/**
    Kreis an die Punkte (2xN-Array) anfitten.
*/
Eigen::Vector3f fit_circle(const Eigen::Array2Xf& points) {
    Eigen::Array2f mean = points.rowwise().mean();
    Eigen::Array2Xf uv(points.colwise() - mean);
    
    Eigen::Array2f Sp2 = uv.square().rowwise().sum();
    Eigen::Array2f Sp3 = uv.cube().rowwise().sum();
    
    float Suv = (uv.row(0) * uv.row(1)).sum();
    float Suvv = (uv.row(0) * uv.row(1).square()).sum();
    float Svuu = (uv.row(0).square() * uv.row(1)).sum();
    
    // Rechte Seite berechnen
    Eigen::Vector2f b((Sp3 + Eigen::Array2f(Suvv, Svuu)) / 2);
    
    // Linke Seite berechnen
    Eigen::Matrix2f A;
    A << Sp2(0), Suv, Suv, Sp2(1);
    
    // Gleichungssystem Lösen
    Eigen::Array2f mid = A.inverse() * b;
    
    // Quadrat des Radius berechnen
    float radius = sqrt(mid.square().sum() + Sp2.sum() / points.cols());
    return (Eigen::Vector3f() << (mid + mean), radius).finished();
}
  void FindObjectOnPlane::generateAngles(
    const cv::Mat& blob_image, const cv::Point2f& test_point,
    std::vector<double>& angles,
    std::vector<double>& max_x,
    std::vector<double>& max_y,
    const image_geometry::PinholeCameraModel& model,
    const jsk_recognition_utils::Plane::Ptr& plane)
  {
    angles.clear();
    const double angle_step = 3;
    std::vector<cv::Point> indices;
    for (int j = 0; j < blob_image.rows; j++) {
        for (int i = 0; i < blob_image.cols; i++) {
          if (blob_image.at<uchar>(j, i) != 0) { // need to check
            indices.push_back(cv::Point(i, j));
          }
        }
    }
    for (double angle = -180; angle < 180; angle += angle_step) {
      Eigen::Vector2f ux(cos(angle/180*M_PI), sin(angle/180*M_PI));
      //Eigen::Vector2f uy(sin(angle/180*M_PI), -cos(angle/180*M_PI));
      cv::Point2d uy_end = getUyEnd(test_point, cv::Point2d(test_point.x + ux[0], test_point.y + ux[1]),
                                    model,
                                    plane);
      Eigen::Vector2f uy(uy_end.x - test_point.x, uy_end.y - test_point.y);

      Eigen::Matrix2f A;
      A << ux[0], uy[0],
        ux[1], uy[1];
      bool excluded = false;
      double max_alpha = -DBL_MAX;
      double max_beta = -DBL_MAX;
      for (size_t i = 0; i < indices.size(); i++) {
        Eigen::Vector2f p_q = Eigen::Vector2f(indices[i].x, indices[i].y) - Eigen::Vector2f(test_point.x, test_point.y);
        Eigen::Vector2f a_b = A.inverse() * p_q;
        double alpha = a_b[0];
        double beta = a_b[1];
        if (alpha < 0 || beta < 0) {
          excluded = true;
          break;
        }
        if (alpha > max_alpha) {
          max_alpha = alpha;
        }
        if (beta > max_beta) {
          max_beta = beta;
        }
        
      }
      if (!excluded) {
        angles.push_back(angle);
        max_x.push_back(max_alpha);
        max_y.push_back(max_beta);
      }
    }
  }
Ejemplo n.º 3
0
    auto match(vfloat2 p, vfloat2 tr_prediction,
	       F A, F B, GD Ag,
	       const int winsize,
	       const float min_ev_th,
	       const int max_interations,
	       const float convergence_delta)
    {
      typedef typename F::value_type V;
      int WS = winsize;
      int ws = winsize;
      int hws = ws/2;

      // Gradient matrix
      Eigen::Matrix2f G = Eigen::Matrix2f::Zero();
      int cpt = 0;
      for(int r = -hws; r <= hws; r++)
	for(int c = -hws; c <= hws; c++)
	  {
	    vfloat2 n = p + vfloat2(r, c);
	    if (A.has(n.cast<int>()))
	      {
		Eigen::Matrix2f m;
		auto g = Ag.linear_interpolate(n);
		float gx = g[0];
		float gy = g[1];
		m <<
		  gx * gx, gx * gy,
		  gx * gy, gy * gy;
		G += m;
		cpt++;
	      }
	  }

      // Check minimum eigenvalue.
      float min_ev = 99999.f;
      auto ev = (G / cpt).eigenvalues();
      for (int i = 0; i < ev.size(); i++)
	if (fabs(ev[i].real()) < min_ev) min_ev = fabs(ev[i].real());

      if (min_ev < min_ev_th)
	return std::pair<vfloat2, float>(vfloat2(-1,-1), FLT_MAX);

      Eigen::Matrix2f G1 = G.inverse();

      // Precompute gs and as.
      vfloat2 prediction_ = p + tr_prediction;
      vfloat2 v = prediction_;
      Eigen::Vector2f nk = Eigen::Vector2f::Ones();

      char gs_buffer[WS * WS * sizeof(vfloat2)];
      vfloat2* gs = (vfloat2*) gs_buffer;
      // was: vfloat2 gs[WS * WS];

      typedef plus_promotion<V> S;
      char as_buffer[WS * WS * sizeof(S)];
      S* as = (S*) as_buffer;
      // was: S as[WS * WS];
      {
	for(int i = 0, r = -hws; r <= hws; r++)
	  {
	    for(int c = -hws; c <= hws; c++)
	      {
		vfloat2 n = p + vfloat2(r, c);
		if (Ag.has(n.cast<int>()))
		  {
		    gs[i] = Ag.linear_interpolate(n).template cast<float>();
		    as[i] = cast<S>(A.linear_interpolate(n));
		  }
		i++;
	      }
	  }
      }
      auto domain = B.domain();// - border(hws + 1);

      // Gradient descent
      for (int k = 0; k <= max_interations && nk.norm() >= convergence_delta; k++)
	{
	  Eigen::Vector2f bk = Eigen::Vector2f::Zero();
	  // Temporal difference.
	  int i = 0;
	  for(int r = -hws; r <= hws; r++)
	    {
	      for(int c = -hws; c <= hws; c++)          
		{
		  vfloat2 n = p + vfloat2(r, c);
		  if (Ag.has(n.cast<int>()))
		    {
		      vfloat2 n2 = v + vfloat2(r, c);
		      auto g = gs[i];
		      float dt = (cast<float>(as[i]) - cast<float>(B.linear_interpolate(n2)));
		      bk += Eigen::Vector2f{g[0] * dt, g[1] * dt};
		    }
		  i++;
		}
	    }

	  nk = G1 * bk;
	  v += vfloat2{nk[0], nk[1]};

	  if (!domain.has(v.cast<int>()))
	    return std::pair<vfloat2, float>(vfloat2(0, 0), FLT_MAX);
	}

      // Compute the SSD.
      float err = 0;
      for(int r = -hws; r <= hws; r++)
	for(int c = -hws; c <= hws; c++)
	  {
	    vfloat2 n2 = v + vfloat2(r, c);
	    int i = (r+hws) * ws + (c+hws);
	    {
	      err += fabs(cast<float>(as[i] - cast<S>(B.linear_interpolate(n2))));
	      cpt++;
	    }
	  }

      return std::pair<vfloat2, float>(v - p, err / (cpt));


    }
 void PoseWithCovarianceStampedToGaussianPointCloud::convert(const geometry_msgs::PoseWithCovarianceStamped::ConstPtr& msg)
 {
   boost::mutex::scoped_lock lock(mutex_);
   pcl::PointCloud<pcl::PointXYZ>::Ptr cloud(new pcl::PointCloud<pcl::PointXYZ>);
   Eigen::Vector2f mean;
   Eigen::Matrix2f S;
   if (cut_plane_ == "xy" || cut_plane_ == "flipped_xy") {
     mean = Eigen::Vector2f(msg->pose.pose.position.x, msg->pose.pose.position.y);
     S << msg->pose.covariance[0], msg->pose.covariance[1], 
       msg->pose.covariance[6], msg->pose.covariance[7];
   }
   else if (cut_plane_ == "yz" || cut_plane_ == "flipped_yz") {
     mean = Eigen::Vector2f(msg->pose.pose.position.y, msg->pose.pose.position.z);
     S << msg->pose.covariance[7], msg->pose.covariance[8], 
       msg->pose.covariance[13], msg->pose.covariance[14];
   }
   else if (cut_plane_ == "zx" || cut_plane_ == "flipped_zx") {
     mean = Eigen::Vector2f(msg->pose.pose.position.z, msg->pose.pose.position.x);
     S << msg->pose.covariance[14], msg->pose.covariance[12], 
       msg->pose.covariance[0], msg->pose.covariance[2];
   }
   double max_sigma = 0;
   for (size_t i = 0; i < 2; i++) {
     for (size_t j = 0; j < 2; j++) {
       double sigma = sqrt(S(i, j));
       if (max_sigma < sigma) {
         max_sigma = sigma;
       }
     }
   }
   Eigen::Matrix2f S_inv = S.inverse();
   double dx = 6.0 * max_sigma / sampling_num_;
   double dy = 6.0 * max_sigma / sampling_num_;
   for (double x = - 3.0 * max_sigma; x <= 3.0 * max_sigma; x += dx) {
     for (double y = - 3.0 * max_sigma; y <= 3.0 * max_sigma; y += dy) {
       Eigen::Vector2f diff(x, y);
       Eigen::Vector2f input = diff + mean;
       float z = gaussian(input, mean, S, S_inv);
       pcl::PointXYZ p;
       if (cut_plane_ == "xy") {
         p.x = input[0];
         p.y = input[1];
         p.z = z + msg->pose.pose.position.z;
       }
       else if (cut_plane_ == "yz") {
         p.y = input[0];
         p.z = input[1];
         p.x = z + msg->pose.pose.position.x;
       }
       else if (cut_plane_ == "zx") {
         p.z = input[0];
         p.x = input[1];
         p.y = z + msg->pose.pose.position.y;
       }
       else if (cut_plane_ == "flipped_xy") {
         p.x = input[0];
         p.y = input[1];
         p.z = -z + msg->pose.pose.position.z;
       }
       else if (cut_plane_ == "flipped_yz") {
         p.y = input[0];
         p.z = input[1];
         p.x = -z + msg->pose.pose.position.x;
       }
       else if (cut_plane_ == "flipped_zx") {
         p.z = input[0];
         p.x = input[1];
         p.y = -z + msg->pose.pose.position.y;
       }
       cloud->points.push_back(p);
     }
   }
   sensor_msgs::PointCloud2 ros_cloud;
   pcl::toROSMsg(*cloud, ros_cloud);
   ros_cloud.header = msg->header;
   pub_.publish(ros_cloud);
 }