Plane detection algorithm implementation from 3D point cloud

This project relies on OpenCV using C++ 11 as the programming language, a multi-plane detection algorithm of point cloud is realized. At the same time, a relatively effective algorithm optimization is proposed. It has a good performance in the accuracy and speed of multi-plane detection.

Author: Yechun Ruan, Wanli Zhong, Changzhen Zhang

Software Environment

OpenCV 4.5.1
g++ 5.4
Python 3.6 + Open3D Python version (optional，used for visualization)

Installation Tutorial: OpenCV, Open3D

Interface Introduction

The interface of 3D point cloud plane detection is:

/**
*
* @param labels  The label that the point belongs to a certain plane, n × 1 matrix, n is equal to the size of the input point cloud (output)
* @param planes  Holds the vector of plane equations, the equation is expressed as ax + by + cz + d = 0 (output)
* @param points3d  Input point cloud data
* @param thr  Threshold
* @param max_iterations  Maximum number of iterations
* @param desired_num_planes  Number of target planes
* @param grid_size  Downsampling grid size, if less than or equal to 0, it means no downsampling
* @param normal  Normal vector constraint, nullptr means no constraint is used, otherwise the detected plane normal vector satisfies the constraint
*/
void get_planes(cv::Mat &labels, std::vector<cv::Vec4f> &planes, cv::InputArray &points3d,
             float thr, int max_iterations, int desired_num_planes, float grid_size, cv::Vec3f *normal);

Explain in Detail:

labels: The parameter type is cv::Mat, n × 1 matrix, single channel, int type data, used to save the label of the point, 0 means not belonging to any plane, positive integer k means belonging to the number k plane
planes: The parameter type is std::vector<cv::Vec4f>, which is used to save the equations of the plane. [a, b, c, d] quaternion corresponds to ax + by + cz + d = 0
points3d: The parameter type is cv::Mat, n × 3 matrix, single channel, float type data, used to store point cloud data
thr: The parameter type is float, if the distance from the point to the plane is less than this value, the point is considered to belong to the inner point of the plane
max_iterations: The parameter type is int, the iteration will stop when the number of iterations reaches this value, and the plane equation with the most interior points will be output
desired_num_planes: The parameter type is int, this value represents the number of planes that you want to find from the point cloud
grid_size: The parameter type is float, the side length of the voxel filtering down-sampling grid, if it is less than or equal to 0, it means no down-sampling processing
normal: The parameter type is cv::Vec3f*, the normal vector of the plane in the three-dimensional space, nullptr means no constraint is used, otherwise the detected plane normal vector satisfies the constraint

Run Demo

Clone

git clone https://github.com/No-Plane-Cannot-Be-Detected/Plane_Detection

Compile

cd Plane_Detection

cmake .

make

Note: The above are the compilation steps for Linux operating system. If it is windows operating system, please modify the ninth line of the CMakeLists.txt file and set the OpenCV directory to the corresponding installation directory.

DEMO 1

./Point-Cloud-Plane-Detection 3 0.2 0.2 1000 ./data/check.ply 0 0 0

DEMO 2

./Point-Cloud-Plane-Detection 3 0.5 0.22 1000 ./data/Cassette_GT_.ply-sampling-0.2.ply 0 0 0

The incoming parameters are the number of target planes, the threshold, the grid size, the maximum number of iterations, the path of the point cloud file, and the normal vector constraint (0, 0, 0 means not using the normal vector constraint).

Point Cloud Visualization

Point cloud visualization can be achieved through Open3D (APP version, C++ version, Python version), PCL (C++ version, Python version), etc.

The Python version of the Open3D visualized point cloud sample code is in ./viz/Pointcloud-Visualization-With-Open3D.py

DEMO

python ./Pointcloud-Visualization-With-Open3D.py -cloud "../data/check.ply" -label "../data/check_label.txt"

Directory Structure

.
├── data (Data input and output directory)
│   ├── Cassette_GT_.ply-sampling-0.2.ply
│   └── check.ply
│   └── check_label.txt
├── images (Document picture directory)
├── include (Header file directory)
│   ├── ransac.h
│   └── utils.h
├── source (Source file directory)
│   ├── main.cpp
│   ├── ransac.cpp
│   └── utils.cpp
└── viz  (Visual sample code directory)
    └── Pointcloud-Visualization-With-Open3D.py

Dataset

Some datasets are from: IQmulus & TerraMobilita Contest. Download links: Zones 0-4, Cassette_idclass.zip.

Another part of the dataset is generated by standard plane equation with noise points.

More Details about our project

Technical Background and Scheme Design

Test overview and analysis

Radar real-time demonstration video of ZhiRen activity room in SUSTech

Project report in Chinese

Reference

R. Adams and L. Bischof. Seeded region growing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(6):641-647, 19 94.
Dorit Borrmann, Jan Elseberg, Kai Lingemann, and Andreas Nuchter. The 3d hough transform for plane detection in point clouds: A review and a new accumulator design. 3D Research, 0202, 06 2011.
Onduřej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Bernd Michaelis and Gerald Krell, editors, Pattern Recognition, pages 236-243, Berlin, Heidelberg, 2003. Springer Berlin Heidelberg.
M. Fischler and R. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981.

Contribution

Fork this warehouse
Create new Feat_xxx branch
Submit the code
New Pull Request

GengCauWong/Plane_Detection