function [omckk,Tckk,Rckk] = compute_extrinsic_init(x_kk,X_kk,fc,cc,kc,alpha_c), %compute_extrinsic % %[omckk,Tckk,Rckk] = compute_extrinsic_init(x_kk,X_kk,fc,cc,kc,alpha_c) % %Computes the extrinsic parameters attached to a 3D structure X_kk given its projection %on the image plane x_kk and the intrinsic camera parameters fc, cc and kc. %Works with planar and non-planar structures. % %INPUT: x_kk: Feature locations on the images % X_kk: Corresponding grid coordinates % fc: Camera focal length % cc: Principal point coordinates % kc: Distortion coefficients % alpha_c: Skew coefficient % %OUTPUT: omckk: 3D rotation vector attached to the grid positions in space % Tckk: 3D translation vector attached to the grid positions in space % Rckk: 3D rotation matrices corresponding to the omc vectors % %Method: Computes the normalized point coordinates, then computes the 3D pose % %Important functions called within that program: % %normalize_pixel: Computes the normalize image point coordinates. % %pose3D: Computes the 3D pose of the structure given the normalized image projection. % %project_points.m: Computes the 2D image projections of a set of 3D points if nargin < 6, alpha_c = 0; if nargin < 5, kc = zeros(5,1); if nargin < 4, cc = zeros(2,1); if nargin < 3, fc = ones(2,1); if nargin < 2, error('Need 2D projections and 3D points (in compute_extrinsic.m)'); return; end; end; end; end; end; %keyboard; % Compute the normalized coordinates: xn = normalize_pixel(x_kk,fc,cc,kc,alpha_c); Np = size(xn,2); %% Check for planarity of the structure: %keyboard; X_mean = mean(X_kk')'; Y = X_kk - (X_mean*ones(1,Np)); YY = Y*Y'; [U,S,V] = svd(YY); r = S(3,3)/S(2,2); %keyboard; if (r < 1e-3)|(Np < 5), %1e-3, %1e-4, %norm(X_kk(3,:)) < eps, % Test of planarity %fprintf(1,'Planar structure detected: r=%f\n',r); % Transform the plane to bring it in the Z=0 plane: R_transform = V'; %norm(R_transform(1:2,3)) if norm(R_transform(1:2,3)) < 1e-6, R_transform = eye(3); end; if det(R_transform) < 0, R_transform = -R_transform; end; T_transform = -(R_transform)*X_mean; X_new = R_transform*X_kk + T_transform*ones(1,Np); % Compute the planar homography: H = compute_homography(xn,X_new(1:2,:)); % De-embed the motion parameters from the homography: sc = mean([norm(H(:,1));norm(H(:,2))]); H = H/sc; % Extra normalization for some reasons... %H(:,1) = H(:,1)/norm(H(:,1)); %H(:,2) = H(:,2)/norm(H(:,2)); if 0, %%% Some tests for myself... the opposite sign solution leads to negative depth!!! % Case#1: no opposite sign: omckk1 = rodrigues([H(:,1:2) cross(H(:,1),H(:,2))]); Rckk1 = rodrigues(omckk1); Tckk1 = H(:,3); Hs1 = [Rckk1(:,1:2) Tckk1]; xn1 = Hs1*[X_new(1:2,:);ones(1,Np)]; xn1 = [xn1(1,:)./xn1(3,:) ; xn1(2,:)./xn1(3,:)]; e1 = xn1 - xn; % Case#2: opposite sign: omckk2 = rodrigues([-H(:,1:2) cross(H(:,1),H(:,2))]); Rckk2 = rodrigues(omckk2); Tckk2 = -H(:,3); Hs2 = [Rckk2(:,1:2) Tckk2]; xn2 = Hs2*[X_new(1:2,:);ones(1,Np)]; xn2 = [xn2(1,:)./xn2(3,:) ; xn2(2,:)./xn2(3,:)]; e2 = xn2 - xn; if 1, %norm(e1) < norm(e2), omckk = omckk1; Tckk = Tckk1; Rckk = Rckk1; else omckk = omckk2; Tckk = Tckk2; Rckk = Rckk2; end; else u1 = H(:,1); u1 = u1 / norm(u1); u2 = H(:,2) - dot(u1,H(:,2)) * u1; u2 = u2 / norm(u2); u3 = cross(u1,u2); RRR = [u1 u2 u3]; omckk = rodrigues(RRR); %omckk = rodrigues([H(:,1:2) cross(H(:,1),H(:,2))]); Rckk = rodrigues(omckk); Tckk = H(:,3); end; %If Xc = Rckk * X_new + Tckk, then Xc = Rckk * R_transform * X_kk + Tckk + T_transform Tckk = Tckk + Rckk* T_transform; Rckk = Rckk * R_transform; omckk = rodrigues(Rckk); Rckk = rodrigues(omckk); else %fprintf(1,'Non planar structure detected: r=%f\n',r); % Computes an initial guess for extrinsic parameters (works for general 3d structure, not planar!!!): % The DLT method is applied here!! J = zeros(2*Np,12); xX = (ones(3,1)*xn(1,:)).*X_kk; yX = (ones(3,1)*xn(2,:)).*X_kk; J(1:2:end,[1 4 7]) = -X_kk'; J(2:2:end,[2 5 8]) = X_kk'; J(1:2:end,[3 6 9]) = xX'; J(2:2:end,[3 6 9]) = -yX'; J(1:2:end,12) = xn(1,:)'; J(2:2:end,12) = -xn(2,:)'; J(1:2:end,10) = -ones(Np,1); J(2:2:end,11) = ones(Np,1); JJ = J'*J; [U,S,V] = svd(JJ); RR = reshape(V(1:9,12),3,3); if det(RR) < 0, V(:,12) = -V(:,12); RR = -RR; end; [Ur,Sr,Vr] = svd(RR); Rckk = Ur*Vr'; sc = norm(V(1:9,12)) / norm(Rckk(:)); Tckk = V(10:12,12)/sc; omckk = rodrigues(Rckk); Rckk = rodrigues(omckk); end;