PCL（ Point Cloud Library）是用于处理2D/3D 图像以及点云的一个大型开源项目。学习PCL最好的途径是阅读其官网文档（Point Cloud Library (PCL)）。虽然PCL的网站文档稍微有点“丑”，但是其内容十分详尽。从应用的角度而言，PCL可以用于点云的分割、分类、校准以及可视化等方面。从理论角度而言，PCL中包含的众多算法能更好得帮助人们理解与创造新的点云算法。无论是工业应用还是科研攻关，PCL都能在三维数据处理领域祝您一臂之力。

激光雷达作为自动驾驶最常用的传感器，经常需要使用激光雷达来做建图、定位和感知等任务。

而这时候使用降低点云规模的预处理方法，可以能够去除无关区域的点以及降低点云规模。并能够给后续的PCL点云分割带来有效的收益。

1. 特征提取

1.1. 三维激光雷达压缩成二维

void filterGroundPlane(const PCLPointCloud& pc, PCLPointCloud& ground, PCLPointCloud& nonground) const{  ground.header = pc.header;  nonground.header = pc.header;  if (pc.size() < 50){    ROS_WARN("Pointcloud in OctomapServer too small, skipping ground plane extraction");    nonground = pc;  } else {      // https://blog.csdn.net/weixin_41552975/article/details/120428619    // 指模型参数，如果是平面的话应该是指a b c d四个参数值    pcl::ModelCoefficients::Ptr coefficients (new pcl::ModelCoefficients);    pcl::PointIndices::Ptr inliers (new pcl::PointIndices);    // 创建分割对象    pcl::SACSegmentation<PCLPoint> seg;    //可选设置    seg.setOptimizeCoefficients (true);    //必须设置    seg.setModelType(pcl::SACMODEL_PERPENDICULAR_PLANE);    seg.setMethodType(pcl::SAC_RANSAC);    // 设置迭代次数的上限    seg.setMaxIterations(200);    // 设置距离阈值    seg.setDistanceThreshold (0.04);    //设置所搜索平面垂直的轴     seg.setAxis(Eigen::Vector3f(0,0,1));    //设置待检测的平面模型和上述轴的最大角度    seg.setEpsAngle(0.15);    // pc 赋值    PCLPointCloud cloud_filtered(pc);    //创建滤波器    pcl::ExtractIndices<PCLPoint> extract;    bool groundPlaneFound = false;    while(cloud_filtered.size() > 10 && !groundPlaneFound){         // 所有点云传入，并通过coefficients提取到所有平面      seg.setInputCloud(cloud_filtered.makeShared());      seg.segment (*inliers, *coefficients);      if (inliers->indices.size () == 0){        ROS_INFO("PCL segmentation did not find any plane.");        break;      }      //输入要滤波的点云      extract.setInputCloud(cloud_filtered.makeShared());      //被提取的点的索引集合      extract.setIndices(inliers);      if (std::abs(coefficients->values.at(3)) < 0.07){        ROS_DEBUG("Ground plane found: %zu/%zu inliers. Coeff: %f %f %f %f", inliers->indices.size(), cloud_filtered.size(),                  coefficients->values.at(0), coefficients->values.at(1), coefficients->values.at(2), coefficients->values.at(3));        //true:滤波结果取反，false，则是取正        extract.setNegative (false);        //获取地面点集合，并传入ground        extract.filter (ground);        // 存在有不是平面的点        if(inliers->indices.size() != cloud_filtered.size()){          extract.setNegative(true);          PCLPointCloud cloud_out;          // 传入cloud_out          extract.filter(cloud_out);          // 不断减少cloud_filtered数目，同时累加nonground数目          cloud_filtered = cloud_out;          nonground += cloud_out;        }        groundPlaneFound = true;      } else{ // 否则提取那些不是平面的，然后剩下的就是平面点        ROS_DEBUG("Horizontal plane (not ground) found: %zu/%zu inliers. Coeff: %f %f %f %f", inliers->indices.size(), cloud_filtered.size(),                  coefficients->values.at(0), coefficients->values.at(1), coefficients->values.at(2), coefficients->values.at(3));        pcl::PointCloud<PCLPoint> cloud_out;        extract.setNegative (false);        extract.filter(cloud_out);        nonground +=cloud_out;        if(inliers->indices.size() != cloud_filtered.size()){          extract.setNegative(true);          cloud_out.points.clear();          extract.filter(cloud_out);          cloud_filtered = cloud_out;        } else{          cloud_filtered.points.clear();        }      }    }    // 由于没有找到平面，则会进入下面    if (!groundPlaneFound){      ROS_WARN("No ground plane found in scan");      // 对高度进行粗略调整，以防止出现虚假障碍物      pcl::PassThrough<PCLPoint> second_pass;      second_pass.setFilterFieldName("z");      second_pass.setFilterLimits(-m_groundFilterPlaneDistance, m_groundFilterPlaneDistance);      second_pass.setInputCloud(pc.makeShared());      second_pass.filter(ground);      second_pass.setFilterLimitsNegative (true);      second_pass.filter(nonground);    }    // Create a set of planar coefficients with X=Y=0,Z=1    pcl::ModelCoefficients::Ptr coefficients1(new pcl::ModelCoefficients());    coefficients1->values.resize(4);    coefficients1->values[0] = 1;    coefficients1->values[1] = 0;    coefficients1->values[2] = 0;    coefficients1->values[3] = 0;    // Create the filtering object    pcl::PointCloud<pcl::PointXYZ>::Ptr cloud_projected(new pcl::PointCloud<pcl::PointXYZ>);    pcl::ProjectInliers<pcl::PointXYZ> proj;    proj.setModelType(pcl::SACMODEL_PLANE);    proj.setInputCloud(nonground);    proj.setModelCoefficients(coefficients1);    proj.filter(*cloud_projected);    if (cloud_projected.size() > 0)             writer.write<PCLPoint>("cloud_projected.pcd",cloud_projected, false);  }}

1.2. 面特征提取

PCL中Sample——consensus模块提供了RANSAC平面拟合模块。

SACMODEL_PLANE 模型：定义为平面模型，共设置四个参数 [normal_x，normal_y，normal_z，d]。其中，（normal_x，normal_y，normal_z）为平面法向量，d为常数项。

pcl::SACSegmentationFromNormals seg;//创建分割时所需要的模型系数对象，coefficients及存储内点的点索引集合对象inliers 
pcl::ModelCoefficients::Ptr coefficients(new pcl::ModelCoefficients); 
pcl::PointIndices::Ptr inliers(new pcl::PointIndices); 
// 创建分割对象 
pcl::SACSegmentation& lt;
pcl::PointXYZ& gt;
// 可选择配置，设置模型系数需要优化
seg.setOptimizeCoefficients(true); 
// 必要的配置，设置分割的模型类型，所用的随机参数估计方法，距离阀值，输入点云 
seg.setModelType(pcl::SACMODEL_PLANE); //设置模型类型 
seg.setMethodType(pcl::SAC_RANSAC);
//设置随机采样一致性方法类型 
seg.setDistanceThreshold(0.01);
//设定距离阀值，距离阀值决定了点被认为是局内点是必须满足的条件国，表示点到估计模型的距离最大值
seg.setInputCloud(cloud);
//引发分割实现，存储分割结果到点几何inliers及存储平面模型的系数coefficients 
seg.segment(*inliers, *coefficients);

1.3. 圆柱体提取

圆柱体的提取也是基于Ransec来实现提取，RANSAC从样本中随机抽选出一个样本子集，使用最小方差估计算法对这个子集计算模型参数，然后计算所有样本与该模型的偏差。

再使用一个预先设定好的阈值与偏差比较，当偏差小于阈值时，该样本点属于模型内样本点（inliers），简称内点，否则为模型外样本点（outliers），简称外点。

pcl::SACSegmentationFromNormals seg;// Create the segmentation object for cylinder segmentation and set all the parameters
seg.setOptimizeCoefficients(true);
seg.setModelType(pcl::SACMODEL_CYLINDER);   // 提取圆柱体的操作
seg.setMethodType(pcl::SAC_RANSAC);
seg.setNormalDistanceWeight(0.1);
seg.setMaxIterations(10000);
seg.setDistanceThreshold(0.05);   // 距离5cm
seg.setRadiusLimits(0, 0.1);    // 半径 10cm
seg.setInputCloud(cloud_filtered2);
seg.setInputNormals(cloud_normals2);// Obtain the cylinder inliers and coefficients
seg.segment(*inliers_cylinder, *coefficients_cylinder);
std::cerr << "Cylinder coefficients: " << *coefficients_cylinder << std::endl;

1.4. 半径近邻

半径内近邻搜索（Neighbors within Radius Search），是指搜索点云中一点在球体半径 R内的所有近邻点。

// Neighbors within radius search
std::vector pointIdxRadiusSearch;
std::vector pointRadiusSquaredDistance;
float radius = 256.0f * rand () / (RAND_MAX + 1.0f);if ( kdtree.radiusSearch (searchPoint, radius, pointIdxRadiusSearch, pointRadiusSquaredDistance) > 0 )
{for (size_t i = 0; i < pointIdxRadiusSearch.size (); ++i)std::cout << "    "  <<   cloud->points[ pointIdxRadiusSearch[i] ].x << " " << cloud->points[ pointIdxRadiusSearch[i] ].y << " " << cloud->points[ pointIdxRadiusSearch[i] ].z << " (squared distance: " << pointRadiusSquaredDistance[i] << ")" << std::endl;
}

1.5. 聚类

首先选取种子点，利用kd-tree对种子点进行半径r邻域搜索，若邻域内存在点，则与种子点归为同一聚类簇Q；

欧式聚类：
void Cvisualization::ShowCloud4()
{//读入点云数据table_scene_lms400.pcdpcl::PCDReader reader;pcl::PointCloud::Ptr cloud (new pcl::PointCloud), cloud_f (new pcl::PointCloud);reader.read ("E:/ai/pcltest/20210903changhuAM-0001.pcd", *cloud);std::cout << "PointCloud before filtering has: " << cloud->points.size () << " data points." << std::endl; //*//    /*从输入的.PCD文件载入数据后，我们创建了一个VoxelGrid滤波器对数据进行下采样，我们在这里进行下采样的原       因是来加速处理过程，越少的点意味着分割循环中处理起来越快。*/// Create the filtering object: downsample the dataset using a leaf size of 1cmpcl::VoxelGrid vg; //体素栅格下采样对象pcl::PointCloud::Ptr cloud_filtered (new pcl::PointCloud);vg.setInputCloud (cloud);vg.setLeafSize (0.01f, 0.01f, 0.01f); //设置采样的体素大小vg.filter (*cloud_filtered);  //执行采样保存数据std::cout << "PointCloud after filtering has: " << cloud_filtered->points.size ()  << " data points." << std::endl; //*// Create the segmentation object for the planar model and set all the parameterspcl::SACSegmentation seg;//创建分割对象pcl::PointIndices::Ptr inliers (new pcl::PointIndices);pcl::ModelCoefficients::Ptr coefficients (new pcl::ModelCoefficients);pcl::PointCloud::Ptr cloud_plane (new pcl::PointCloud ());pcl::PCDWriter writer;seg.setOptimizeCoefficients (true);  //设置对估计的模型参数进行优化处理seg.setModelType (pcl::SACMODEL_PLANE);//设置分割模型类别seg.setMethodType (pcl::SAC_RANSAC);//设置用哪个随机参数估计方法seg.setMaxIterations (100);  //设置最大迭代次数seg.setDistanceThreshold (0.02);    //设置判断是否为模型内点的距离阈值int i=0, nr_points = (int) cloud_filtered->points.size ();while (cloud_filtered->points.size () > 0.3 * nr_points){// Segment the largest planar component from the remaining cloud//      /*为了处理点云中包含多个模型，我们在一个循环中执行该过程，并在每次模型被提取后，我们保存剩余的点，进行迭代。模型内点通过分割过程获取，如下*/seg.setInputCloud (cloud_filtered);seg.segment (*inliers, *coefficients);if (inliers->indices.size () == 0){std::cout << "Could not estimate a planar model for the given dataset." << std::endl;break;}//移去平面局内点，提取剩余点云pcl::ExtractIndices extract;   //创建点云提取对象extract.setInputCloud (cloud_filtered);    //设置输入点云extract.setIndices (inliers);   //设置分割后的内点为需要提取的点集extract.setNegative (false); //设置提取内点而非外点// Get the points associated with the planar surfaceextract.filter (*cloud_plane);   //提取输出存储到cloud_planestd::cout << "PointCloud representing the planar component: " << cloud_plane->points.size () << " data points." << std::endl;// Remove the planar inliers, extract the restextract.setNegative (true);extract.filter (*cloud_f);*cloud_filtered = *cloud_f;}// Creating the KdTree object for the search method of the extractionpcl::search::KdTree::Ptr tree (new pcl::search::KdTree);tree->setInputCloud (cloud_filtered); //创建点云索引向量，用于存储实际的点云信息std::vector cluster_indices;pcl::EuclideanClusterExtraction ec;ec.setClusterTolerance (0.2); //设置近邻搜索的搜索半径为2cmec.setMinClusterSize (100);//设置一个聚类需要的最少点数目为100ec.setMaxClusterSize (25000);//设置一个聚类需要的最大点数目为25000ec.setSearchMethod (tree);//设置点云的搜索机制ec.setInputCloud (cloud_filtered);ec.extract (cluster_indices);//从点云中提取聚类，并将点云索引保存在cluster_indices中//    /* 为了从点云索引向量中分割出每个聚类，必须迭代访问点云索引，每次创建一个新的点云数据集，并且将所有当前聚类的点写入到点云数据集中 *///迭代访问点云索引cluster_indices，直到分割出所有聚类int j = 0;for (std::vector::const_iterator it = cluster_indices.begin (); it != cluster_indices.end (); ++it){pcl::PointCloud::Ptr cloud_cluster (new pcl::PointCloud);//创建新的点云数据集cloud_cluster，将所有当前聚类写入到点云数据集中for (std::vector::const_iterator pit = it->indices.begin (); pit != it->indices.end (); ++pit)cloud_cluster->points.push_back (cloud_filtered->points[*pit]); //*cloud_cluster->width = cloud_cluster->points.size ();cloud_cluster->height = 1;cloud_cluster->is_dense = true;std::cout << "PointCloud representing the Cluster: " << cloud_cluster->points.size () << " data points." << std::endl;std::stringstream ss;ss << "E:/ai/pcltest/cloud_cluster_" << j << ".pcd";writer.write (ss.str (), *cloud_cluster, false);j++;}pcl::visualization::PCLVisualizer::Ptr viewer(new pcl::visualization::PCLVisualizer("HelloMyFirstVisualPCL"));viewer->addPointCloud(cloud, "sample cloud");while (!viewer->wasStopped()){viewer->spinOnce(100);boost::this_thread::sleep(boost::posix_time::microseconds(100000));}
}

1.6. 区域生长

区域生长的基本思想是将具有相似性质的点集合起来构成区域。

首先对每个需要分割的区域找出一个种子作为生长的起点,然后将种子周围邻域中与种子有相同或相似性质的点(根据事先确定的生长或相似准则来确定，多为法向量、曲率)归并到种子所在的区域中。

#include 
#include 
#include 
#include 
#include 
#include 
#include 
#include int main()
{pcl::PointCloud::Ptr cloud(new pcl::PointCloud);if (pcl::io::loadPCDFile("data//table_scene_lms400.pcd", *cloud) == -1){std::cout << "Cloud reading failed." << std::endl;return (-1);}// 设置搜索方式为kdTreepcl::search::Search::Ptr tree(new pcl::search::KdTree);// 计算法向量pcl::PointCloud ::Ptr normals(new pcl::PointCloud );pcl::NormalEstimation normal_estimator;normal_estimator.setSearchMethod(tree);normal_estimator.setInputCloud(cloud);normal_estimator.setKSearch(50);normal_estimator.compute(*normals);//直通滤波在Z轴的0到1米之间pcl::IndicesPtr indices(new std::vector );pcl::PassThrough pass;pass.setInputCloud(cloud);pass.setFilterFieldName("z");pass.setFilterLimits(0.0, 1.0);pass.filter(*indices);// 欧式聚类pcl::RegionGrowing reg;reg.setMinClusterSize(5000);     //最小的聚类的点数reg.setMaxClusterSize(1000000);  //最大的聚类的点数reg.setSearchMethod(tree);       //搜索方式reg.setNumberOfNeighbours(30);   //设置搜索的邻域点的个数reg.setInputCloud(cloud);        //输入点//reg.setIndices (indices);reg.setInputNormals(normals);    //输入的法线reg.setSmoothnessThreshold(3.0 / 180.0 * M_PI);  //设置平滑度reg.setCurvatureThreshold(1.0);  //设置曲率的阀值// 获取聚类的结果，分割结果保存在点云索引的向量中std::vector  clusters;reg.extract(clusters);//输出聚类的数量std::cout << "Number of clusters is equal to " << clusters.size() << std::endl;// 输出第一个聚类的数量std::cout << "First cluster has " << clusters[0].indices.size() << " points." << endl;std::cout << "These are the indices of the points of the initial" <::Ptr colored_cloud = reg.getColoredCloud();pcl::visualization::CloudViewer viewer("Cluster viewer");viewer.showCloud(colored_cloud);while (!viewer.wasStopped()){}return (0);
}

1.7. 线特征拟合

一般线特征拟合的方式前提是先要滤除不必要的点，而这个就需要使用K-D tree来先实现搜索

#include 
#include 
#include 
#include 
#include 
#include 
#include using namespace std::chrono_literals;pcl::visualization::PCLVisualizer::Ptr
simpleVis(pcl::PointCloud::ConstPtr cloud)
{// --------------------------------------------// -----Open 3D viewer and add point cloud-----// --------------------------------------------pcl::visualization::PCLVisualizer::Ptr viewer(new pcl::visualization::PCLVisualizer("3D Viewer"));viewer->setBackgroundColor(0, 0, 0);viewer->addPointCloud(cloud, "sample cloud");viewer->setPointCloudRenderingProperties(pcl::visualization::PCL_VISUALIZER_POINT_SIZE, 3, "sample cloud");// viewer->addCoordinateSystem (1.0, "global");//viewer->initCameraParameters();return (viewer);
}pcl::PointCloud::Ptr
create_line(double x0, double y0, double z0, double a, double b, double c, double point_size = 1000, double step = 0.1)
{pcl::PointCloud::Ptr cloud_line(new pcl::PointCloud);cloud_line->width = point_size;cloud_line->height = 1;cloud_line->resize(cloud_line->width * cloud_line->height);for (std::size_t i = 0; i < cloud_line->points.size(); ++i) {cloud_line->points[i].x = x0 + a / std::pow(a * a + b * b + c * c, 0.5) * i * 0.1;cloud_line->points[i].y = y0 + b / std::pow(a * a + b * b + c * c, 0.5) * i * 0.1;cloud_line->points[i].z = z0 + c / std::pow(a * a + b * b + c * c, 0.5) * i * 0.1;}return cloud_line;
}void fit_line(pcl::PointCloud::Ptr& cloud, double distance_threshold)
{// fit line from a point cloudpcl::ModelCoefficients::Ptr coefficients1(new pcl::ModelCoefficients);pcl::PointIndices::Ptr inliers1(new pcl::PointIndices);pcl::SACSegmentation seg;seg.setOptimizeCoefficients(true);seg.setModelType(pcl::SACMODEL_LINE);seg.setMethodType(pcl::SAC_RANSAC);seg.setMaxIterations(1000);seg.setDistanceThreshold(distance_threshold);seg.setInputCloud(cloud);seg.segment(*inliers1, *coefficients1);// line parametersdouble x0, y0, z0, a, b, c;x0 = coefficients1->values[0];y0 = coefficients1->values[1];z0 = coefficients1->values[2];a = coefficients1->values[3];b = coefficients1->values[4];c = coefficients1->values[5];std::cout << "model parameters1:"<< "   (x - " << x0 << ") / " << a << " = (y - " << y0 << ") / " << b<< " = (z - " << z0 << ") / " << c << std::endl;// extract segmentation partpcl::PointCloud::Ptr cloud_line1(new pcl::PointCloud);pcl::ExtractIndices extract;extract.setInputCloud(cloud);extract.setIndices(inliers1);extract.setNegative(false);extract.filter(*cloud_line1);// extract remain pointcloudpcl::PointCloud::Ptr cloud_remain(new pcl::PointCloud);extract.setNegative(true);extract.filter(*cloud_remain);//显示原始点云pcl::visualization::PCLVisualizer::Ptr viewer_ori;viewer_ori = simpleVis(cloud);while (!viewer_ori->wasStopped()) {viewer_ori->spinOnce(100);std::this_thread::sleep_for(100ms);}pcl::visualization::PCLVisualizer::Ptr viewer(new pcl::visualization::PCLVisualizer("3D Viewer"));viewer->setBackgroundColor(0, 0, 0);viewer->addPointCloud(cloud_remain, "cloud_remain");viewer->setPointCloudRenderingProperties(pcl::visualization::PCL_VISUALIZER_POINT_SIZE, 2, "cloud_remain");viewer->addPointCloud(cloud_line1, "cloud_line1");viewer->setPointCloudRenderingProperties(pcl::visualization::PCL_VISUALIZER_POINT_SIZE, 5, "cloud_line1");viewer->setPointCloudRenderingProperties(pcl::visualization::PCL_VISUALIZER_COLOR, 1.0, 0.5, 0.5, "cloud_line1");while (!viewer->wasStopped()) {viewer->spinOnce(100);std::this_thread::sleep_for(100ms);}
}void demo()
{// line parametersdouble x0 = -2, y0 = -2, z0 = 0, a = 1, b = 1, c = 0;auto line_pcd_create = create_line(x0, y0, z0, a, b, c);pcl::PointCloud::Ptr cloud_noise(new pcl::PointCloud);std::size_t noise_points_size = line_pcd_create->points.size() / 10;cloud_noise->width = noise_points_size;cloud_noise->height = 1;cloud_noise->points.resize(cloud_noise->width * cloud_noise->height);// add noisefor (std::size_t i = 0; i < noise_points_size; ++i) {int random_num = line_pcd_create->points.size() * rand() / (RAND_MAX + 1.0f);cloud_noise->points[i].x =line_pcd_create->points[random_num].x + 10 * rand() / (RAND_MAX + 1.0f) - 5;cloud_noise->points[i].y =line_pcd_create->points[random_num].y + 10 * rand() / (RAND_MAX + 1.0f) - 5;cloud_noise->points[i].z =line_pcd_create->points[random_num].z + 10 * rand() / (RAND_MAX + 1.0f) - 5;}pcl::PointCloud::Ptr line_with_noise(new pcl::PointCloud);*line_with_noise = *cloud_noise + *line_pcd_create;fit_line(line_with_noise, 1);
}int main(int argc, char* argv[])
{if (argc < 3) {std::cout << "please input parametars:\nfilepath\ndistance_threshold" << std::endl;demo();return -1;}std::string file_path = argv[1];double distance_threshold = atof(argv[2]);pcl::PointCloud::Ptr cloud(new pcl::PointCloud);if (pcl::io::loadPLYFile(file_path, *cloud) < 0) {std::cout << "can not read file " << file_path << std::endl;return -1;}std::cout << "point size: " << cloud->points.size() << std::endl;fit_line(cloud, distance_threshold);return 0;
}

1.8. 点特征提取

点特征的提取和线特征的提取原理一样

    pcl::HarrisKeypoint3D<pcl::PointXYZ, pcl::PointXYZI, pcl::Normal> harris;    harris.setInputCloud(cloud);//设置输入点云 指针    harris.setNonMaxSupression(true);    harris.setRadius(0.6f);//　块体半径    harris.setThreshold(0.01f);//数量阈值    //新建的点云必须初始化，清零，否则指针会越界    //注意Harris的输出点云必须是有强度(I)信息的 pcl::PointXYZI，因为评估值保存在I分量里    pcl::PointCloud<pcl::PointXYZI>::Ptr cloud_out_ptr(new pcl::PointCloud<pcl::PointXYZI>);    // 计算特征点    harris.compute(*cloud_out_ptr);

参考文献

自动驾驶-激光雷达预处理/特征提取

PCL入门系列一——PCL简介及PCL安装 - 知乎

pcl教程（五）聚类_紫沐衙的博客-CSDN博客