00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025
00026
00027
00028
00029
00030
00031
00032
00033
00034
00035
00036 #include <jsk_pcl_ros/region_adjacency_graph.h>
00037 #include <jsk_topic_tools/log_utils.h>
00038
00039 namespace jsk_pcl_ros
00040 {
00041 RegionAdjacencyGraph::RegionAdjacencyGraph()
00042 {
00043
00044 }
00045
00046 void RegionAdjacencyGraph::generateRAG(
00047 const std::vector<pcl::PointCloud<PointT>::Ptr> &cloud_clusters,
00048 const std::vector<pcl::PointCloud<pcl::Normal>::Ptr> &normal_clusters,
00049 const pcl::PointCloud<pcl::PointXYZ>::Ptr centroids,
00050 std::vector<std::vector<int> > &neigbor_indices,
00051 const int edge_weight_criteria)
00052 {
00053 if (cloud_clusters.empty() || normal_clusters.empty() ||
00054 centroids->empty() || neigbor_indices.empty()) {
00055 ROS_ERROR("ERROR: Cannot Generate RAG of empty data...");
00056 return;
00057 }
00058 const int comparision_points_size = 100;
00059 if (cloud_clusters.size() == neigbor_indices.size()) {
00060 std::vector<VertexDescriptor> vertex_descriptor;
00061 for (int j = 0; j < centroids->size(); j++) {
00062 VertexDescriptor v_des = boost::add_vertex(
00063 VertexProperty(j, centroids->points[j], -1), this->graph_);
00064 vertex_descriptor.push_back(v_des);
00065 }
00066 for (int j = 0; j < neigbor_indices.size(); j++) {
00067 VertexDescriptor r_vd = vertex_descriptor[j];
00068 std::vector<Eigen::Vector3f> center_point;
00069 std::vector<Eigen::Vector3f> center_normal;
00070 cv::Mat r_histogram;
00071 if (edge_weight_criteria == RAG_EDGE_WEIGHT_CONVEX_CRITERIA) {
00072 this->sampleRandomPointsFromCloudCluster(
00073 cloud_clusters[j],
00074 normal_clusters[j],
00075 center_point,
00076 center_normal,
00077 comparision_points_size);
00078 } else if (edge_weight_criteria == RAG_EDGE_WEIGHT_DISTANCE) {
00079 if (cloud_clusters[j]->size() > sizeof(char) &&
00080 normal_clusters[j]->size() > sizeof(char)) {
00081 this->computeCloudClusterRPYHistogram(
00082 cloud_clusters[j],
00083 normal_clusters[j],
00084 r_histogram);
00085 }
00086 } else {
00087 ROS_ERROR("Incorrect Measurement type");
00088 return;
00089 }
00090 for (int i = 0; i < neigbor_indices[j].size(); i++) {
00091 int n_index = neigbor_indices[j][i];
00092 VertexDescriptor vd = vertex_descriptor[n_index];
00093 float distance = 0.0f;
00094 if (edge_weight_criteria == RAG_EDGE_WEIGHT_CONVEX_CRITERIA) {
00095 std::vector<Eigen::Vector3f> n1_point;
00096 std::vector<Eigen::Vector3f> n1_normal;
00097 this->sampleRandomPointsFromCloudCluster(
00098 cloud_clusters[n_index],
00099 normal_clusters[n_index],
00100 n1_point,
00101 n1_normal,
00102 comparision_points_size);
00103
00104 int commonIndex = this->getCommonNeigbour(
00105 neigbor_indices[j],
00106 neigbor_indices[n_index]);
00107 if (commonIndex == -1) {
00108
00109
00110 } else {
00111 std::vector<Eigen::Vector3f> n2_point;
00112 std::vector<Eigen::Vector3f> n2_normal;
00113 this->sampleRandomPointsFromCloudCluster(
00114 cloud_clusters[commonIndex],
00115 normal_clusters[commonIndex],
00116 n2_point,
00117 n2_normal,
00118 comparision_points_size);
00119 std::vector<std::vector<Eigen::Vector3f> > _points;
00120 std::vector<std::vector<Eigen::Vector3f> > _normals;
00121 _points.push_back(center_point);
00122 _points.push_back(n1_point);
00123
00124 _normals.push_back(center_normal);
00125 _normals.push_back(n1_normal);
00126
00127 distance = this->getCloudClusterWeightFunction<float>(
00128 _points, _normals);
00129 }
00130 } else if (edge_weight_criteria == RAG_EDGE_WEIGHT_DISTANCE) {
00131 if (cloud_clusters[j]->size() > sizeof(char) &&
00132 cloud_clusters[n_index]->size() > sizeof(char)) {
00133 cv::Mat n_histogram;
00134 this->computeCloudClusterRPYHistogram(
00135 cloud_clusters[n_index],
00136 normal_clusters[n_index],
00137 n_histogram);
00138 distance = static_cast<float>(
00139 cv::compareHist(
00140 r_histogram, n_histogram, CV_COMP_CORREL));
00141 } else {
00142 distance = 0.0f;
00143 }
00144 } else {
00145 distance = 0.0f;
00146 }
00147 if (r_vd != vd) {
00148 bool found = false;
00149 EdgeDescriptor e_descriptor;
00150 boost::tie(e_descriptor, found) = boost::edge(r_vd, vd, this->graph_);
00151 if (!found) {
00152 boost::add_edge(
00153 r_vd, vd, EdgeProperty(distance), this->graph_);
00154 }
00155 }
00156 }
00157 }
00158 } else {
00159 ROS_WARN("Elements not same size..");
00160 }
00161 }
00162
00163 template<typename T>
00164 T RegionAdjacencyGraph::getCloudClusterWeightFunction(
00165 const std::vector<std::vector<Eigen::Vector3f> > &_points,
00166 const std::vector<std::vector<Eigen::Vector3f> > &_normal)
00167 {
00168 #define ANGLE_THRESHOLD (10)
00169 if (_points.size() == 2 && _points.size() == _normal.size()) {
00170 T weights_ = -1.0f;
00171 int concave_ = 0;
00172 int convex_ = 0;
00173 for (int i = 0; i < _points[0].size(); i++) {
00174 T convexC_ij = this->convexityCriterion<T>(
00175 _points[0][i], _points[1][i], _normal[0][i], _normal[1][i]);
00176 float angle_ = getVectorAngle(_normal[0][i], _normal[1][i]);
00177 if (convexC_ij < 0.0f && angle_ < ANGLE_THRESHOLD) {
00178 convexC_ij = abs(convexC_ij);
00179 }
00180 if (convexC_ij > 0.0) {
00181 convex_++;
00182 }
00183 if (convexC_ij <= 0.0 || std::isnan(convexC_ij)) {
00184 concave_++;
00185 }
00186
00187
00188
00189
00190 }
00191 if (concave_ < convex_ + 20) {
00192 weights_ = 1.0f;
00193 }
00194 return weights_;
00195 } else if (_points.size() == 3) {
00196 T weights_ = FLT_MIN;
00197 for (int i = 0; i < _points[0].size(); i++) {
00198 T convexC_ij = this->convexityCriterion<T>(
00199 _points[0][i], _points[1][i], _normal[0][i], _normal[1][i]);
00200 T convexC_ic = this->convexityCriterion<T>(
00201 _points[0][i], _points[2][i], _normal[0][i], _normal[2][i]);
00202 T convexC_jc = this->convexityCriterion<T>(
00203 _points[1][i], _points[2][i], _normal[1][i], _normal[2][i]);
00204
00205
00206
00207
00208 weights_ = std::max(convexC_ij,
00209 std::max(convexC_ic, convexC_jc));
00210 }
00211 return weights_;
00212 }
00213 }
00214
00215 float RegionAdjacencyGraph::getVectorAngle(
00216 const Eigen::Vector3f &vector1,
00217 const Eigen::Vector3f &vector2,
00218 bool indegree)
00219 {
00220 float angle_ = acos(vector1.dot(vector2));
00221 if (indegree) {
00222 return angle_ * 180/M_PI;
00223 } else {
00224 return angle_;
00225 }
00226 }
00227
00228 template<typename T>
00229 T RegionAdjacencyGraph::convexityCriterion(
00230 const Eigen::Vector3f ¢er_point,
00231 const Eigen::Vector3f &n1_point,
00232 const Eigen::Vector3f ¢er_normal,
00233 const Eigen::Vector3f &neigbour_normal)
00234 {
00235 Eigen::Vector3f difference_ = center_point - n1_point;
00236 difference_ /= difference_.norm();
00237 T convexityc = static_cast<T>(center_normal.dot(difference_) -
00238 neigbour_normal.dot(difference_));
00239 return convexityc;
00240 }
00241
00242 void RegionAdjacencyGraph::sampleRandomPointsFromCloudCluster(
00243 pcl::PointCloud<PointT>::Ptr cloud,
00244 pcl::PointCloud<pcl::Normal>::Ptr normal,
00245 std::vector<Eigen::Vector3f> &point_vector,
00246 std::vector<Eigen::Vector3f> &normal_vector,
00247 int gen_sz)
00248 {
00249 for (int i = 0; i < std::max(gen_sz, (int)cloud->size()); i++) {
00250 int _idx = rand() % cloud->size();
00251 Eigen::Vector3f cv = cloud->points[_idx].getVector3fMap();
00252 Eigen::Vector3f nv = Eigen::Vector3f(
00253 normal->points[_idx].normal_x,
00254 normal->points[_idx].normal_y,
00255 normal->points[_idx].normal_z);
00256 point_vector.push_back(cv);
00257 normal_vector.push_back(nv);
00258 }
00259 }
00260
00261 void RegionAdjacencyGraph::splitMergeRAG(
00262 const int _threshold)
00263 {
00264 if (num_vertices(this->graph_) == 0) {
00265 ROS_ERROR("ERROR: Cannot Merge Empty RAG ...");
00266 return;
00267 }
00268 IndexMap index_map = get(boost::vertex_index, this->graph_);
00269 EdgePropertyAccess edge_weights = get(boost::edge_weight, this->graph_);
00270 VertexIterator i, end;
00271 int label = -1;
00272 for (boost::tie(i, end) = vertices(this->graph_); i != end; i++) {
00273 if (this->graph_[*i].v_label == -1) {
00274 graph_[*i].v_label = ++label;
00275 }
00276 AdjacencyIterator ai, a_end;
00277 boost::tie(ai, a_end) = adjacent_vertices(*i, this->graph_);
00278 for (; ai != a_end; ++ai) {
00279 bool found = false;
00280 EdgeDescriptor e_descriptor;
00281 boost::tie(e_descriptor, found) = boost::edge(*i, *ai, this->graph_);
00282 if (found) {
00283 EdgeValue edge_val = boost::get(
00284 boost::edge_weight, this->graph_, e_descriptor);
00285 float weights_ = edge_val;
00286 if (weights_ < _threshold) {
00287 remove_edge(e_descriptor, this->graph_);
00288 } else {
00289 if (this->graph_[*ai].v_label == -1) {
00290 this->graph_[*ai].v_label = this->graph_[*i].v_label;
00291 }
00292 }
00293 }
00294 }
00295 }
00296 #ifdef DEBUG
00297
00298 std::cout << "\nPRINT INFO. \n --Graph Size: "
00299 << num_vertices(this->graph_) <<
00300 std::endl << "--Total Label: " << label << "\n\n";
00301 #endif // DEBUG
00302 }
00303
00304 int RegionAdjacencyGraph::getCommonNeigbour(
00305 const std::vector<int> &c1_neigbour,
00306 const std::vector<int> &c2_neigbour)
00307 {
00308 int commonIndex = -1;
00309 for (int j = 0; j < c1_neigbour.size(); j++) {
00310 int c1_val = c1_neigbour[j];
00311 for (int i = 0; i < c2_neigbour.size(); i++) {
00312 int c2_val = c2_neigbour[i];
00313 if (c1_val == c2_val) {
00314 commonIndex = c1_val;
00315 break;
00316 }
00317 }
00318 }
00319 return commonIndex;
00320 }
00321
00322 void RegionAdjacencyGraph::getCloudClusterLabels(
00323 std::vector<int> &labelMD)
00324 {
00325 labelMD.clear();
00326 VertexIterator i, end;
00327 for (boost::tie(i, end) = vertices(this->graph_); i != end; ++i) {
00328 labelMD.push_back(static_cast<int>(this->graph_[*i].v_label));
00329 }
00330 }
00331
00332 void RegionAdjacencyGraph::printGraph(
00333 const Graph &_graph)
00334 {
00335 VertexIterator i, end;
00336 for (boost::tie(i, end) = vertices(_graph); i != end; ++i) {
00337 AdjacencyIterator ai, a_end;
00338 boost::tie(ai, a_end) = adjacent_vertices(*i, _graph);
00339 std::cout << *i << "\t" << _graph[*i].v_label << std::endl;
00340 }
00341 }
00342
00343 void RegionAdjacencyGraph::computeCloudClusterRPYHistogram(
00344 const pcl::PointCloud<PointT>::Ptr _cloud,
00345 const pcl::PointCloud<pcl::Normal>::Ptr _normal,
00346 cv::Mat &_histogram)
00347 {
00348 pcl::VFHEstimation<PointT,
00349 pcl::Normal,
00350 pcl::VFHSignature308> vfh;
00351 vfh.setInputCloud(_cloud);
00352 vfh.setInputNormals(_normal);
00353 pcl::search::KdTree<PointT>::Ptr tree(
00354 new pcl::search::KdTree<PointT>);
00355 vfh.setSearchMethod(tree);
00356 pcl::PointCloud<pcl::VFHSignature308>::Ptr _vfhs(
00357 new pcl::PointCloud<pcl::VFHSignature308>());
00358 vfh.compute(*_vfhs);
00359 _histogram = cv::Mat(sizeof(char), 308, CV_32F);
00360 for (int i = 0; i < _histogram.cols; i++) {
00361 _histogram.at<float>(0, i) = _vfhs->points[0].histogram[i];
00362 }
00363 float curvature_ = 0.0f;
00364 for (int i = 0; i < _normal->size(); i++) {
00365 curvature_ += _normal->points[i].curvature;
00366 }
00367 curvature_ /= static_cast<float>(_normal->size());
00368 cv::normalize(
00369 _histogram, _histogram, 0, 1, cv::NORM_MINMAX, -1, cv::Mat());
00370 }
00371 }