一文解读自动驾驶中的激光雷达点云分割算法

/*
    @brief Extract initial seeds of the given pointcloud sorted segment.
    This function filter ground seeds points accoring to heigt.
    This function will set the `g_ground_pc` to `g_seed_pc`.
    @param p_sorted: sorted pointcloud
    
    @param ::num_lpr_: num of LPR points
    @param ::th_seeds_: threshold distance of seeds
    @param ::
*/
void PlaneGroundFilter::extract_initial_seeds_(const pcl::PointCloud<VPoint> &p_sorted)
{
    // LPR is the mean of low point representative
    double sum = 0;
    int cnt = 0;
    // Calculate the mean height value.
    for (int i = 0; i < p_sorted.points.size() && cnt < num_lpr_; i++)
    {
        sum += p_sorted.points[i].z;
        cnt++;
    }
    double lpr_height = cnt != 0 ? sum / cnt : 0; // in case divide by 0
    g_seeds_pc->clear();
    // iterate pointcloud, filter those height is less than lpr.height+th_seeds_
    for (int i = 0; i < p_sorted.points.size(); i++)
    {
        if (p_sorted.points[i].z < lpr_height + th_seeds_)
        {
            g_seeds_pc->points.push_back(p_sorted.points[i]);
        }
    }
    // return seeds points
}

/*
    @brief The function to estimate plane model. The
    model parameter `normal_` and `d_`, and `th_dist_d_`
    is set here.
    The main step is performed SVD(UAV) on covariance matrix.
    Taking the sigular vector in U matrix according to the smallest
    sigular value in A, as the `normal_`. `d_` is then calculated 
    according to mean ground points.


    @param g_ground_pc:global ground pointcloud ptr.
    
*/
void PlaneGroundFilter::estimate_plane_(void)
{
    // Create covarian matrix in single pass.
    // TODO: compare the efficiency.
    Eigen::Matrix3f cov;
    Eigen::Vector4f pc_mean;
    pcl::computeMeanAndCovarianceMatrix(*g_ground_pc, cov, pc_mean);
    // Singular Value Decomposition: SVD
    JacobiSVD<MatrixXf> svd(cov, Eigen::DecompositionOptions::ComputeFullU);
    // use the least singular vector as normal
    normal_ = (svd.matrixU().col(2));
    // mean ground seeds value
    Eigen::Vector3f seeds_mean = pc_mean.head<3>();


    // according to normal.T*[x,y,z] = -d
    d_ = -(normal_.transpose() * seeds_mean)(0, 0);
    // set distance threhold to `th_dist - d`
    th_dist_d_ = th_dist_ - d_;


    // return the equation parameters
}

extract_initial_seeds_(laserCloudIn);
g_ground_pc = g_seeds_pc;
// Ground plane fitter mainloop
for (int i = 0; i < num_iter_; i++)
{
    estimate_plane_();
    g_ground_pc->clear();
    g_not_ground_pc->clear();


    //pointcloud to matrix
    MatrixXf points(laserCloudIn_org.points.size(), 3);
    int j = 0;
    for (auto p : laserCloudIn_org.points)
    {
        points.row(j++) << p.x, p.y, p.z;
    }
    // ground plane model
    VectorXf result = points * normal_;
    // threshold filter
    for (int r = 0; r < result.rows(); r++)
    {
        if (result[r] < th_dist_d_)
        {
            g_all_pc->points[r].label = 1u; // means ground
            g_ground_pc->points.push_back(laserCloudIn_org[r]);
        }
        else
        {
            g_all_pc->points[r].label = 0u; // means not ground and non clusterred
            g_not_ground_pc->points.push_back(laserCloudIn_org[r]);
        }
    }
}

/*!
 *
 * @param[in] in_cloud Input Point Cloud to be organized in radial segments
 * @param[out] out_organized_points Custom Point Cloud filled with XYZRTZColor data
 * @param[out] out_radial_divided_indices Indices of the points in the original cloud for each radial segment
 * @param[out] out_radial_ordered_clouds Vector of Points Clouds, each element will contain the points ordered
 */
void PclTestCore::XYZI_to_RTZColor(const pcl::PointCloud<pcl::PointXYZI>::Ptr in_cloud,
                                   PointCloudXYZIRTColor &out_organized_points,
                                   std::vector<pcl::PointIndices> &out_radial_divided_indices,
                                   std::vector<PointCloudXYZIRTColor> &out_radial_ordered_clouds)
{
    out_organized_points.resize(in_cloud->points.size());
    out_radial_divided_indices.clear();
    out_radial_divided_indices.resize(radial_dividers_num_);
    out_radial_ordered_clouds.resize(radial_dividers_num_);

    for (size_t i = 0; i < in_cloud->points.size(); i++)
    {
        PointXYZIRTColor new_point;
//计算radius和theta
//方便转到柱坐标下。
        auto radius = (float)sqrt(
            in_cloud->points[i].x * in_cloud->points[i].x + in_cloud->points[i].y * in_cloud->points[i].y);
        auto theta = (float)atan2(in_cloud->points[i].y, in_cloud->points[i].x) * 180 / M_PI;
        if (theta < 0)
        {
            theta += 360;
        }
        //角度的微分
        auto radial_div = (size_t)floor(theta / RADIAL_DIVIDER_ANGLE);
        //半径的微分
        auto concentric_div = (size_t)floor(fabs(radius / concentric_divider_distance_));

        new_point.point = in_cloud->points[i];
        new_point.radius = radius;
        new_point.theta = theta;
        new_point.radial_div = radial_div;
        new_point.concentric_div = concentric_div;
        new_point.original_index = i;

        out_organized_points[i] = new_point;

        //radial divisions更加角度的微分组织射线
        out_radial_divided_indices[radial_div].indices.push_back(i);

        out_radial_ordered_clouds[radial_div].push_back(new_point);

    } //end for

    //将同一根射线上的点按照半径（距离）排序
#pragma omp for
    for (size_t i = 0; i < radial_dividers_num_; i++)
    {
        std::sort(out_radial_ordered_clouds[i].begin(), out_radial_ordered_clouds[i].end(),
                  [](const PointXYZIRTColor &a, const PointXYZIRTColor &b) { return a.radius < b.radius; });
    }
}

/*!
 * Classifies Points in the PointCoud as Ground and Not Ground
 * @param in_radial_ordered_clouds Vector of an Ordered PointsCloud ordered by radial distance from the origin
 * @param out_ground_indices Returns the indices of the points classified as ground in the original PointCloud
 * @param out_no_ground_indices Returns the indices of the points classified as not ground in the original PointCloud
 */
void PclTestCore::classify_pc(std::vector<PointCloudXYZIRTColor> &in_radial_ordered_clouds,
                              pcl::PointIndices &out_ground_indices,
                              pcl::PointIndices &out_no_ground_indices)
{
    out_ground_indices.indices.clear();
    out_no_ground_indices.indices.clear();
#pragma omp for
    for (size_t i = 0; i < in_radial_ordered_clouds.size(); i++) //sweep through each radial division 遍历每一根射线
    {
        float prev_radius = 0.f;
        float prev_height = -SENSOR_HEIGHT;
        bool prev_ground = false;
        bool current_ground = false;
        for (size_t j = 0; j < in_radial_ordered_clouds[i].size(); j++) //loop through each point in the radial div
        {
            float points_distance = in_radial_ordered_clouds[i][j].radius - prev_radius;//计算当前点和上一个点的水平距离pointdistance
            float height_threshold = tan(DEG2RAD(local_max_slope_)) * points_distance;//根据local_max_slope_和pointdistance计算当前的坡度差阈值height_threshold
            float current_height = in_radial_ordered_clouds[i][j].point.z;
            float general_height_threshold = tan(DEG2RAD(general_max_slope_)) * in_radial_ordered_clouds[i][j].radius;//根据general_max_slope_和当前点的水平距离计算整个地面的高度差阈值general_height_threshold

            //for points which are very close causing the height threshold to be tiny, set a minimum value
            if (points_distance > concentric_divider_distance_ && height_threshold < min_height_threshold_)
            {
                height_threshold = min_height_threshold_;
            }

            //check current point height against the LOCAL threshold (previous point)
            if (current_height <= (prev_height + height_threshold) && current_height >= (prev_height - height_threshold))
            {
                //Check again using general geometry (radius from origin) if previous points wasn't ground
                if (!prev_ground)
                {
                    if (current_height <= (-SENSOR_HEIGHT + general_height_threshold) && current_height >= (-SENSOR_HEIGHT - general_height_threshold))
                    {
                        current_ground = true;
                    }
                    else
                    {
                        current_ground = false;
                    }
                }
                else
                {
                    current_ground = true;
                }
            }
            else
            {
                //check if previous point is too far from previous one, if so classify again
                if (points_distance > reclass_distance_threshold_ &&
                    (current_height <= (-SENSOR_HEIGHT + height_threshold) && current_height >= (-SENSOR_HEIGHT - height_threshold)))
                {
                    current_ground = true;
                }
                else
                {
                    current_ground = false;
                }
            }

            if (current_ground)
            {
                out_ground_indices.indices.push_back(in_radial_ordered_clouds[i][j].original_index);
                prev_ground = true;
            }
            else
            {
                out_no_ground_indices.indices.push_back(in_radial_ordered_clouds[i][j].original_index);
                prev_ground = false;
            }

            prev_radius = in_radial_ordered_clouds[i][j].radius;
            prev_height = in_radial_ordered_clouds[i][j].point.z;
        }
    }
}