Beispiel #1
0
void EKFOA::process(const double delta_t, cv::Mat & frame, Eigen::Vector3d & rW, Eigen::Vector4d & qWR, Eigen::Matrix3d & axes_orientation_and_confidence, std::vector<Point3d> (& XYZs)[3], Delaunay & triangulation, Point3d & closest_point){
	double time_total;
	std::vector<cv::Point2f> features_to_add;
	std::vector<Features_extra> features_extra;

	/*
	 * EKF prediction (state and measurement prediction)
	 */
	time_total = (double)cv::getTickCount();
	double time_prediction = (double)cv::getTickCount();
	filter.predict_state_and_covariance(delta_t);
	filter.compute_features_h(cam, features_extra);
	time_prediction = (double)cv::getTickCount() - time_prediction;
//	std::cout << "predict = " << time_prediction/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;

	/*
	 * Sense and map management (delete features from EKF)
	 */
	double time_tracker = (double)cv::getTickCount();
	motion_tracker.process(frame, features_extra, features_to_add);
	//TODO: Why is optical flow returning points outside the image???
	time_tracker = (double)cv::getTickCount() - time_tracker;

	time_total = time_total + time_tracker; //do not count the time spent by the tracker

	//Delete no longer seen features from the state, covariance matrix and the features_extra:
	double time_del = (double)cv::getTickCount();
	filter.delete_features(features_extra);
	time_del = (double)cv::getTickCount() - time_del;
//	std::cout << "delete  = " << time_del/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;

	/*
	 * EKF Update step and map management (add new features to EKF)
	 */
	double time_update = (double)cv::getTickCount();
	filter.update(cam, features_extra);
	time_update = (double)cv::getTickCount() - time_update;
//	std::cout << "update  = " << time_update/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;


	//Add new features
	double time_add = (double)cv::getTickCount();
	filter.add_features_inverse_depth(cam, features_to_add);
	time_add = (double)cv::getTickCount() - time_add;
//	std::cout << "add_fea = " << time_add/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;


	/*
	 * Triangulation, surface and GUI data setting:
	 */
	double time_triangulation = (double)cv::getTickCount();

	std::vector< std::pair<Point2d, size_t> > triangle_list;
	std::list<Triangle> triangles_list_3d;

	const Eigen::VectorXd & x_k_k = filter.x_k_k();
	const Eigen::MatrixXd & p_k_k = filter.p_k_k();

	//Set the position, so the GUI can draw it:
	rW = x_k_k.segment<3>(0);//current position

	//Set the axes orientation and confidence:
	axes_orientation_and_confidence.setIdentity();//axes_orientation_and_confidence stores in each column one axis (X, Y, Z)
	axes_orientation_and_confidence *= 5; //make the lines larger, so they are actually informative
	//Apply rotation matrix:
	Eigen::Matrix3d qWR_R;//Rotation matrix of current orientation quaternion
	qWR = x_k_k.segment<4>(3);
	MotionModel::quaternion_matrix(qWR, qWR_R);
	axes_orientation_and_confidence.applyOnTheLeft(qWR_R); // == R * axes_orientation_and_confidence
	for (int axis=0 ; axis<axes_orientation_and_confidence.cols() ; axis++){
		//Set the length to be 3*sigma:
		axes_orientation_and_confidence.col(axis) *= 3*std::sqrt(p_k_k(axis, axis)); //the first 3 positions of the cov matrix define the confidence for the position
		//Translate origin:
		axes_orientation_and_confidence.col(axis) += rW;
	}

	int num_features = (x_k_k.rows()-13)/6;
	XYZs[0].resize(num_features);
	XYZs[1].resize(num_features);
	XYZs[2].resize(num_features);


	//Compute the 3d positions and inverse depth variances of all the points in the state
	int i=0; //Feature counter
	for (int start_feature=13 ; start_feature<x_k_k.rows() ; start_feature+=6){
		const int feature_inv_depth_index = start_feature + 5;

		//As with any normal distribution, nearly all (99.73%) of the possible depths lie within three standard deviations of the mean!
		const double sigma_3 = std::sqrt(p_k_k(feature_inv_depth_index, feature_inv_depth_index)); //sqrt(depth_variance)

		const Eigen::VectorXd & yi = x_k_k.segment(start_feature, 6);
		Eigen::VectorXd point_close(x_k_k.segment(start_feature, 6));
		Eigen::VectorXd point_far(x_k_k.segment(start_feature, 6));

		//Change the depth of the feature copy, so that it is possible to represent the range between -3*sigma and 3*sigma:
		point_close(5) += sigma_3;
		point_far(5) -= sigma_3;

		Eigen::Vector3d XYZ_mu = (Feature::compute_cartesian(yi)); //mu (mean)
		Eigen::Vector3d XYZ_close = (Feature::compute_cartesian(point_close)); //mean + 3*sigma. (since inverted signs are also inverted)
		Eigen::Vector3d XYZ_far = (Feature::compute_cartesian(point_far)); //mean - 3*sigma

		//The center of the model is ALWAYS the current position of the camera/robot, so have to 'cancel' the current orientation (R_inv) and translation (rWC = x_k_k.head(3)):
		//Note: It is nicer to do this in the GUI class, as it is only a presention/perspective change. But due to the structure, it was easier to do it here.
		XYZs[0][i] = Point3d(XYZ_mu(0), XYZ_mu(1), XYZ_mu(2)); //mu (mean)
		XYZs[1][i] = Point3d(XYZ_close(0), XYZ_close(1), XYZ_close(2)); //mean + 3*sigma. (since inverted signs are also inverted)
		XYZs[2][i] = Point3d(XYZ_far(0), XYZ_far(1), XYZ_far(2)); //mean - 3*sigma

		//If the size that contains the 99.73% of the inverse depth distribution is smaller than the current inverse depth, add it to the surface:
		const double size_sigma_3 = std::abs(1.0/(x_k_k(feature_inv_depth_index)-sigma_3) - 1.0/(x_k_k(feature_inv_depth_index)+sigma_3));
		if (size_sigma_3 < 1/x_k_k(feature_inv_depth_index)){
			triangle_list.push_back(std::make_pair(Point2d(features_extra[i].z(0), features_extra[i].z(1)), i));
		}

		if (x_k_k(feature_inv_depth_index) < 0 ){
			std::cout << "feature behind the camera!!! : idx=" << i << ", value=" << x_k_k(feature_inv_depth_index) << std::endl;
		}
		i++;
	}

	triangulation.insert(triangle_list.begin(), triangle_list.end());

	cv::Scalar delaunay_color = cv::Scalar(255, 0, 0); //blue
	for(Delaunay::Finite_faces_iterator fit = triangulation.finite_faces_begin(); fit != triangulation.finite_faces_end(); ++fit) {
		const Delaunay::Face_handle & face = fit;
		//face->vertex(i)->info() = index of the point in the observation list.
		line(frame, features_extra[face->vertex(0)->info()].z_cv, features_extra[face->vertex(1)->info()].z_cv, delaunay_color, 1);
		line(frame, features_extra[face->vertex(1)->info()].z_cv, features_extra[face->vertex(2)->info()].z_cv, delaunay_color, 1);
		line(frame, features_extra[face->vertex(2)->info()].z_cv, features_extra[face->vertex(0)->info()].z_cv, delaunay_color, 1);

		//Add the face of the linked 3d points of this 2d triangle:
		triangles_list_3d.push_back(Triangle(XYZs[1][face->vertex(0)->info()], XYZs[1][face->vertex(1)->info()], XYZs[1][face->vertex(2)->info()])); //XYZs[1] == close
	}

	// constructs AABB tree
	Tree tree(triangles_list_3d.begin(), triangles_list_3d.end());

	if (tree.size()>0){
		// compute closest point and squared distance
		Point3d point_query(rW[0], rW[1], rW[2]);
		closest_point = tree.closest_point(point_query);
//		FT sqd = tree.squared_distance(point_query);

		Eigen::Vector3d last_displacement_vector = last_position - rW;

//		double repealing_force = 0;
//		if (std::sqrt(sqd) < 0.2){
//			std::cout << "can crash! " << std::endl;
//			repealing_force = 1/std::sqrt(sqd);
//		}
//		std::cout << "distance = [distance, " << std::sqrt(sqd) << "];" << std::endl;
//		std::cout << "repealing_force = [repealing_force, " << repealing_force << "];" << std::endl;
	}


	//remember this position
	last_position = rW;
//	std::cout << "certaint= " << p_k_k.diagonal().sum() << std::endl;

	time_triangulation = (double)cv::getTickCount() - time_triangulation;
//	std::cout << "Triang  = " << time_triangulation/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;

	time_total = (double)cv::getTickCount() - time_total;
//	std::cout << "EKF     = " << time_total/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;

//	std::cout << "tracker = " << time_tracker/((double)cvGetTickFrequency()*1000.) << "ms" << std::endl;
}
Beispiel #2
0
void
opengv::absolute_pose::modules::gp3p_main(
    const Eigen::Matrix3d & f,
    const Eigen::Matrix3d & v,
    const Eigen::Matrix3d & p,
    transformations_t & solutions)
{
  Eigen::Matrix<double,48,85> groebnerMatrix =
      Eigen::Matrix<double,48,85>::Zero();
  gp3p::init(groebnerMatrix,f,v,p);
  gp3p::compute(groebnerMatrix);

  Eigen::Matrix<double,8,8> M = Eigen::Matrix<double,8,8>::Zero();
  M.block<6,8>(0,0) = -groebnerMatrix.block<6,8>(36,77);
  M(6,0) = 1.0;
  M(7,6) = 1.0;

  Eigen::ComplexEigenSolver< Eigen::Matrix<double,8,8> > Eig(M,true);
  Eigen::Matrix<std::complex<double>,8,1> D = Eig.eigenvalues();
  Eigen::Matrix<std::complex<double>,8,8> V = Eig.eigenvectors();

  for( int c = 0; c < V.cols(); c++ )
  {
    std::complex<double> eigValue = D[c];

    if( eigValue.imag() < 0.0001 )
    {
      cayley_t cayley;
      Eigen::Vector3d n;

      for(size_t i = 0; i < 3; i++)
      {
        std::complex<double> cay = V(i+4,c)/V(7,c);
        cayley[2-i] = cay.real();
        std::complex<double> depth = V(i+1,c)/V(7,c);
        n[2-i] = depth.real();
      }

      rotation_t rotation = math::cayley2rot(cayley);
      //the groebner problem was set up to find the transpose!
      rotation.transposeInPlace();

      point_t center_cam = Eigen::Vector3d::Zero();
      point_t center_world = Eigen::Vector3d::Zero();
      for( size_t i = 0; i < (size_t) f.cols(); i++ )
      {
        point_t temp = rotation*(n[i]*f.col(i)+v.col(i));
        center_cam = center_cam + temp;
        center_world = center_world + p.col(i);
      }

      center_cam = center_cam/f.cols();
      center_world = center_world/f.cols();
      translation_t translation = center_world - center_cam;

      transformation_t transformation;
      transformation.block<3,3>(0,0) = rotation;
      transformation.col(3) = translation;
      solutions.push_back(transformation);
    }
  }
}