Point Cloud Registration Pipeline

Artifact for paper Tigris: Architecture and Algorithms for 3D Perception in Point Clouds (MICRO'19): a general and flexible Point Cloud Registration pipeline built with Point Cloud Library (PCL).

Our implementation is excellent for development and evaluation of point-cloud registration algorithms as it is representative as well as highly configurable: components of the pipeline could be configured with various PCL kernels, or even your own implementations!

Pipeline Overview

Registration is a key building block in virtually all point cloud-based applications. It is a process that finds the transformation matrix that aligns two point cloud frames to form a globally consistent point cloud.

As shown in the figure above, our pipeline adopts a common two-phase design consisting of an initial estimation phase and a fine-tuning phase. The first phase performs an initial estimation of the transformation matrix, which is then fine-tuned in the second phase until the convergence.

Get Started (Before Running)

0. Prequisites:
   Make sure to install PCL and its dependencies correctly (the visualization module is not needed here). PCL 1.9.0 is preferred (we will use the shared objects only).

1. Project Settings:
   Three IMPORTANT specifications need to be made in the CMakeLists.txt in order to build the project successfully.
   

   A). set the path to PCLConfig.cmake correctly
   In the third line of the default CMakeLists.txt, set the path to PCLConfig.cmake correctly:
   set(PCL_DIR $PATH_TO_CONFIG_FILE)
   

   Below is an example on my machine: the path to the PCLConfig.cmake file is /u/txu17/local/pcl-1.9.0/share/pcl-1.9 (benefits of installing PCL under user space: more control of where and what to build, no need for sudo access).
   

   So in the CMakeLists.txt file, PCL_DIR is set to that path:
   (CMake has default paths to find PCLConfig.cmake, but they are not necessarily where PCLConfig.cmake is on your machine)).
   

   B). set the PCL version correctly
   Set the PCL version in the find_package line.

   C). set the included headers correctly
   In the include_directories("./pcl-1.7;$PATH_TO_EIGEN;" include) line, please make sure ./pcl-1.7 is always there, and change $PATH_TO_EIGEN to the path to EIGEN. Example when EIGEN is installed at /usr/include/eigen3:

2. Dataset:
   Two functions, i.e., loadTXTFile and loadBINFile, are provided to load point cloud data (the implementations are in src/utils.cpp):
   loadTXTFile is designed to load points in text files. An input file is considered to be a point cloud frame, in which each line should contain the coordinates of a point, formatted as: x_coord,y_corrd,z_corrd. Example below:
   

   loadBINFile is designed specifically for KITTI binary data. It also does the coordinate system transformation.
   

   Depending on the format of the input data, you could modify src/pc_pipeline.cpp to switch between these two functions.

3. Pipeline configuration:
   There are several ways to configure the registration pipeline, i.e., choosing kernels and specifying algorithm parameters of different stages:
   

   a). To use the configuration file
   By default, the configuration file is named 'config.txt' (in the bin directory). It specifies kernel parameters as well as some paths in the format of: $KEY: $VALUE. For example, this line in the configuration file:
   Normal_Search_Radius: 0.75
   sets the parameter Search Radius in the Normal Estimation stage to 0.75.
   
   Below is a complete list of parameters that could be set by the configuration file (which could be extended!):
   

   Parameter Name	Stage	Meaning	Values
   Normal_Search_Radius	Normal Estimation	Search radius	A floating point, e.g., 0.75
   Normal_Use_Customized_KDTree	Normal Estimation	Use customized KD Tree or not	{true, false}
   Normal_Max_Leaf_Size	Normal Estimation	Max leaf size for customized KD-Tree	An integer number, e.g., 32
   Key_Point_Detection_Module	Key-point Detection	Detection algorithm	{NARF, SIFT, HARRIS}
   Feature_Search_Radius	Descriptor Calculation	Search radius	A floating point, e.g., 0.85
   Feature_Module	Descriptor Calculation	Feature descriptor type	{SHOT, FPFH}
   Corr_Est_Use_Reciprocal_Search	Key-Point Correspondence Estimation	To do reciprocal search or not	{true, false}
   Ransac_Threshold	Correspondence Rejection	Distance threshold	A floating point, e.g., 0.20
   Ransac_Max_Iteration	Correspondence Rejection	Maximum iteration number	An integer, e.g., 10000
   ICP_Solver	Fine-Tuning (ICP)	Solver type	{SVD, LM}
   ICP_Max_Iteration	Fine-Tuning (ICP)	Maximum iteration number	An integer, e.g., 15
   ICP_Use_Ransac	Fine-Tuning (ICP)	To do rejection in each iteration or not	{true, false}
   ICP_Use_Reciprocal_Search	Fine-Tuning (ICP)	To do reciprocal search in each iteration or not	{true, false}
   ICP_Transformation_Epsilon	Fine-Tuning (ICP)	Transformation matrix epsilon	A floating point, e.g., 1e-9
   ICP_Max_Correspondence_Distance	Fine-Tuning (ICP)	Correspondences with higher distances will be ignored	A floating point, e.g., 1.20
   ICP_Euclidean_Fitness_Epsilon	Fine-Tuning (ICP)	Euclidean distance difference epsilon	A floating point, e.g., 1e-6
   ICP_Ransac_Outlier_Rejection_Threshold	Fine-Tuning (ICP)		An integer number, e.g., 120
   ICP_Use_Customized_KDTree	Fine-Tuning (ICP)	Use customized KD-Tree or not	{true, false}
   ICP_Max_Leaf_Size	Fine-Tuning (ICP)	Max leaf size for customized KD-Tree	An integer number, e.g., 32
   Approx_Radius_Search_Para	Normal Estimation	Threshold for approximated radius search	A floating point under 1.0, e.g., 0.20
   Approx_Nearest_Search_Para	Fine-Tuning (ICP)	Threshold for approximated nearest search	A floating point, e.g., 1.20
   Save_Approx_Data	Normal Estimation / Fine-Tuning (ICP)	Record data for approximation search or not	{true, false}

   b). To modify the source code

   • change to different PCL kernels: ICP example.
   • add implementations.

Build & Run

1. Switch to the project directory (where the Makefile is) if you are not, and make. By default, if everything set correctly, it will build the project and generate the executable in the bin directory.

2. Enter the bin directory, and set the config.txt. Especially, please set the path to dataset correctly as the example below (suppose the data is at /data/pointCloud/):

3. Make sure the configurations are all set, then execute ./pc_pipeline (default name). Hopefully the output will be similar to what's shown below:
   

++ Switch to the kdtree_module directory to take a look at our implementation (mainly based on FLANN's) of KD-Tree construction and search. If the parameter Normal_Use_Customized_KDTree is set to be True, the pipeline would use our version of KD-Tree search in the Normal Estimation Stage. ICP_Use_Customized_KDTree works samely.

Publication

This project implements the configurable point cloud registration pipeline described in the following paper:

T. Xu*, B. Tian*, and Y. Zhu, "Tigris: Architecture and Algorithms for 3D Perception in Point Clouds", In Proc. of MICRO, 2019. (* co-primary authors)

Please kindly consider citing this paper in your publications if it helps your research.

@inproceedings{xu2019tigris,
  title={Tigris: Architecture and Algorithms for 3D Perception in Point Clouds},
  author={Xu, Tiancheng and Tian, Boyuan and Zhu, Yuhao},
  booktitle={Proceedings of the 52th International Symposium on Microarchitecture},
  year={2019},
  organization={ACM}
}

Acknowledgement