# pcl_learning **Repository Path**: weipushan/pcl_learning ## Basic Information - **Project Name**: pcl_learning - **Description**: 个人 PCL 点云库学习仓库。 - **Primary Language**: Unknown - **License**: Not specified - **Default Branch**: master - **Homepage**: None - **GVP Project**: No ## Statistics - **Stars**: 0 - **Forks**: 0 - **Created**: 2022-10-20 - **Last Updated**: 2023-10-16 ## Categories & Tags **Categories**: Uncategorized **Tags**: None ## README # PCL_Learning 个人学习 PCL (Point Cloud Library) 的仓库. ## Install Note - 安装配置 - Ubuntu 18.04 - GPU RTX 3060 + CUDA 11.1 - PCL 1.9.1 + VTK 8.2.0 - 主要参考 PCL [安装教程](https://blog.csdn.net/qq_42257666/article/details/124574029) - 安装依赖 ```bash sudo apt-get update sudo apt-get install git build-essential linux-libc-dev sudo apt-get install cmake cmake-gui sudo apt-get install libusb-1.0-0-dev libusb-dev libudev-dev sudo apt-get install mpi-default-dev openmpi-bin openmpi-common sudo apt-get install libflann1.9 libflann-dev sudo apt-get install libeigen3-dev sudo apt-get install libboost-all-dev sudo apt-get install libvtk7.1-qt sudo apt-get install libvtk7.1 sudo apt-get install libvtk7-qt-dev sudo apt-get install libqhull* libgtest-dev sudo apt-get install freeglut3-dev pkg-config sudo apt-get install libxmu-dev libxi-dev sudo apt-get install mono-complete sudo apt-get install openjdk-8-jdk openjdk-8-jre ``` - 安装 VTK - 安装依赖 ```bash sudo apt-get install cmake-curses-gui sudo apt-get install freeglut3-dev ``` - 安装 [VTK-8.2.0](https://vtk.org/download/) - mkdir build - 通过 cmake-gui 配置依赖 - 选择source 和 build 路径 - 勾选grouped advanced - 点击configuration - 红色区域中勾选 Module_vtkGUISupportQt、VTK_Group_Qt - 重新 configure, generate - 如果有问题就删除build内容, 重新configure generate - sudo make - sudo make install - 编译 PCL ```bash $ mkdir -p build/installed $ cd build $ cmake \ > -DCMAKE_BUILD_TYPE=None \ > -DCMAKE_INSTALL_PREFIX=./installed \ > -DBUILD_GPU=ON \ > -DBUILD_apps=ON \ > -DBUILD_examples=ON \ > .. ``` - `Error generating file pcl_gpu_octree_generated_knn_search.cu.o` - sm_30 sm_70 及以上不支持, 删掉 `cmake/pcl_find_cuda.cmake` 中sm_30, sm_70 及以上 ```cmake if(NOT ${CUDA_VERSION_STRING} VERSION_LESS "10.0") # set(__cuda_arch_bin "3.0 3.5 5.0 5.2 5.3 6.0 6.1 7.0 7.2 7.5") set(__cuda_arch_bin "3.5 5.0 5.2 5.3 6.0 6.1") ``` - `No CMAKE_CUDA_COMPILER could be found` - `.zshrc` 中添加 `export PATH=$PATH:/usr/local/cuda-11.1/bin/` # PCL Official Tutorial [PCL Official Tutorial](https://pcl.readthedocs.io/projects/tutorials/en/master/) ## 1 Basic Usage ### 1.1 PCL Walkthrough PCL 点云库包括以下模块: - Filters - 点云滤波, 如降采样, 过滤噪音离群点等 - Features - 点云特征, 比如局部点云的法线, 弯曲度等 - Keypoints - 点云关键点 - Registration - 点云配准 - KdTree - Kd树 - Octree - 八叉树 - Segmentation - 点云分割 - Sample Consensus - 样本一致性 如[RANSAC](https://zhuanlan.zhihu.com/p/402727549) - Surface - 点云表面重建 - Range Image - 深度图 - IO - 点云输入输出, 如读写PCD - Visualization - 点云可视化 - Common - PCL的通用模块 - Search - 点云搜索库 - Binaries - 常用的PCL工具, 如pcl_viewer ### 1.2 Getting Started / Basic Structures - PointCloud - organized point cloud - 表示像图片一样分为行列的点数据 - projectable point cloud - 可以通过相机模型转换为图像UV坐标 - $u = fx/z, v = fy/z$ - `width (int)` - 可以指代 organized point cloud 的一行点数量, 即宽度 - 可以指代 unorganized point cloud 的点总数 - `height (int)` - 可以指代 organized point cloud 的一列点数量, 即高度 - unorganized point cloud 时设置为1 - `points (std::vector)` - 保存点云中所有的点 - `is_dense (bool)` - 表示所有点的数据都是有限的, finit (true), Inf/NaN (fasle) - `sensor_origin_ (Eigen::Vector4f)` - 点云传感器位姿 translation - `sensor_orientation_ (Eigen::Quaternionf)` - 点云传感器位姿 orientation - `is_Organized()` - 判断是否是 organized point cloud ### 1.3 Using PCL in your own project - CMakeLists.txt 指定 PCL 安装位置 ```cmake find_package(PCL 1.9.1 REQUIRED COMPONENTS common io PATHS "/home/shan/App/pcl/pcl-pcl-1.9.1/build/installed" NO_DEFAULT_PATH ) ``` - `error while loading shared libraries: libvtkCommonMisc-8.2.so.1` ```bash # 方法1: 将 /usr/local/lib 加入到 /etc/ld.so.conf, 然后执行ldconfig sudo chmod 777 /etc/ld.so.conf sudo echo "/usr/local/lib" >> /etc/ld.so.conf sudo ldconfig # 方法2: 临时添加到 LD_LIBRARY_PATH, 在 .zshrc 中添加以下内容 export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH ``` - example: `write_pcd.cpp` ### 1.4 Using a matrix to transform a point cloud - example: `matrix_transform.cpp` - 转换一个点或一个向量 ```bash # 点 添加1在末尾 [10, 5, 0, 1] * 4x4_transformation_matrix # 向量 添加0在末尾 [3, 0, -1, 0] * 4x4_transformation_matrix ``` ## 2 Advanced Usage ### 2.1 Adding your own custom PointT type #### PointT types available in PCL - PointXYZ, - member: float x, y, z - 通过联合体实现 SSE alignment (数据地址对齐) - ``` union { float data[4]; struct { float x; float y; float z; }; }; ``` - PointXYZI - member: float x, y, z, intensity - 单独设置一个联合体用于Intensity, 因为xyz的联合体最后一项data[3]会被设置为0/1用于transformation - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { struct { float intensity; }; float data_c[4]; }; ``` - PointXYZRGBA - member: float x, y, z; std::uint32_t rgba - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { union { struct { std::uint8_t b; std::uint8_t g; std::uint8_t r; std::uint8_t a; }; float rgb; }; std::uint32_t rgba; }; ``` - PointXYZRGB - same as PointXYZRGBA - member: float x, y, z; std::uint32_t rgba - PointXY - member: float x, y - ``` struct { float x; float y; }; ``` - InterestPoint - member: float x, y, z, strength - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { struct { float strength; }; float data_c[4]; }; ``` - Normal - member: float normal[3], curvature - ``` union { float data_n[4]; float normal[3]; struct { float normal_x; float normal_y; float normal_z; }; } union { struct { float curvature; }; float data_c[4]; }; ``` - PointNormal - member: float x, y, z; float normal[3], curvature - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { float data_n[4]; float normal[3]; struct { float normal_x; float normal_y; float normal_z; }; }; union { struct { float curvature; }; float data_c[4]; }; ``` - PointXYZRGBNormal - member: float x, y, z, normal[3], curvature; std::uint32_t rgba - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { float data_n[4]; float normal[3]; struct { float normal_x; float normal_y; float normal_z; }; } union { struct { union { union { struct { std::uint8_t b; std::uint8_t g; std::uint8_t r; std::uint8_t a; }; float rgb; }; std::uint32_t rgba; }; float curvature; }; float data_c[4]; }; ``` - PointXYZINormal - member: float x, y, z, intensity, normal[3], curvature - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { float data_n[4]; float normal[3]; struct { float normal_x; float normal_y; float normal_z; }; } union { struct { float intensity; float curvature; }; float data_c[4]; }; ``` - PointWithRange (depth map) - member: float x, y, z (union with float point[4]), range - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { struct { float range; }; float data_c[4]; }; ``` - PointWithViewpoint - member: float x, y, z, vp_x, vp_y, vp_z - ``` union { float data[4]; struct { float x; float y; float z; }; }; union { struct { float vp_x; float vp_y; float vp_z; }; float data_c[4]; }; ``` - MomentInvariants - member: float j1, j2, j3 - Simple point type holding the 3 moment invariants at a surface patch - ``` struct { float j1, j2, j3; }; ``` - PrincipalRadiiRSD - member: float r_min, r_max - Simple point type holding the 2 RSD radii at a surface patch - ``` struct { float r_min, r_max; }; ``` - Boundary - member: std::uint8_t boundary_point - Simple point type holding whether the point is lying on a surface boundary or not - ``` struct { std::uint8_t boundary_point; }; ``` - PrincipalCurvatures - member: float principal_curvature[3], pc1, pc2 - Simple point type holding the principal curvatures of a given point. - PFHSignature125 - member: float pfh[125] - Simple point type holding the PFH (Point Feature Histogram) of a given point. - FPFHSignature33 - member: float fpfh[33] - Simple point type holding the FPFH (Fast Point Feature Histogram) of a given point. - VFHSignature308 - member: float vfh[308] - Simple point type holding the VFH (Viewpoint Feature Histogram) of a given point. - Narf36 - member: float x, y, z, roll, pitch, yaw; float descriptor[36] - Simple point type holding the NARF (Normally Aligned Radius Feature) of a given point. - BorderDescription - member: int x, y; BorderTraits traits - Simple point type holding the border type of a given point. - IntensityGradient - member: float gradient[3] - Simple point type holding the intensity gradient of a given point. - Histogram - member: float histogram[N] - General purpose n-D histogram placeholder - PointWithScale - member: float x, y, z, scale - Similar to PointXYZI - PointSurfel - member: float x, y, z, normal[3], rgba, radius, confidence, curvature #### Add new PointT type example: 02-advanced_usage/01_adding/custom_ptype.cpp ### 2.2 Writing a new PCL class example: - pcl/filters/include/pcl/filters/bilateral.h - pcl/filters/include/pcl/filters/impl/bilateral.hpp - pcl/filters/src/bilateral.cpp ## 3 Visualization example - 01_simple_pcd_viewer.cpp - 简单的 pcd viewer - 02_multithread_pcd_viewer.cpp - 多线程的 pcd viewer - 03_range_image_viewer.cpp - 点云转化为深度图, 并可视化点云和深度图 - 效果与预期不符, 可能是视角的问题 - 04_pcl_visualizer.cpp - 实现给点云上色, 显示法线, 添加形状, 多视图, 键鼠交互等 ## 4 Feature ### 4.1 How 3D Features work in PCL - just cartesian coordiantes is not enough - other characteristics and metrics - intensity - surface remission - color - good point feature representation 好的点云特征 - rigid transformations - varying sampling density - noise - KD-树搜索 [KD-Tree](https://zhuanlan.zhihu.com/p/529487972) - k-search - k nearst neighbors - radius-search - all neast neightnors in a sphere of radius r - ranged k-search - k nearst neighbors in a range #### 4.1.1 How to pass the input - setInputCloud (PointCloudConstPtr &) - 输入点云 - setIndices (IndicesConstPtr &) - 输入索引, 用来指定需要操作的部分点云的索引 - setSearchSurface (PointCloudConstPtr &) - 输入搜索范围, 用来作为Input点云的搜索面 - 比如通过关键点去计算整个点云的特征 - 原始点云通过 setSearchSurface() 输入 - 关键点通过 setInputCloud() 输入 ![image-20221021160941662](readme.assets/image-20221021160941662.png) #### 4.1.2 An example for normal estimation - example: 03_pcl_feature/01_normal_estimation.cpp ### 4.2 Estimating Surface Normals in a PointCloud - example: 03_pcl_feature/01_normal_estimation.cpp 通过最小二乘的原理, 可以使用 PCA 来求解 covariance matrix 获得 eigen value 和 eigen vector 来求得法向量 normal, 选择最小特征值对应的特征向量，并进行单位归一化，则该向量为点云法向量 ```cpp // Placeholder for the 3x3 covariance matrix at each surface patch Eigen::Matrix3f covariance_matrix; // 16-bytes aligned placeholder for the XYZ centroid of a surface patch Eigen::Vector4f xyz_centroid; // Estimate the XYZ centroid compute3DCentroid (cloud, xyz_centroid); // Compute the 3x3 covariance matrix computeCovarianceMatrix (cloud, xyz_centroid, covariance_matrix); ``` 需要注意法向量的正反方向, 通过设置 viewpoint 可以统一法向量的方向 ```cpp flipNormalTowardsViewpoint ( const PointT &point, float vp_x, float vp_y, float vp_z, Eigen::Vector4f &normal); ``` 计算法向量时需要选择 k 个最近点 或 r 搜索半径, 这决定了法向量的计算结果, 下图左是较小的k/r, 右是太大的k/r ![image-20221021180126715](readme.assets/image-20221021180126715.png) 计算法向量的伪代码 ```pseudocode for each point p in cloud P 1. get the nearest neighbors of p 2. compute the surface normal n of p 3. check if n is consistently oriented towards the viewpoint and flip otherwise ``` 设置视角点 ```cpp setViewPoint (float vpx, float vpy, float vpz); ``` 计算单个点的法向量 ```cpp computePointNormal ( const pcl::PointCloud &cloud, const std::vector &indices, Eigen::Vector4f &plane_parameters, float &curvature); ``` 使用 OpenMP 加速 normal estimation - pcl::NormalEstimationOMP ### 4.3 Normal Estimation Using Integral Images - example: 03_pcl_feature/02_normal_estimation_using_integral_images.cpp ### 4.4 Point Feature Histograms (PFH) descriptors - PFH 点特征直方图 - 只有法线不足以表现点云的细节特征 - PFH 通过使用多维的直方图来生成某点周围的平均曲率来表现某点的K个邻点的几何关系 ![image-20221024101417703](readme.assets/image-20221024101417703.png) - 对某点的K个邻点, 前提是他们的法向量已知, 我们需要两两计算两点以及他们的法线之间的差距, 首先在每两个点之间设定一个坐标系UVW ![image-20221024101607431](readme.assets/image-20221024101607431.png) $$ \begin{aligned} \mathrm{\boldsymbol{u}} &=\boldsymbol{n}_s \\ \mathrm{\boldsymbol{v}} &=\mathrm{\boldsymbol{u}} \times \frac{\left(\boldsymbol{p}_t-\boldsymbol{p}_s\right)}{\left\|\boldsymbol{p}_t-\boldsymbol{p}_s\right\|_2} \\ \mathrm{\boldsymbol{w}} &=\mathrm{\boldsymbol{u}} \times \mathrm{\boldsymbol{v}} \end{aligned} $$ - 通过UVW计算两个点之间的四个特征 $(\alpha, \phi, \theta, d)$, 如此可将两个点的坐标与法线信息从(3+3)x2=12个减少到4个 $$ \begin{aligned} \alpha &=\mathbf{v} \cdot \boldsymbol{n}_t \\ \phi &=\mathbf{u} \cdot \frac{\left(\boldsymbol{p}_t-\boldsymbol{p}_s\right)}{d} \\ \theta &=\arctan \left(\mathbf{w} \cdot \boldsymbol{n}_t, \mathbf{u} \cdot \boldsymbol{n}_t\right) \\ d &= \left\|\boldsymbol{p}_t-\boldsymbol{p}_s\right\|_2 \end{aligned} $$ - 单独计算两个点的PFH - ```c++ computePairFeatures (const Eigen::Vector4f &p1, const Eigen::Vector4f &n1, const Eigen::Vector4f &p2, const Eigen::Vector4f &n2, float &f1, float &f2, float &f3, float &f4); ``` - 将以上计算的四种特征, 每种都分为b份, 并画成直方图, 就是PFH的完整表达 - example: 03_pcl_feature/03_pfh_estimation.cpp - Pseudo code ```pseudocode for each point p in cloud P 1. get the nearest neighbors of p 2. for each pair of neighbors, compute the three angular values 3. bin all the results in an output histogram ``` - 计算点云中单个点的PFH ```c++ computePointPFHSignature ( const pcl::PointCloud &cloud, const pcl::PointCloud &normals, const std::vector &indices, int nr_split, Eigen::VectorXf &pfh_histogram); ``` - 计算PFH之前最好检查一下数据是否finit ### 4.5 Fast Point Feature Histograms (FPFH) descriptors - 传统PFH算法的时间复杂度 $O(nk^2)$, 非常耗时 - FPFH 的时间复杂度为 $O(nk)$ - Step1: 计算点云中每个查询点 $p_q$ 与其K个邻点的 $\alpha, \phi, \theta$, 这步被称为 Simplified Point Feature Histogram (SPFH) - Sept2: 对于每个点, 重新确定K个邻点, 使用相邻的 $p_q$ 的 SPFH值用来加权, 权重 $w_i$ 表示查询点 $p_q$ 与相邻点 $p_i$ 之间的某个度量距离 $$ F P F H\left(\boldsymbol{p}_q\right)=S P F H\left(\boldsymbol{p}_q\right)+\frac{1}{k} \sum_{i=1}^k \frac{1}{\omega_i} \cdot S P F H\left(\boldsymbol{p}_i\right) $$ - FPFH vs PFH - FPFH 没有完全覆盖PFH的数据 - FPFH 还包括一些在r范围外的数据 - FPFH 时间复杂度比 PFH 小很多, 可以实现实时应用 - 下图中红色线表示第一次 $p_q$ 与其邻点的SPFH值, 随后其邻点也同样进行该操作, 表示为图中的五个圈, 随后 $p_q$ 再一次计算与其邻点的SPFH值, 但这次要同样考虑其邻点的SPFH值, ![image-20221024120609035](readme.assets/image-20221024120609035.png) - Pseudo Code ```pseudocode for each point p in cloud P step 1: 1. get the nearest neighbors of p 2. for each pair of p, p_i (where p_i is a neighbor of p, compute the three angular values 3. bin all the results in an output SPFH histogram step 2: 1. get the nearest neighbors of p 2. use each SPFH of p with a weighting scheme to assemble the FPFH p: ``` - check NAN/Infinite before **FPFHEstimation** - spedding FPFH with OpenMP ### 4.6 Viewpoint Feature Histogram (VFH) - VFH 是一种新颖的特征方法, 用于点云识别和 6Dof 姿态估计 - VFH 在 FPFH的基础上, 通过统计视角点与各点的角度, 和各点的法向量之间的角度差 ![image-20221024133740373](readme.assets/image-20221024133740373.png) ![image-20221024133723256](readme.assets/image-20221024133723256.png) ![image-20221024133823778](readme.assets/image-20221024133823778.png) ## 5 Filter example - PassThrough: - example: 01_filter_passThrough.cpp - 可以设置 filter limit negative - VoxelGrid: - example: 02_filter_voxelgrid.cpp - 降采样 - 设置采样格子大小 - StatisticalOutlierRemoval - example: 03_filter_StatisticalOutlierRemoval.cpp - 设置搜索范围K和标准差倍数, 在这个标准差外面的点会被标记为Outlier - 可以设置 filter negative, 保存 outlier - RadiusOutlierRemoval - example: 04_filter_RadiusOutlierRemoval.cpp - 半径搜索过滤， 通过设置半径搜索范围内的K个邻点数量来判断是否是outlier - ConditionalRemoval - example: 05_filter_ConditionalRemoval.cpp - 条件过滤，比较灵活地设置各种过滤条件 - BilateralFilter - example: 06_filter_BilateralFilter.cpp - 一种非线性滤波，可以达到保持边缘的同时对图像进行平滑的效果 - 时间复杂度较高，大的点云不适合 - 只适用于有序点云 - ProjectInliers - example: 07_filter_ProjectInliers.cpp - 对点云进行投影，示例中将点云投影到XY平面上 - ExtractIndices - exmaple: 08_filter_ExtractIndices.cpp - 通过segmentation获取对应的indices, 通过indices对点云进行分割 ## 6 Recognition ### 6.1 点云识别任务的特点 - 为每个点分配一个语义标记。点云的分类是将点云分类到不同的点云集。同一个点云集具有相似或相同的属性，例如地面、树木、人等。也叫做点云语义分割 - 有label数据 ### 6.2 3D Object Recognition based on Correspondence Grouping - 简介 - 使用 Correspondence Grouping algorithms (对应分组算法) 实现 3D Object Recognition - 可视化显示场景中可能存在的模型, 并得出其6DOF转换矩阵 - Code: - exmaple: 06-PCL_recognition/01_recognition_CorrespondenceGrouping.cpp - Demo流程 1. 解析命令, 加载场景点云与物体点云 2. 计算法线 - 获得 - 场景点云法线 - 物体点云法线 3. 通过 UniformSampling 获得点云关键点 - 均匀下采样 - 对点云数据创建一个三维体素栅格，然后，在每个体素保留一个最接近体素中心的点，代替体素中所有点 - 获得 - 场景 uniform sampling 关键点 - 物体 uniform sampling 关键点 4. 通过SHOTEstimationOMP计算关键点描述子 [SHOT 描述子](https://cloud.tencent.com/developer/article/1475995) - 获得 - 场景 SHOT352 描述子 - 物体 SHOT352 描述子 5. 通过 KdTree 和描述子寻找关键点之间的匹配关系 - 匹配距离小于 0.25, (匹配距离在0-1之间) - 保存匹配关系为 pcl::Correspondence 6. 通过Hough3D法进行点云间的配准 1. 通过 BOARDLocalReferenceFrameEstimation 计算 reference frame - 获得 - 场景 reference frame - 物体 reference frame 2. 通过 Hough3DGrouping 实现点云间的识别(配准), - 获得 - 物体在场景中的转换矩阵 - 聚类后的点云匹配关系 7. 可视化 - 物体在场景中转换后的点云 - 物体点云与场景的对应关系 ### 6.3 Implicit Shape Model 隐式形状模型 - [Implicit Shape Model](https://pcl.readthedocs.io/projects/tutorials/en/master/implicit_shape_model.html#implicit-shape-model) - 主要是通过训练数据学习到一类物体点云的特征, 并通过学习到的特征去**判断新的物体的点云中心** - 训练步骤 1. 对训练集点云 voxelgrid 降采样作为关键点 2. 对关键点提取FPFH特征 3. 对提取到的特征进行K-means聚类, 获得视觉特征词(visual word)字典(即聚类后的每一类特征), 每个特征都是对应一类特征的实例(instance) 4. 计算每个关键点到点云质心的距离 5. 针对每个视觉特征词(visual word), 计算其统计权重 - ![image-20221102160544130](readme.assets/image-20221102160544130.png) 6. 对每个特征点计算学习权重 - ![image-20221102161130381](readme.assets/image-20221102161130381.png) - 识别步骤 1. 获取关键点 2. 获取关键点的特征(FPFH) 3. 对每个关键点的特征, 寻找其最近的视觉特征词(visual word) 4. 对每个特征, 寻找到对应的视觉特征词后(visual word), 计算对应类别和视觉特征词的权重 - ![image-20221102162916523](readme.assets/image-20221102162916523.png) 5. 通过NMS分析第四部的结果 ### 6.4 Hypothesis Verification for 3D Object Recognition - 简介 - 验证杂乱和严重遮挡的 3D 场景中的模型假设 - 匹配描述子 - Correspondence Grouping - 确认各个物体假设的度量距离 - 减少False Positive 的可能 - [Hypothesis Verification](https://pcl.readthedocs.io/projects/tutorials/en/master/global_hypothesis_verification.html#global-hypothesis-verification) - ![image-20221102180307384](readme.assets/image-20221102180307384.png) ## 7 Registration ### 7.1 点云配准任务的特点 - 通过计算得到完美的坐标变换，将处于不同视角下的点云数据经过旋转平移等刚性变换统一整合到指定坐标系之下的过程 - 关键点匹配是点云配准的关键任务 - 点云配准一般流程 ![image-20221102181803187](readme.assets/image-20221102181803187.png) - 关键点提取 - NARF, SIFT, FAST - 关键点描述子提取 - NARF, FPFH, BRIEF, SIFT - 关键点匹配 - 点匹配 (使用点坐标) - brute force matching - kd-tree nearest neighbor search (FLANN) - searching in the image space of organized data - searching in the index space of organized data - 特征匹配 - brute force matching and - kd-tree nearest neighbor search (FLANN). - 两种特征匹配的策略 - Direct correspondence estimation - A 匹配 B - “Reciprocal” correspondence estimation - A 匹配 B 之后, 再通过他们的交集 B 匹配 A - 去除错误匹配 - RANSAC 随机一致性模型 - 使用一部分匹配的结果 - 估计转换矩阵 - 优化方法 - SVD - ICP - ... - Iterative Closest Point 流程 1. Search for correspondences. 2. Reject bad correspondences. 3. Estimate a transformation using the good correspondences. 4. Iterate. - Feature based registration 1. use SIFT Keypoints (pcl::SIFT…something) 2. use FPFH descriptors (pcl::FPFHEstimation) at the keypoints (see our tutorials for that, like http://www.pointclouds.org/media/rss2011.html) 3. get the FPFH descriptors and estimate correspondences using pcl::CorrespondenceEstimation 4. reject bad correspondences using one or many of the pcl::CorrespondenceRejectionXXX methods 5. finally get a transformation as mentioned above ### 7.2 iterative closest point example: 07-pcl_registration_ICP.cpp ### 7.3 How to incrementally register pairs of clouds example: 07-pcl_registration_IncrementRegist.cpp ![incremental_registration](readme.assets/incremental_registration.gif) ## 8 Segmentation ### 8.1 点云分割任务的特点 - 根据空间、几何和纹理等特征点进行划分，同一划分内的点云拥有相似的特征。点云分割的目的是分块，从而便于单独处理。 - 没有label数据?