough votes or not...
230  /// now just taking the first three clusters
231  result.clear ();
232  size_t max_clusters = (clusters.size () < 3) ? clusters.size () : 3;
233  for (size_t cluster_i = 0; cluster_i < max_clusters; ++ cluster_i)
234  {
235  PCL_INFO ("Winning cluster has #votes: %d and #poses voted: %d.\n", cluster_votes[cluster_i].second, clusters[cluster_votes[cluster_i].first].size ());
236  Eigen::Vector3f translation_average (0.0, 0.0, 0.0);
237  Eigen::Vector4f rotation_average (0.0, 0.0, 0.0, 0.0);
238  for (typename PoseWithVotesList::iterator v_it = clusters[cluster_votes[cluster_i].first].begin (); v_it != clusters[cluster_votes[cluster_i].first].end (); ++ v_it)
239  {
240  translation_average += v_it->pose.translation ();
241  /// averaging rotations by just averaging the quaternions in 4D space - reference "On Averaging Rotations" by CLAUS GRAMKOW
242  rotation_average += Eigen::Quaternionf (v_it->pose.rotation ()).coeffs ();
243  }
244 
245  translation_average /= static_cast<float> (clusters[cluster_votes[cluster_i].first].size ());
246  rotation_average /= static_cast<float> (clusters[cluster_votes[cluster_i].first].size ());
247 
248  Eigen::Affine3f transform_average;
249  transform_average.translation ().matrix () = translation_average;
250  transform_average.linear ().matrix () = Eigen::Quaternionf (rotation_average).normalized().toRotationMatrix ();
251 
252  result.push_back (PoseWithVotes (transform_average, cluster_votes[cluster_i].second));
253  }
254 }
255 
256 
257 //////////////////////////////////////////////////////////////////////////////////////////////
258 template <typename PointSource, typename PointTarget> bool
260  Eigen::Affine3f &pose2)
261 {
262  float position_diff = (pose1.translation () - pose2.translation ()).norm ();
263  Eigen::AngleAxisf rotation_diff_mat ((pose1.rotation ().inverse ().lazyProduct (pose2.rotation ()).eval()));
264 
265  float rotation_diff_angle = fabsf (rotation_diff_mat.angle ());
266 
267  if (position_diff < clustering_position_diff_threshold_ && rotation_diff_angle < clustering_rotation_diff_threshold_)
268  return true;
269  else return false;
270 }
271 
272 
273 //////////////////////////////////////////////////////////////////////////////////////////////
274 template <typename PointSource, typename PointTarget> bool
277 {
278  return (a.votes > b.votes);
279 }
280 
281 
282 //////////////////////////////////////////////////////////////////////////////////////////////
283 template <typename PointSource, typename PointTarget> bool
285  const std::pair<size_t, unsigned int> &b)
286 {
287  return (a.second > b.second);
288 }
289 
290 //#define PCL_INSTANTIATE_PPFRegistration(PointSource,PointTarget) template class PCL_EXPORTS pcl::PPFRegistration<PointSource, PointTarget>;
291 
292 #endif // PCL_REGISTRATION_IMPL_PPF_REGISTRATION_H_
PointCloudTarget::ConstPtr PointCloudTargetConstPtr
boost::shared_ptr< KdTreeFLANN< PointT, Dist > > Ptr
Definition: kdtree_flann.h:88
void setInputTarget(const PointCloudTargetConstPtr &cloud) override
Provide a pointer to the input target (e.g., the point cloud that we want to align the input source t...
void transformPointCloud(const pcl::PointCloud< PointT > &cloud_in, pcl::PointCloud< PointT > &cloud_out, const Eigen::Transform< Scalar, 3, Eigen::Affine > &transform, bool copy_all_fields=true)
Apply an affine transform defined by an Eigen Transform.
Definition: transforms.hpp:215
const PointT & front() const
Definition: point_cloud.h:469
Registration represents the base registration class for general purpose, ICP-like methods...
Definition: registration.h:61
Structure for storing a pose (represented as an Eigen::Affine3f) and an integer for counting votes...
std::vector< PoseWithVotes, Eigen::aligned_allocator< PoseWithVotes > > PoseWithVotesList
Class that registers two point clouds based on their sets of PPFSignatures.
PCL_EXPORTS bool computePairFeatures(const Eigen::Vector4f &p1, const Eigen::Vector4f &n1, const Eigen::Vector4f &p2, const Eigen::Vector4f &n2, float &f1, float &f2, float &f3, float &f4)
Compute the 4-tuple representation containing the three angles and one distance between two points re...