inline Eigen::Matrix<double, Eigen::Dynamic, Eigen::Dynamic> wishart_rng(const double nu, const Eigen::Matrix<double,Eigen::Dynamic,Eigen::Dynamic>& S, RNG& rng) { static const char* function = "stan::prob::wishart_rng(%1%)"; using stan::math::check_size_match; using stan::math::check_positive; check_positive(function,nu,"degrees of freedom",(double*)0); check_size_match(function, S.rows(), "Rows of scale parameter", S.cols(), "columns of scale parameter", (double*)0); Eigen::Matrix<double,Eigen::Dynamic,Eigen::Dynamic> B(S.rows(), S.cols()); B.setZero(); for(int i = 0; i < S.cols(); i++) { B(i,i) = std::sqrt(chi_square_rng(nu - i, rng)); for(int j = 0; j < i; j++) B(j,i) = normal_rng(0,1,rng); } return stan::math::multiply_lower_tri_self_transpose(S.llt().matrixL() * B); }
bool CameraPoseFinderICP::minimizePointToPlaneErrFunc(unsigned level,Eigen::Matrix<float, 6, 1> &six_dof,const Mat44& cur_transform,const Mat44& last_transform_inv) { cudaCalPointToPlaneErrSolverParams( level,cur_transform,last_transform_inv, _camera_params_pyramid[level],AppParams::instance()->_icp_params.fDistThres, AppParams::instance()->_icp_params.fNormSinThres); CudaMap1D<float> buf=(CudaDeviceDataMan::instance()->rigid_align_buf_reduced).clone(CPU); int shift = 0; Eigen::Matrix<float, 6, 6, Eigen::RowMajor> ATA; Eigen::Matrix<float, 6, 1> ATb; for (int i = 0; i < 6; ++i) //rows { for (int j = i; j < 7; ++j) // cols + b { float value = buf.at(shift++); if (j == 6) { ATb(i) = value; } else { ATA(i, j) = value; ATA(j, i) = value; } } } //cout<<ATb<<endl; if (ATA.determinant()<1E-10) { return false; } six_dof = ATA.llt().solve(ATb).cast<float>(); return true; }
inline Eigen::VectorXd multi_normal_rng(const Eigen::Matrix<double,Eigen::Dynamic,1>& mu, const Eigen::Matrix<double,Eigen::Dynamic,Eigen::Dynamic>& S, RNG& rng) { using boost::variate_generator; using boost::normal_distribution; static const char* function = "stan::prob::multi_normal_rng(%1%)"; using stan::math::check_positive; using stan::math::check_finite; using stan::math::check_symmetric; check_positive(function, S.rows(), "Covariance matrix rows"); check_symmetric(function, S, "Covariance matrix"); check_finite(function, mu, "Location parameter"); variate_generator<RNG&, normal_distribution<> > std_normal_rng(rng, normal_distribution<>(0,1)); Eigen::VectorXd z(S.cols()); for(int i = 0; i < S.cols(); i++) z(i) = std_normal_rng(); return mu + S.llt().matrixL() * z; }
void CCubicGrids::extractRTFromBuffer(const cuda::GpuMat& cvgmSumBuf_, Eigen::Matrix3f* peimRw_, Eigen::Vector3f* peivTw_) const{ Mat cvmSumBuf; cvgmSumBuf_.download(cvmSumBuf); double* aHostTmp = (double*)cvmSumBuf.data; //declare A and b Eigen::Matrix<double, 6, 6, Eigen::RowMajor> A; Eigen::Matrix<double, 6, 1> b; //retrieve A and b from cvmSumBuf short sShift = 0; for (int i = 0; i < 6; ++i){ // rows for (int j = i; j < 7; ++j) { // cols + b double value = aHostTmp[sShift++]; if (j == 6) // vector b b.data()[i] = value; else A.data()[j * 6 + i] = A.data()[i * 6 + j] = value; }//for each col }//for each row //checking nullspace double dDet = A.determinant(); if (fabs(dDet) < 1e-15 || dDet != dDet){ if (dDet != dDet) PRINTSTR("Failure -- dDet cannot be qnan. "); //reset (); return; }//if dDet is rational //float maxc = A.maxCoeff(); Eigen::Matrix<float, 6, 1> result = A.llt().solve(b).cast<float>(); //Eigen::Matrix<float, 6, 1> result = A.jacobiSvd(ComputeThinU | ComputeThinV).solve(b); float alpha = result(0); float beta = result(1); float gamma = result(2); Eigen::Matrix3f Rinc = (Eigen::Matrix3f)Eigen::AngleAxisf(gamma, Eigen::Vector3f::UnitZ()) * Eigen::AngleAxisf(beta, Eigen::Vector3f::UnitY()) * Eigen::AngleAxisf(alpha, Eigen::Vector3f::UnitX()); Eigen::Vector3f tinc = result.tail<3>(); //compose //eivTwCur = Rinc * eivTwCur + tinc; //eimrmRwCur = Rinc * eimrmRwCur; Eigen::Vector3f eivTinv = -peimRw_->transpose()* (*peivTw_); Eigen::Matrix3f eimRinv = peimRw_->transpose(); eivTinv = Rinc * eivTinv + tinc; eimRinv = Rinc * eimRinv; *peivTw_ = -eimRinv.transpose() * eivTinv; *peimRw_ = eimRinv.transpose(); }
void dare(const Eigen::Matrix<double,xDim,xDim> &A, const Eigen::Matrix<double,xDim,uDim> &B, Eigen::Matrix<double,xDim,xDim> &P) { Eigen::Matrix<double,xDim,xDim> Ainv = A.inverse(); Eigen::Matrix<double,xDim,xDim> ABRB = Ainv * B * _R.llt().solve(B.transpose()); Eigen::Matrix<double,2*xDim,2*xDim> Z; Z.block(0,0,xDim,xDim) = Ainv; Z.block(0,xDim,xDim,xDim) = ABRB; Z.block(xDim,0,xDim,xDim) = _Q * Ainv; Z.block(xDim,xDim,xDim,xDim) = A.transpose() + _Q * ABRB; Eigen::ComplexEigenSolver <Eigen::Matrix<double,2*xDim,2*xDim> > ces; ces.compute(Z); Eigen::Matrix<std::complex<double>,2*xDim,1> eigVal = ces.eigenvalues(); Eigen::Matrix<std::complex<double>,2*xDim,2*xDim> eigVec = ces.eigenvectors(); Eigen::Matrix<std::complex<double>,2*xDim,xDim> unstableEigVec; int ctr = 0; for (int i = 0; i < 2*xDim; i++) { if (eigVal(i).real()*eigVal(i).real() + eigVal(i).imag()*eigVal(i).imag() > 1) { unstableEigVec.col(ctr) = eigVec.col(i); ctr++; if (ctr > xDim) break; } } Eigen::Matrix<std::complex<double>,xDim,xDim> U21inv = unstableEigVec.block(0,0,xDim,xDim).inverse(); Eigen::Matrix<std::complex<double>,xDim,xDim> PP = unstableEigVec.block(xDim,0,xDim,xDim) * U21inv; for (int i = 0; i < xDim; i++) { for (int j = 0; j < xDim; j++) { P(i,j) = PP(i,j).real(); } } }
template<int N> void calculateGaussPdf(Eigen::Matrix<double,Eigen::Dynamic,N> X, Eigen::Matrix<double,1,N> mu, Eigen::Matrix<double,N,N> C, double* result){ Eigen::Matrix<double,N,N> L = C.llt().matrixL().transpose(); // cholesky decomposition Eigen::Matrix<double,N,N> Linv = L.inverse(); double det = L.diagonal().prod(); //determinant of L is equal to square rooot of determinant of C double lognormconst = -log(2 * M_PI)*X.cols()/2 - log(fabs(det)); Eigen::Matrix<double,1,N> x = mu; Eigen::Matrix<double,1,N> tmp = x; for (int i=0; i<X.rows(); i++){ x.noalias() = X.row(i) - mu; tmp.noalias() = x*Linv; double exponent = -0.5 * (tmp.cwiseProduct(tmp)).sum(); result[i] = lognormconst+exponent; } /* Eigen::Matrix<double,Eigen::Dynamic,N> X0 = (X.rowwise() - mu)*Linv; Eigen::Map<Eigen::Matrix<double,Eigen::Dynamic,1> > resultMap(result, X.rows()); resultMap = (X0.rowwise().squaredNorm()).array() * (-0.5) + lognormconst; */ fmath::expd_v(result, X.rows()); }
void LipmVarHeightPlanner::backwardPass(const Traj<3,3> &com) { Eigen::Matrix<double,6,6> Qxx; Eigen::Matrix<double,3,3> Quu; Eigen::Matrix<double,3,3> invQuu; Eigen::Matrix<double,3,6> Qux; Eigen::Matrix<double,6,3> Qxu; Eigen::Matrix<double,6,1> Qx; Eigen::Matrix<double,3,1> Qu; Eigen::Matrix<double,6,6> A; Eigen::Matrix<double,6,3> B; Eigen::Matrix<double,6,6> Vxx = _Vxx; Eigen::Matrix<double,6,1> Vx = Eigen::Matrix<double,6,1>::Zero(); Eigen::Matrix<double,6,1> cur_x; Eigen::Matrix<double,3,1> cur_u; Eigen::Matrix<double,6,1> x_h; Eigen::Matrix<double,3,1> u_h; //NEW_GSL_MATRIX(tmpXX0, tmpXX0_d, tmpXX0_v, 6, 6); //NEW_GSL_MATRIX(tmpUX0, tmpUX0_d, tmpUX0_v, 3, 6); const double *x0, *u0; for (int t = (int)_K.size()-1; t >= 0; t--) { //printf("======================================\n"); // get x u x0 = _traj0[t].x; u0 = _traj0[t].u; //printf("%g %g %g %g %g %g\n", x0[0], x0[1], x0[2], x0[3], x0[4], x0[5]); //printf("%g %g %g\n", u0[0], u0[1], u0[2]); for (int i = 0; i < 6; i++) { cur_x(i) = x0[i]; x_h(i) = x0[i] - _zmp_d[t].x[i]; } for (int i = 0; i < 3; i++) { cur_u(i) = u0[i]; u_h(i) = u0[i] - _zmp_d[t].u[i]; } // get A B getAB(x0, u0, _z0[t], A, B); //printGSLMatrix("A", A); //printGSLMatrix("B", B); // Qx = Q*x_hat + A'*Vx Qx = _Q*x_h + A.transpose()*Vx; // Qu = R*u_hat + B'*Vx Qu = _R*u_h + B.transpose()*Vx; // Qxx = Q + A'*Vxx*A Qxx = _Q + A.transpose()*Vxx*A; // Qxu = A'*Vxx*B Qxu = A.transpose()*Vxx*B; // Quu = R + B'*Vxx*B Quu = _R + B.transpose()*Vxx*B; // Qux = Qxu' Qux = B.transpose()*Vxx*A; // K = -inv(Quu)*Qux _K[t] = -(Quu.llt().solve(Qux)); // du = -inv(Quu)*Qu _du[t] = -(Quu.llt().solve(Qu)); // Vx = Qx + K'*Qu Vx = Qx + _K[t].transpose()*Qu; // Vxx = Qxx + Qxu*K Vxx = Qxx + Qxu*_K[t]; //getchar(); //assert(!hasInfNan(_K[t])); //assert(!hasInfNan(_du[t])); } }
bool pcl::gpu::KinfuTracker::operator() (const DepthMap& depth_raw, Eigen::Affine3f *hint) { device::Intr intr (fx_, fy_, cx_, cy_); if (!disable_icp_) { { //ScopeTime time(">>> Bilateral, pyr-down-all, create-maps-all"); //depth_raw.copyTo(depths_curr[0]); device::bilateralFilter (depth_raw, depths_curr_[0]); if (max_icp_distance_ > 0) device::truncateDepth(depths_curr_[0], max_icp_distance_); for (int i = 1; i < LEVELS; ++i) device::pyrDown (depths_curr_[i-1], depths_curr_[i]); for (int i = 0; i < LEVELS; ++i) { device::createVMap (intr(i), depths_curr_[i], vmaps_curr_[i]); //device::createNMap(vmaps_curr_[i], nmaps_curr_[i]); computeNormalsEigen (vmaps_curr_[i], nmaps_curr_[i]); } pcl::device::sync (); } //can't perform more on first frame if (global_time_ == 0) { Matrix3frm init_Rcam = rmats_[0]; // [Ri|ti] - pos of camera, i.e. Vector3f init_tcam = tvecs_[0]; // transform from camera to global coo space for (i-1)th camera pose Mat33& device_Rcam = device_cast<Mat33> (init_Rcam); float3& device_tcam = device_cast<float3>(init_tcam); Matrix3frm init_Rcam_inv = init_Rcam.inverse (); Mat33& device_Rcam_inv = device_cast<Mat33> (init_Rcam_inv); float3 device_volume_size = device_cast<const float3>(tsdf_volume_->getSize()); //integrateTsdfVolume(depth_raw, intr, device_volume_size, device_Rcam_inv, device_tcam, tranc_dist, volume_); device::integrateTsdfVolume(depth_raw, intr, device_volume_size, device_Rcam_inv, device_tcam, tsdf_volume_->getTsdfTruncDist(), tsdf_volume_->data(), depthRawScaled_); for (int i = 0; i < LEVELS; ++i) device::tranformMaps (vmaps_curr_[i], nmaps_curr_[i], device_Rcam, device_tcam, vmaps_g_prev_[i], nmaps_g_prev_[i]); ++global_time_; return (false); } /////////////////////////////////////////////////////////////////////////////////////////// // Iterative Closest Point Matrix3frm Rprev = rmats_[global_time_ - 1]; // [Ri|ti] - pos of camera, i.e. Vector3f tprev = tvecs_[global_time_ - 1]; // tranfrom from camera to global coo space for (i-1)th camera pose Matrix3frm Rprev_inv = Rprev.inverse (); //Rprev.t(); //Mat33& device_Rprev = device_cast<Mat33> (Rprev); Mat33& device_Rprev_inv = device_cast<Mat33> (Rprev_inv); float3& device_tprev = device_cast<float3> (tprev); Matrix3frm Rcurr; Vector3f tcurr; if(hint) { Rcurr = hint->rotation().matrix(); tcurr = hint->translation().matrix(); } else { Rcurr = Rprev; // transform to global coo for ith camera pose tcurr = tprev; } { //ScopeTime time("icp-all"); for (int level_index = LEVELS-1; level_index>=0; --level_index) { int iter_num = icp_iterations_[level_index]; MapArr& vmap_curr = vmaps_curr_[level_index]; MapArr& nmap_curr = nmaps_curr_[level_index]; //MapArr& vmap_g_curr = vmaps_g_curr_[level_index]; //MapArr& nmap_g_curr = nmaps_g_curr_[level_index]; MapArr& vmap_g_prev = vmaps_g_prev_[level_index]; MapArr& nmap_g_prev = nmaps_g_prev_[level_index]; //CorespMap& coresp = coresps_[level_index]; for (int iter = 0; iter < iter_num; ++iter) { Mat33& device_Rcurr = device_cast<Mat33> (Rcurr); float3& device_tcurr = device_cast<float3>(tcurr); Eigen::Matrix<double, 6, 6, Eigen::RowMajor> A; Eigen::Matrix<double, 6, 1> b; #if 0 device::tranformMaps(vmap_curr, nmap_curr, device_Rcurr, device_tcurr, vmap_g_curr, nmap_g_curr); findCoresp(vmap_g_curr, nmap_g_curr, device_Rprev_inv, device_tprev, intr(level_index), vmap_g_prev, nmap_g_prev, distThres_, angleThres_, coresp); device::estimateTransform(vmap_g_prev, nmap_g_prev, vmap_g_curr, coresp, gbuf_, sumbuf_, A.data(), b.data()); //cv::gpu::GpuMat ma(coresp.rows(), coresp.cols(), CV_32S, coresp.ptr(), coresp.step()); //cv::Mat cpu; //ma.download(cpu); //cv::imshow(names[level_index] + string(" --- coresp white == -1"), cpu == -1); #else estimateCombined (device_Rcurr, device_tcurr, vmap_curr, nmap_curr, device_Rprev_inv, device_tprev, intr (level_index), vmap_g_prev, nmap_g_prev, distThres_, angleThres_, gbuf_, sumbuf_, A.data (), b.data ()); #endif //checking nullspace double det = A.determinant (); if (fabs (det) < 1e-15 || pcl_isnan (det)) { if (pcl_isnan (det)) cout << "qnan" << endl; reset (); return (false); } //float maxc = A.maxCoeff(); Eigen::Matrix<float, 6, 1> result = A.llt ().solve (b).cast<float>(); //Eigen::Matrix<float, 6, 1> result = A.jacobiSvd(ComputeThinU | ComputeThinV).solve(b); float alpha = result (0); float beta = result (1); float gamma = result (2); Eigen::Matrix3f Rinc = (Eigen::Matrix3f)AngleAxisf (gamma, Vector3f::UnitZ ()) * AngleAxisf (beta, Vector3f::UnitY ()) * AngleAxisf (alpha, Vector3f::UnitX ()); Vector3f tinc = result.tail<3> (); //compose tcurr = Rinc * tcurr + tinc; Rcurr = Rinc * Rcurr; } } } //save transform rmats_.push_back (Rcurr); tvecs_.push_back (tcurr); } else /* if (disable_icp_) */ { if (global_time_ == 0) ++global_time_; Matrix3frm Rcurr = rmats_[global_time_ - 1]; Vector3f tcurr = tvecs_[global_time_ - 1]; rmats_.push_back (Rcurr); tvecs_.push_back (tcurr); } Matrix3frm Rprev = rmats_[global_time_ - 1]; Vector3f tprev = tvecs_[global_time_ - 1]; Matrix3frm Rcurr = rmats_.back(); Vector3f tcurr = tvecs_.back(); /////////////////////////////////////////////////////////////////////////////////////////// // Integration check - We do not integrate volume if camera does not move. float rnorm = rodrigues2(Rcurr.inverse() * Rprev).norm(); float tnorm = (tcurr - tprev).norm(); const float alpha = 1.f; bool integrate = (rnorm + alpha * tnorm)/2 >= integration_metric_threshold_; if (disable_icp_) integrate = true; /////////////////////////////////////////////////////////////////////////////////////////// // Volume integration float3 device_volume_size = device_cast<const float3> (tsdf_volume_->getSize()); Matrix3frm Rcurr_inv = Rcurr.inverse (); Mat33& device_Rcurr_inv = device_cast<Mat33> (Rcurr_inv); float3& device_tcurr = device_cast<float3> (tcurr); if (integrate) { //ScopeTime time("tsdf"); //integrateTsdfVolume(depth_raw, intr, device_volume_size, device_Rcurr_inv, device_tcurr, tranc_dist, volume_); integrateTsdfVolume (depth_raw, intr, device_volume_size, device_Rcurr_inv, device_tcurr, tsdf_volume_->getTsdfTruncDist(), tsdf_volume_->data(), depthRawScaled_); } /////////////////////////////////////////////////////////////////////////////////////////// // Ray casting Mat33& device_Rcurr = device_cast<Mat33> (Rcurr); { //ScopeTime time("ray-cast-all"); raycast (intr, device_Rcurr, device_tcurr, tsdf_volume_->getTsdfTruncDist(), device_volume_size, tsdf_volume_->data(), vmaps_g_prev_[0], nmaps_g_prev_[0]); for (int i = 1; i < LEVELS; ++i) { resizeVMap (vmaps_g_prev_[i-1], vmaps_g_prev_[i]); resizeNMap (nmaps_g_prev_[i-1], nmaps_g_prev_[i]); } pcl::device::sync (); } ++global_time_; return (true); }
bool pcl::gpu::KinfuTracker::operator() (const DepthMap& depth_raw) { device::Intr intr (fx_, fy_, cx_, cy_); { //ScopeTime time(">>> Bilateral, pyr-down-all, create-maps-all"); //depth_raw.copyTo(depths_curr[0]); device::bilateralFilter (depth_raw, depths_curr_[0]); if (max_icp_distance_ > 0) device::truncateDepth(depths_curr_[0], max_icp_distance_); for (int i = 1; i < LEVELS; ++i) device::pyrDown (depths_curr_[i-1], depths_curr_[i]); for (int i = 0; i < LEVELS; ++i) { device::createVMap (intr(i), depths_curr_[i], vmaps_curr_[i]); //device::createNMap(vmaps_curr_[i], nmaps_curr_[i]); computeNormalsEigen (vmaps_curr_[i], nmaps_curr_[i]); } pcl::device::sync (); } //can't perform more on first frame if (global_time_ == 0) { Matrix3frm initial_cam_rot = rmats_[0]; // [Ri|ti] - pos of camera, i.e. Matrix3frm initial_cam_rot_inv = initial_cam_rot.inverse (); Vector3f initial_cam_trans = tvecs_[0]; // transform from camera to global coo space for (i-1)th camera pose Mat33& device_initial_cam_rot = device_cast<Mat33> (initial_cam_rot); Mat33& device_initial_cam_rot_inv = device_cast<Mat33> (initial_cam_rot_inv); float3& device_initial_cam_trans = device_cast<float3>(initial_cam_trans); float3 device_volume_size = device_cast<const float3>(tsdf_volume_->getSize()); device::integrateTsdfVolume(depth_raw, intr, device_volume_size, device_initial_cam_rot_inv, device_initial_cam_trans, tsdf_volume_->getTsdfTruncDist(), tsdf_volume_->data(), getCyclicalBufferStructure (), depthRawScaled_); /* Matrix3frm init_Rcam = rmats_[0]; // [Ri|ti] - pos of camera, i.e. Vector3f init_tcam = tvecs_[0]; // transform from camera to global coo space for (i-1)th camera pose Mat33& device_Rcam = device_cast<Mat33> (init_Rcam); float3& device_tcam = device_cast<float3>(init_tcam); Matrix3frm init_Rcam_inv = init_Rcam.inverse (); Mat33& device_Rcam_inv = device_cast<Mat33> (init_Rcam_inv); float3 device_volume_size = device_cast<const float3>(tsdf_volume_->getSize ()); //integrateTsdfVolume(depth_raw, intr, device_volume_size, device_Rcam_inv, device_tcam, tranc_dist, volume_); device::integrateTsdfVolume(depth_raw, intr, device_volume_size, device_Rcam_inv, device_tcam, tsdf_volume_->getTsdfTruncDist (), tsdf_volume_->data (), getCyclicalBufferStructure (), depthRawScaled_); */ for (int i = 0; i < LEVELS; ++i) device::tranformMaps (vmaps_curr_[i], nmaps_curr_[i], device_initial_cam_rot, device_initial_cam_trans, vmaps_g_prev_[i], nmaps_g_prev_[i]); if(perform_last_scan_) finished_ = true; ++global_time_; return (false); } /////////////////////////////////////////////////////////////////////////////////////////// // Iterative Closest Point // GET PREVIOUS GLOBAL TRANSFORM // Previous global rotation Matrix3frm cam_rot_global_prev = rmats_[global_time_ - 1]; // [Ri|ti] - pos of camera, i.e. // Previous global translation Vector3f cam_trans_global_prev = tvecs_[global_time_ - 1]; // transform from camera to global coo space for (i-1)th camera pose // Previous global inverse rotation Matrix3frm cam_rot_global_prev_inv = cam_rot_global_prev.inverse (); // Rprev.t(); // GET CURRENT GLOBAL TRANSFORM Matrix3frm cam_rot_global_curr = cam_rot_global_prev; // transform to global coo for ith camera pose Vector3f cam_trans_global_curr = cam_trans_global_prev; // CONVERT TO DEVICE TYPES //LOCAL PREVIOUS TRANSFORM Mat33& device_cam_rot_local_prev_inv = device_cast<Mat33> (cam_rot_global_prev_inv); float3& device_cam_trans_local_prev_tmp = device_cast<float3> (cam_trans_global_prev); float3 device_cam_trans_local_prev; device_cam_trans_local_prev.x = device_cam_trans_local_prev_tmp.x - (getCyclicalBufferStructure ())->origin_metric.x; device_cam_trans_local_prev.y = device_cam_trans_local_prev_tmp.y - (getCyclicalBufferStructure ())->origin_metric.y; device_cam_trans_local_prev.z = device_cam_trans_local_prev_tmp.z - (getCyclicalBufferStructure ())->origin_metric.z; /* Matrix3frm Rprev = rmats_[global_time_ - 1]; // [Ri|ti] - pos of camera, i.e. Vector3f tprev = tvecs_[global_time_ - 1]; // tranfrom from camera to global coo space for (i-1)th camera pose Matrix3frm Rprev_inv = Rprev.inverse (); //Rprev.t(); //Mat33& device_Rprev = device_cast<Mat33> (Rprev); Mat33& device_Rprev_inv = device_cast<Mat33> (Rprev_inv); float3& device_tprev = device_cast<float3> (tprev); Matrix3frm Rcurr = Rprev; // tranform to global coo for ith camera pose Vector3f tcurr = tprev; */ { //ScopeTime time("icp-all"); for (int level_index = LEVELS-1; level_index>=0; --level_index) { int iter_num = icp_iterations_[level_index]; // current maps MapArr& vmap_curr = vmaps_curr_[level_index]; MapArr& nmap_curr = nmaps_curr_[level_index]; // previous maps MapArr& vmap_g_prev = vmaps_g_prev_[level_index]; MapArr& nmap_g_prev = nmaps_g_prev_[level_index]; // We need to transform the maps from global to the local coordinates Mat33& rotation_id = device_cast<Mat33> (rmats_[0]); // Identity Rotation Matrix. Because we only need translation float3 cube_origin = (getCyclicalBufferStructure ())->origin_metric; cube_origin.x = -cube_origin.x; cube_origin.y = -cube_origin.y; cube_origin.z = -cube_origin.z; MapArr& vmap_temp = vmap_g_prev; MapArr& nmap_temp = nmap_g_prev; device::tranformMaps (vmap_temp, nmap_temp, rotation_id, cube_origin, vmap_g_prev, nmap_g_prev); /* MapArr& vmap_curr = vmaps_curr_[level_index]; MapArr& nmap_curr = nmaps_curr_[level_index]; //MapArr& vmap_g_curr = vmaps_g_curr_[level_index]; //MapArr& nmap_g_curr = nmaps_g_curr_[level_index]; MapArr& vmap_g_prev = vmaps_g_prev_[level_index]; MapArr& nmap_g_prev = nmaps_g_prev_[level_index]; */ //CorespMap& coresp = coresps_[level_index]; for (int iter = 0; iter < iter_num; ++iter) { /* Mat33& device_Rcurr = device_cast<Mat33> (Rcurr); float3& device_tcurr = device_cast<float3>(tcurr); */ //CONVERT TO DEVICE TYPES // CURRENT LOCAL TRANSFORM Mat33& device_cam_rot_local_curr = device_cast<Mat33> (cam_rot_global_curr);/// We have not dealt with changes in rotations float3& device_cam_trans_local_curr_tmp = device_cast<float3> (cam_trans_global_curr); float3 device_cam_trans_local_curr; device_cam_trans_local_curr.x = device_cam_trans_local_curr_tmp.x - (getCyclicalBufferStructure ())->origin_metric.x; device_cam_trans_local_curr.y = device_cam_trans_local_curr_tmp.y - (getCyclicalBufferStructure ())->origin_metric.y; device_cam_trans_local_curr.z = device_cam_trans_local_curr_tmp.z - (getCyclicalBufferStructure ())->origin_metric.z; Eigen::Matrix<double, 6, 6, Eigen::RowMajor> A; Eigen::Matrix<double, 6, 1> b; estimateCombined (device_cam_rot_local_curr, device_cam_trans_local_curr, vmap_curr, nmap_curr, device_cam_rot_local_prev_inv, device_cam_trans_local_prev, intr (level_index), vmap_g_prev, nmap_g_prev, distThres_, angleThres_, gbuf_, sumbuf_, A.data (), b.data ()); /* estimateCombined (device_Rcurr, device_tcurr, vmap_curr, nmap_curr, device_Rprev_inv, device_tprev, intr (level_index), vmap_g_prev, nmap_g_prev, distThres_, angleThres_, gbuf_, sumbuf_, A.data (), b.data ()); */ //checking nullspace double det = A.determinant (); if ( fabs (det) < 1e-15 || pcl_isnan (det) ) { if (pcl_isnan (det)) cout << "qnan" << endl; PCL_ERROR ("LOST...\n"); reset (); return (false); } //float maxc = A.maxCoeff(); Eigen::Matrix<float, 6, 1> result = A.llt ().solve (b).cast<float>(); //Eigen::Matrix<float, 6, 1> result = A.jacobiSvd(ComputeThinU | ComputeThinV).solve(b); float alpha = result (0); float beta = result (1); float gamma = result (2); Eigen::Matrix3f cam_rot_incremental = (Eigen::Matrix3f)AngleAxisf (gamma, Vector3f::UnitZ ()) * AngleAxisf (beta, Vector3f::UnitY ()) * AngleAxisf (alpha, Vector3f::UnitX ()); Vector3f cam_trans_incremental = result.tail<3> (); //compose cam_trans_global_curr = cam_rot_incremental * cam_trans_global_curr + cam_trans_incremental; cam_rot_global_curr = cam_rot_incremental * cam_rot_global_curr; /* tcurr = Rinc * tcurr + tinc; Rcurr = Rinc * Rcurr; */ } } } //save tranform rmats_.push_back (cam_rot_global_curr); tvecs_.push_back (cam_trans_global_curr); /* rmats_.push_back (Rcurr); tvecs_.push_back (tcurr); */ //check for shift bool has_shifted = cyclical_.checkForShift(tsdf_volume_, getCameraPose (), 0.6 * volume_size_, true, perform_last_scan_); if(has_shifted) PCL_WARN ("SHIFTING\n"); // get NEW local rotation Matrix3frm cam_rot_local_curr_inv = cam_rot_global_curr.inverse (); Mat33& device_cam_rot_local_curr_inv = device_cast<Mat33> (cam_rot_local_curr_inv); Mat33& device_cam_rot_local_curr = device_cast<Mat33> (cam_rot_global_curr); // get NEW local translation float3& device_cam_trans_local_curr_tmp = device_cast<float3> (cam_trans_global_curr); float3 device_cam_trans_local_curr; device_cam_trans_local_curr.x = device_cam_trans_local_curr_tmp.x - (getCyclicalBufferStructure ())->origin_metric.x; device_cam_trans_local_curr.y = device_cam_trans_local_curr_tmp.y - (getCyclicalBufferStructure ())->origin_metric.y; device_cam_trans_local_curr.z = device_cam_trans_local_curr_tmp.z - (getCyclicalBufferStructure ())->origin_metric.z; /////////////////////////////////////////////////////////////////////////////////////////// // Integration check - We do not integrate volume if camera does not move. float rnorm = rodrigues2(cam_rot_global_curr.inverse() * cam_rot_global_prev).norm(); float tnorm = (cam_trans_global_curr - cam_trans_global_prev).norm(); const float alpha = 1.f; bool integrate = (rnorm + alpha * tnorm)/2 >= integration_metric_threshold_; //~ if(integrate) //~ std::cout << "\tCamera movement since previous frame was " << (rnorm + alpha * tnorm)/2 << " integrate is set to " << integrate << std::endl; //~ else //~ std::cout << "Camera movement since previous frame was " << (rnorm + alpha * tnorm)/2 << " integrate is set to " << integrate << std::endl; /////////////////////////////////////////////////////////////////////////////////////////// // Volume integration float3 device_volume_size = device_cast<const float3> (tsdf_volume_->getSize()); /* Matrix3frm Rcurr_inv = Rcurr.inverse (); Mat33& device_Rcurr_inv = device_cast<Mat33> (Rcurr_inv); float3& device_tcurr = device_cast<float3> (tcurr);*/ if (integrate) { //integrateTsdfVolume(depth_raw, intr, device_volume_size, device_Rcurr_inv, device_tcurr, tranc_dist, volume_); integrateTsdfVolume (depth_raw, intr, device_volume_size, device_cam_rot_local_curr_inv, device_cam_trans_local_curr, tsdf_volume_->getTsdfTruncDist (), tsdf_volume_->data (), getCyclicalBufferStructure (), depthRawScaled_); } /////////////////////////////////////////////////////////////////////////////////////////// // Ray casting /*Mat33& device_Rcurr = device_cast<Mat33> (Rcurr);*/ { raycast (intr, device_cam_rot_local_curr, device_cam_trans_local_curr, tsdf_volume_->getTsdfTruncDist (), device_volume_size, tsdf_volume_->data (), getCyclicalBufferStructure (), vmaps_g_prev_[0], nmaps_g_prev_[0]); // POST-PROCESSING: We need to transform the newly raycasted maps into the global space. Mat33& rotation_id = device_cast<Mat33> (rmats_[0]); /// Identity Rotation Matrix. Because we only need translation float3 cube_origin = (getCyclicalBufferStructure ())->origin_metric; //~ PCL_INFO ("Raycasting with cube origin at %f, %f, %f\n", cube_origin.x, cube_origin.y, cube_origin.z); MapArr& vmap_temp = vmaps_g_prev_[0]; MapArr& nmap_temp = nmaps_g_prev_[0]; device::tranformMaps (vmap_temp, nmap_temp, rotation_id, cube_origin, vmaps_g_prev_[0], nmaps_g_prev_[0]); for (int i = 1; i < LEVELS; ++i) { resizeVMap (vmaps_g_prev_[i-1], vmaps_g_prev_[i]); resizeNMap (nmaps_g_prev_[i-1], nmaps_g_prev_[i]); } pcl::device::sync (); } if(has_shifted && perform_last_scan_) extractAndMeshWorld (); ++global_time_; return (true); }
bool MyPointCloud::alignPointClouds(std::vector<Matrix3frm>& Rcam, std::vector<Vector3f>& tcam, MyPointCloud *globalPreviousPointCloud, device::Intr& intrinsics, int globalTime) { Matrix3frm Rprev = Rcam[globalTime - 1]; // [Ri|ti] - pos of camera, i.e. Vector3f tprev = tcam[globalTime - 1]; // tranfrom from camera to global coo space for (i-1)th camera pose Matrix3frm Rprev_inv = Rprev.inverse(); //Rprev.t(); device::Mat33& device_Rprev_inv = device_cast<device::Mat33> (Rprev_inv); float3& device_tprev = device_cast<float3> (tprev); Matrix3frm Rcurr = Rprev; // tranform to global coo for ith camera pose Vector3f tcurr = tprev; #pragma omp parallel for for(int level = LEVELS - 1; level >= 0; --level) { int iterations = icpIterations_[level]; #pragma omp parallel for for(int iteration = 0; iteration < iterations; ++iteration) { { printf(" ICP level %d iteration %d", level, iteration); pcl::ScopeTime time(""); device::Mat33& device_Rcurr = device_cast<device::Mat33> (Rcurr); float3& device_tcurr = device_cast<float3>(tcurr); Eigen::Matrix<float, 6, 6, Eigen::RowMajor> A; Eigen::Matrix<float, 6, 1> b; if(level == 2 && iteration == 0) error_.create(rows_ * 4, cols_); device::estimateCombined (device_Rcurr, device_tcurr, vmaps_[level], nmaps_[level], device_Rprev_inv, device_tprev, intrinsics (level), globalPreviousPointCloud->getVertexMaps()[level], globalPreviousPointCloud->getNormalMaps()[level], distThres_, angleThres_, gbuf_, sumbuf_, A.data (), b.data (), error_); //checking nullspace float det = A.determinant (); if (fabs (det) < 1e-15 || !pcl::device::valid_host (det)) { // printf("ICP failed at level %d, iteration %d (global time %d)\n", level, iteration, globalTime); return (false); } //else printf("ICP succeed at level %d, iteration %d (global time %d)\n", level, iteration, globalTime); Eigen::Matrix<float, 6, 1> result = A.llt ().solve (b); //Eigen::Matrix<float, 6, 1> result = A.jacobiSvd(ComputeThinU | ComputeThinV).solve(b); float alpha = result (0); float beta = result (1); float gamma = result (2); Eigen::Matrix3f Rinc = (Eigen::Matrix3f)AngleAxisf (gamma, Vector3f::UnitZ ()) * AngleAxisf (beta, Vector3f::UnitY ()) * AngleAxisf (alpha, Vector3f::UnitX ()); Vector3f tinc = result.tail<3> (); //compose tcurr = Rinc * tcurr + tinc; Rcurr = Rinc * Rcurr; } } } //save tranform Rcam[globalTime] = Rcurr; tcam[globalTime] = tcurr; return (true); }
inline bool pcl::gpu::kinfuLS::KinfuTracker::performPairWiseICP(const Intr cam_intrinsics, Matrix3frm& resulting_rotation , Vector3f& resulting_translation) { // we assume that both v and n maps are in the same coordinate space // initialize rotation and translation to respectively identity and zero Matrix3frm previous_rotation = Eigen::Matrix3f::Identity (); Matrix3frm previous_rotation_inv = previous_rotation.inverse (); Vector3f previous_translation = Vector3f(0,0,0); /////////////////////////////////////////////// // Convert pose to device type Mat33 device_cam_rot_prev_inv; float3 device_cam_trans_prev; convertTransforms(previous_rotation_inv, previous_translation, device_cam_rot_prev_inv, device_cam_trans_prev); // Initialize output pose to current pose (i.e. identity and zero translation) Matrix3frm current_rotation = previous_rotation; Vector3f current_translation = previous_translation; /////////////////////////////////////////////// // Run ICP { //ScopeTime time("icp-all"); for (int level_index = LEVELS-1; level_index>=0; --level_index) { int iter_num = icp_iterations_[level_index]; // current vertex and normal maps MapArr& vmap_curr = vmaps_curr_[level_index]; MapArr& nmap_curr = nmaps_curr_[level_index]; // previous vertex and normal maps MapArr& vmap_prev = vmaps_prev_[level_index]; MapArr& nmap_prev = nmaps_prev_[level_index]; // no need to transform maps from global to local since they are both in camera coordinates // run ICP for iter_num iterations (return false when lost) for (int iter = 0; iter < iter_num; ++iter) { //CONVERT POSES TO DEVICE TYPES // CURRENT LOCAL POSE Mat33 device_current_rotation = device_cast<Mat33> (current_rotation); float3 device_current_translation_local = device_cast<float3> (current_translation); Eigen::Matrix<double, 6, 6, Eigen::RowMajor> A; Eigen::Matrix<double, 6, 1> b; // call the ICP function (see paper by Kok-Lim Low "Linear Least-squares Optimization for Point-to-Plane ICP Surface Registration") estimateCombined (device_current_rotation, device_current_translation_local, vmap_curr, nmap_curr, device_cam_rot_prev_inv, device_cam_trans_prev, cam_intrinsics (level_index), vmap_prev, nmap_prev, distThres_, angleThres_, gbuf_, sumbuf_, A.data (), b.data ()); // checking nullspace double det = A.determinant (); if ( fabs (det) < 1e-15 || pcl_isnan (det) ) { if (pcl_isnan (det)) cout << "qnan" << endl; PCL_WARN ("ICP PairWise LOST...\n"); //reset (); return (false); } Eigen::Matrix<float, 6, 1> result = A.llt ().solve (b).cast<float>(); float alpha = result (0); float beta = result (1); float gamma = result (2); // deduce incremental rotation and translation from ICP's results Eigen::Matrix3f cam_rot_incremental = (Eigen::Matrix3f)AngleAxisf (gamma, Vector3f::UnitZ ()) * AngleAxisf (beta, Vector3f::UnitY ()) * AngleAxisf (alpha, Vector3f::UnitX ()); Vector3f cam_trans_incremental = result.tail<3> (); //compose global pose current_translation = cam_rot_incremental * current_translation + cam_trans_incremental; current_rotation = cam_rot_incremental * current_rotation; } } } //////////////////////////////////////////////////////////////////////////////////////////////////////////////////// // since raw depthmaps are quite noisy, we make sure the estimated transform is big enought to be taken into account float rnorm = rodrigues2(current_rotation).norm(); float tnorm = (current_translation).norm(); const float alpha = 1.f; bool integrate = (rnorm + alpha * tnorm)/2 >= integration_metric_threshold_ * 2.0f; if(integrate) { resulting_rotation = current_rotation; resulting_translation = current_translation; } else { resulting_rotation = Eigen::Matrix3f::Identity (); resulting_translation = Vector3f(0,0,0); } // ICP has converged return (true); }
inline bool pcl::gpu::kinfuLS::KinfuTracker::performICP(const Intr& cam_intrinsics, Matrix3frm& previous_global_rotation, Vector3f& previous_global_translation, Matrix3frm& current_global_rotation , Vector3f& current_global_translation) { if(disable_icp_) { lost_=false; return (true); } // Compute inverse rotation Matrix3frm previous_global_rotation_inv = previous_global_rotation.inverse (); // Rprev.t(); /////////////////////////////////////////////// // Convert pose to device type Mat33 device_cam_rot_local_prev_inv; float3 device_cam_trans_local_prev; convertTransforms(previous_global_rotation_inv, previous_global_translation, device_cam_rot_local_prev_inv, device_cam_trans_local_prev); device_cam_trans_local_prev -= getCyclicalBufferStructure ()->origin_metric; ; // Use temporary pose, so that we modify the current global pose only if ICP converged Matrix3frm resulting_rotation; Vector3f resulting_translation; // Initialize output pose to current pose current_global_rotation = previous_global_rotation; current_global_translation = previous_global_translation; /////////////////////////////////////////////// // Run ICP { //ScopeTime time("icp-all"); for (int level_index = LEVELS-1; level_index>=0; --level_index) { int iter_num = icp_iterations_[level_index]; // current vertex and normal maps MapArr& vmap_curr = vmaps_curr_[level_index]; MapArr& nmap_curr = nmaps_curr_[level_index]; // previous vertex and normal maps MapArr& vmap_g_prev = vmaps_g_prev_[level_index]; MapArr& nmap_g_prev = nmaps_g_prev_[level_index]; // We need to transform the maps from global to local coordinates Mat33& rotation_id = device_cast<Mat33> (rmats_[0]); // Identity Rotation Matrix. Because we only need translation float3 cube_origin = (getCyclicalBufferStructure ())->origin_metric; cube_origin = -cube_origin; MapArr& vmap_temp = vmap_g_prev; MapArr& nmap_temp = nmap_g_prev; transformMaps (vmap_temp, nmap_temp, rotation_id, cube_origin, vmap_g_prev, nmap_g_prev); // run ICP for iter_num iterations (return false when lost) for (int iter = 0; iter < iter_num; ++iter) { //CONVERT POSES TO DEVICE TYPES // CURRENT LOCAL POSE Mat33 device_current_rotation = device_cast<Mat33> (current_global_rotation); // We do not deal with changes in rotations float3 device_current_translation_local = device_cast<float3> (current_global_translation); device_current_translation_local -= getCyclicalBufferStructure ()->origin_metric; Eigen::Matrix<double, 6, 6, Eigen::RowMajor> A; Eigen::Matrix<double, 6, 1> b; // call the ICP function (see paper by Kok-Lim Low "Linear Least-squares Optimization for Point-to-Plane ICP Surface Registration") estimateCombined (device_current_rotation, device_current_translation_local, vmap_curr, nmap_curr, device_cam_rot_local_prev_inv, device_cam_trans_local_prev, cam_intrinsics (level_index), vmap_g_prev, nmap_g_prev, distThres_, angleThres_, gbuf_, sumbuf_, A.data (), b.data ()); // checking nullspace double det = A.determinant (); if ( fabs (det) < 1e-15 /*100000 */ || pcl_isnan (det) ) //TODO find a threshold that makes ICP track well, but prevents it from generating wrong transforms { if (pcl_isnan (det)) cout << "qnan" << endl; if(lost_ == false) PCL_ERROR ("ICP LOST... PLEASE COME BACK TO THE LAST VALID POSE (green)\n"); //reset (); //GUI will now show the user that ICP is lost. User needs to press "R" to reset the volume lost_ = true; return (false); } Eigen::Matrix<float, 6, 1> result = A.llt ().solve (b).cast<float>(); float alpha = result (0); float beta = result (1); float gamma = result (2); // deduce incremental rotation and translation from ICP's results Eigen::Matrix3f cam_rot_incremental = (Eigen::Matrix3f)AngleAxisf (gamma, Vector3f::UnitZ ()) * AngleAxisf (beta, Vector3f::UnitY ()) * AngleAxisf (alpha, Vector3f::UnitX ()); Vector3f cam_trans_incremental = result.tail<3> (); //compose global pose current_global_translation = cam_rot_incremental * current_global_translation + cam_trans_incremental; current_global_rotation = cam_rot_incremental * current_global_rotation; } } } // ICP has converged lost_ = false; return (true); }