视觉十四讲：第八讲_直接法

1.直接法的推导：

考虑某个空间点P和两个时刻的相机，P的世界坐标为[X,Y,Z],它在两个相机上成像，记非齐次像素坐标为(p_{1},p_{2})，目标是求第一个相机到第二个相机的相对位姿变换。

直接法的思路是根据当前相机的位姿估计值，来寻找(p_{2})的位置，误差项为光度误差
(e = I_{1}(p_{1}) - I_{2}(p_{2}))
考虑多个空间点P，那个整个相机位姿估计问题会变成：

主要问题是求误差e相对于位姿的导数关系。使用李代数的扰动模型，给(exp(xi))左乘一个小扰动(exp(delta xi))得：

记：(q = delta xi^{Lambda} exp(xi^{Lambda}) P),
(u = frac{1}{Z_{2}}Kq)
q为P在扰动之后，位于第二个相机坐标系下的坐标，而u为它的像素坐标，利用一阶泰勒展开，有：


故误差相对于李代数的雅可比矩阵为：

2.直接法分类：

1.稀疏关键点：通常使用数百个至上千个关键点，像LK光流那样，这种稀疏直接法不必计算描述子，因此速度最快，但只能计算稀疏的重构。
2.半稠密直接法：考虑图像内带有梯度的像素点，舍弃像素梯度不明显的地方，重构一个半稠密结构。
3.稠密直接法：考虑所有像素点，需要GPU加速。

3.实践直接法：

1.单层稀疏直接法：

class JacobianAccumulator {
public:
    JacobianAccumulator(
        const cv::Mat &img1_,
        const cv::Mat &img2_,
        const VecVector2d &px_ref_,
        const vector<double> depth_ref_,
        Sophus::SE3d &T21_) :
        img1(img1_), img2(img2_), px_ref(px_ref_), depth_ref(depth_ref_), T21(T21_) {
        projection = VecVector2d(px_ref.size(), Eigen::Vector2d(0, 0));
    }

    /// accumulate jacobians in a range
    void accumulate_jacobian(const cv::Range &range);

    /// get hessian matrix
    Matrix6d hessian() const { return H; }

    /// get bias
    Vector6d bias() const { return b; }

    /// get total cost
    double cost_func() const { return cost; }

    /// get projected points
    VecVector2d projected_points() const { return projection; }

    /// reset h, b, cost to zero
    void reset() {
        H = Matrix6d::Zero();
        b = Vector6d::Zero();
        cost = 0;
    }

private:
    const cv::Mat &img1;
    const cv::Mat &img2;
    const VecVector2d &px_ref;
    const vector<double> depth_ref;
    Sophus::SE3d &T21;
    VecVector2d projection; // projected points

    std::mutex hessian_mutex;
    Matrix6d H = Matrix6d::Zero();
    Vector6d b = Vector6d::Zero();
    double cost = 0;
};
void DirectPoseEstimationSingleLayer(
    const cv::Mat &img1,
    const cv::Mat &img2,
    const VecVector2d &px_ref,
    const vector<double> depth_ref,
    Sophus::SE3d &T21) {

    const int iterations = 10;
    double cost = 0, lastCost = 0;
    auto t1 = chrono::steady_clock::now();
    JacobianAccumulator jaco_accu(img1, img2, px_ref, depth_ref, T21);

    for (int iter = 0; iter < iterations; iter++) {  //10次迭代次数
        jaco_accu.reset();
        cv::parallel_for_(cv::Range(0, px_ref.size()),  //循环2000个点
                          std::bind(&JacobianAccumulator::accumulate_jacobian, &jaco_accu, std::placeholders::_1));
        
		Matrix6d H = jaco_accu.hessian();
        Vector6d b = jaco_accu.bias();

        //10次迭代获取更新值
        Vector6d update = H.ldlt().solve(b);
		//更新姿态
        T21 = Sophus::SE3d::exp(update) * T21;
		
        cost = jaco_accu.cost_func();

        if (std::isnan(update[0])) {
            // sometimes occurred when we have a black or white patch and H is irreversible
            cout << "update is nan" << endl;
            break;
        }
        if (iter > 0 && cost > lastCost) {
            cout << "cost increased: " << cost << ", " << lastCost << endl;
            break;
        }
        if (update.norm() < 1e-3) {
            // converge
            break;
        }

        lastCost = cost;
        cout << "iteration: " << iter << ", cost: " << cost << endl;
    }

    cout << "T21 = 
" << T21.matrix() << endl;
    auto t2 = chrono::steady_clock::now();
    auto time_used = chrono::duration_cast<chrono::duration<double>>(t2 - t1);
    cout << "direct method for single layer: " << time_used.count() << endl;

    // plot the projected pixels here
    cv::Mat img2_show;
    cv::cvtColor(img2, img2_show, CV_GRAY2BGR);
    VecVector2d projection = jaco_accu.projected_points();
    for (size_t i = 0; i < px_ref.size(); ++i) {
        auto p_ref = px_ref[i];
        auto p_cur = projection[i];
        if (p_cur[0] > 0 && p_cur[1] > 0) {
            cv::circle(img2_show, cv::Point2f(p_cur[0], p_cur[1]), 2, cv::Scalar(0, 250, 0), 2);
            cv::line(img2_show, cv::Point2f(p_ref[0], p_ref[1]), cv::Point2f(p_cur[0], p_cur[1]),
                     cv::Scalar(0, 250, 0));
        }
    }
    cv::imshow("current", img2_show);
    cv::waitKey();
}
void JacobianAccumulator::accumulate_jacobian(const cv::Range &range) {

    // parameters
    const int half_patch_size = 1; //区块大小
    int cnt_good = 0;
    Matrix6d hessian = Matrix6d::Zero();
    Vector6d bias = Vector6d::Zero();
    double cost_tmp = 0;

    for (size_t i = range.start; i < range.end; i++) {

        //第一张图的像素坐标转相机归一化坐标，乘深度转为相机3D点
        Eigen::Vector3d point_ref =
            depth_ref[i] * Eigen::Vector3d((px_ref[i][0] - cx) / fx, (px_ref[i][1] - cy) / fy, 1);
        //将第一张图的点根据估计的位姿，转到第二张图的坐标系下
		Eigen::Vector3d point_cur = T21 * point_ref;
        if (point_cur[2] < 0)   // depth invalid
            continue;
		//将相机坐标系转为像素坐标系
        float u = fx * point_cur[0] / point_cur[2] + cx, v = fy * point_cur[1] / point_cur[2] + cy;
        //去除边界点
		if (u < half_patch_size || u > img2.cols - half_patch_size || v < half_patch_size ||
            v > img2.rows - half_patch_size)
            continue;
        
		
        projection[i] = Eigen::Vector2d(u, v);
        double X = point_cur[0], Y = point_cur[1], Z = point_cur[2],
		//深度平方，用于雅可比矩阵的计算
            Z2 = Z * Z, Z_inv = 1.0 / Z, Z2_inv = Z_inv * Z_inv;
			
        cnt_good++;

        //2个for循环主要是是为了计算图像梯度
        for (int x = -half_patch_size; x <= half_patch_size; x++)
            for (int y = -half_patch_size; y <= half_patch_size; y++) {
				
				//重投影光度误差，
                double error = GetPixelValue(img1, px_ref[i][0] + x, px_ref[i][1] + y) -
                               GetPixelValue(img2, u + x, v + y);
                Matrix26d J_pixel_xi;
                Eigen::Vector2d J_img_pixel;
				//计算李代数雅可比矩阵
                J_pixel_xi(0, 0) = fx * Z_inv;
                J_pixel_xi(0, 1) = 0;
                J_pixel_xi(0, 2) = -fx * X * Z2_inv;
                J_pixel_xi(0, 3) = -fx * X * Y * Z2_inv;
                J_pixel_xi(0, 4) = fx + fx * X * X * Z2_inv;
                J_pixel_xi(0, 5) = -fx * Y * Z_inv;

                J_pixel_xi(1, 0) = 0;
                J_pixel_xi(1, 1) = fy * Z_inv;
                J_pixel_xi(1, 2) = -fy * Y * Z2_inv;
                J_pixel_xi(1, 3) = -fy - fy * Y * Y * Z2_inv;
                J_pixel_xi(1, 4) = fy * X * Y * Z2_inv;
                J_pixel_xi(1, 5) = fy * X * Z_inv;

				//图像梯度
                J_img_pixel = Eigen::Vector2d(
                    0.5 * (GetPixelValue(img2, u + 1 + x, v + y) - GetPixelValue(img2, u - 1 + x, v + y)),
                    0.5 * (GetPixelValue(img2, u + x, v + 1 + y) - GetPixelValue(img2, u + x, v - 1 + y))
                );

                // 计算误差对于李代数的雅可比矩阵
                Vector6d J = -1.0 * (J_img_pixel.transpose() * J_pixel_xi).transpose();

                hessian += J * J.transpose();
                bias += -error * J;
                cost_tmp += error * error;
            }
    }

    if (cnt_good) {
        // set hessian, bias and cost
        unique_lock<mutex> lck(hessian_mutex);   //多线程互斥量,自动释放
        H += hessian;
        b += bias;
        cost += cost_tmp / cnt_good;
    }
}

2.多层稀疏直接法：

类似于光流，把单层直接法扩展到金字塔式的多层直接法上，用Coarse-to-fine的过程计算相对运动。

void DirectPoseEstimationMultiLayer(
    const cv::Mat &img1,
    const cv::Mat &img2,
    const VecVector2d &px_ref,
    const vector<double> depth_ref,
    Sophus::SE3d &T21) {

    // parameters
    int pyramids = 4;
    double pyramid_scale = 0.5;
    double scales[] = {1.0, 0.5, 0.25, 0.125};

    // create pyramids
    vector<cv::Mat> pyr1, pyr2; // image pyramids
    for (int i = 0; i < pyramids; i++) {
        if (i == 0) {
            pyr1.push_back(img1);
            pyr2.push_back(img2);
        } else {
            cv::Mat img1_pyr, img2_pyr;
            cv::resize(pyr1[i - 1], img1_pyr,
                       cv::Size(pyr1[i - 1].cols * pyramid_scale, pyr1[i - 1].rows * pyramid_scale));
            cv::resize(pyr2[i - 1], img2_pyr,
                       cv::Size(pyr2[i - 1].cols * pyramid_scale, pyr2[i - 1].rows * pyramid_scale));
            pyr1.push_back(img1_pyr);
            pyr2.push_back(img2_pyr);
        }
    }

    double fxG = fx, fyG = fy, cxG = cx, cyG = cy;  // backup the old values
    for (int level = pyramids - 1; level >= 0; level--) {
        VecVector2d px_ref_pyr; // set the keypoints in this pyramid level
        for (auto &px: px_ref) {
            px_ref_pyr.push_back(scales[level] * px);
        }

        // scale fx, fy, cx, cy in different pyramid levels
        fx = fxG * scales[level];
        fy = fyG * scales[level];
        cx = cxG * scales[level];
        cy = cyG * scales[level];
        DirectPoseEstimationSingleLayer(pyr1[level], pyr2[level], px_ref_pyr, depth_ref, T21);
    }

}