vins-mono学习笔记

void FeatureManager::triangulate(Vector3d Ps[], Vector3d tic[], Matrix3d ric[])
{
    for (auto &it_per_id : feature)
    {
        it_per_id.used_num = it_per_id.feature_per_frame.size();
        if (!(it_per_id.used_num >= 2 && it_per_id.start_frame < WINDOW_SIZE - 2))
            continue;

        if (it_per_id.estimated_depth > 0)
            continue;
        int imu_i = it_per_id.start_frame, imu_j = imu_i - 1;

        ROS_ASSERT(NUM_OF_CAM == 1);
        Eigen::MatrixXd svd_A(2 * it_per_id.feature_per_frame.size(), 4);
        int svd_idx = 0;

        //R0 t0为第i帧相机坐标系到世界坐标系的变换矩阵
        Eigen::Matrix<double, 3, 4> P0;
        Eigen::Vector3d t0 = Ps[imu_i] + Rs[imu_i] * tic[0];
        Eigen::Matrix3d R0 = Rs[imu_i] * ric[0];
        P0.leftCols<3>() = Eigen::Matrix3d::Identity();
        P0.rightCols<1>() = Eigen::Vector3d::Zero();

        for (auto &it_per_frame : it_per_id.feature_per_frame)
        {
            imu_j++;
            //R t为第j帧相机坐标系到第i帧相机坐标系的变换矩阵，P为i到j的变换矩阵
            Eigen::Vector3d t1 = Ps[imu_j] + Rs[imu_j] * tic[0];
            Eigen::Matrix3d R1 = Rs[imu_j] * ric[0];
            Eigen::Vector3d t = R0.transpose() * (t1 - t0);
            Eigen::Matrix3d R = R0.transpose() * R1;
            Eigen::Matrix<double, 3, 4> P;
            P.leftCols<3>() = R.transpose();
            P.rightCols<1>() = -R.transpose() * t;
            Eigen::Vector3d f = it_per_frame.point.normalized();
            //P = [P1 P2 P3]^T 
            //AX=0      A = [A(2*i) A(2*i+1) A(2*i+2) A(2*i+3) ...]^T
            //A(2*i)   = x(i) * P3 - z(i) * P1
            //A(2*i+1) = y(i) * P3 - z(i) * P2
            svd_A.row(svd_idx++) = f[0] * P.row(2) - f[2] * P.row(0);
            svd_A.row(svd_idx++) = f[1] * P.row(2) - f[2] * P.row(1);

            if (imu_i == imu_j)
                continue;
        }
        //对A的SVD分解得到其最小奇异值对应的单位奇异向量(x,y,z,w)，深度为z/w
        ROS_ASSERT(svd_idx == svd_A.rows());
        Eigen::Vector4d svd_V = Eigen::JacobiSVD<Eigen::MatrixXd>(svd_A, Eigen::ComputeThinV).matrixV().rightCols<1>();
        double svd_method = svd_V[2] / svd_V[3];
        //it_per_id->estimated_depth = -b / A;
        //it_per_id->estimated_depth = svd_V[2] / svd_V[3];

        it_per_id.estimated_depth = svd_method;
        //it_per_id->estimated_depth = INIT_DEPTH;

        if (it_per_id.estimated_depth < 0.1)
        {
            it_per_id.estimated_depth = INIT_DEPTH;
        }

    }
}

         vins这里三角化求出的it_per_id.estimated_depth,是特征点在开始帧图像中的深度,特征点在开始帧相机坐标系下,特征点距离相机光心的距离.

//三角化两帧间某个对应特征点的深度
void GlobalSFM::triangulatePoint(Eigen::Matrix<double, 3, 4> &Pose0, Eigen::Matrix<double, 3, 4> &Pose1,
						Vector2d &point0, Vector2d &point1, Vector3d &point_3d)
{
	Matrix4d design_matrix = Matrix4d::Zero();
	design_matrix.row(0) = point0[0] * Pose0.row(2) - Pose0.row(0);
	design_matrix.row(1) = point0[1] * Pose0.row(2) - Pose0.row(1);
	design_matrix.row(2) = point1[0] * Pose1.row(2) - Pose1.row(0);
	design_matrix.row(3) = point1[1] * Pose1.row(2) - Pose1.row(1);
	Vector4d triangulated_point;
	triangulated_point =
		      design_matrix.jacobiSvd(Eigen::ComputeFullV).matrixV().rightCols<1>();
	point_3d(0) = triangulated_point(0) / triangulated_point(3);
	point_3d(1) = triangulated_point(1) / triangulated_point(3);
	point_3d(2) = triangulated_point(2) / triangulated_point(3);
}

     vins这个三角化应该是求解特征点在世界坐标下的坐标,利用两帧相机关于世界的变换和特征点在两帧相机下的归一化坐标,求解特征点世界坐标.注,程序中以l帧作为世界坐标系.