Summary of plane extraction method in point cloud

Real-Time Plane Segmentation using RGB-D Cameras

This paper is based on the plane detection of the depth image,Cluster the planes based on their depth neighborhood computed normal vectors and their distance from the origin.
The specific process is as follows:

1. Create 2 tangent vectors in the depth image x, y direction (calculate the tangent vector with the integral image) calculate the normal according to the 2 tangent vectors
2. For normal space $(n^x,n^y,n^z)$are projected to the spherical coordinate system for re-clustering to obtain candidate plane clusters
3. Cluster planes with similar local surface normals in distance space (distance of planes from origin)

View details

Fast Plane Detection and Polygonalization in noisy 3D Range Images

This paper performs plane extraction based on depth images, including 2 parts of plane fitting and polygonization.A linear method for computing the covariance matrix and MSE is proposed
Specific steps:

1. The point cloud randomly selects two points to form a point set
2. Find the point closest to this point set, and judge the MSE from the current point to the point set and the distance to the current point plane. Meet the requirements to join the current point set
3. The number of point sets greater than the threshold is valid
4. until all points are traversed

View details

Plane Detection in Point Cloud Data

The paper uses MDL to divide the point cloud into small blocks for RANSAC
Specific steps:

1. Divide the point cloud into different matrix blocks
2. In each matrix block, it is determined that there are 0-3 planes in each matrix block according to the RANSAC and MDL principles
3. Merging local-scale adjacent planes using region growing

View details

Fast and Accurate Plane Segmentation in Depth Maps for Indoor Scenes

A variety of plane extraction methods are proposed: based on depth extraction (DS), based on normal extraction (NS),Based on depth + normal blending (FS), based on the plane prior (PS),Based on RANSAC (RS)
Focus on FS and RS:
FS specific steps:

1. Calculate the gradient map of the image according to the depth
2. Calculate the gradient map of the image based on the normal
3. Combining the output of depth-based and normal-based methods for segmentation, post-fusion (watershed algorithm) can be applied to produce accurate and stable results

RS specific steps:

1. did not understand

View details

Robust Plane Detection Using Depth Information From a Consumer Depth Camera

A robust depth-based plane detection (DPD) algorithm is proposed, which consists of two parts: region-growing based plane detection and two-stage refinement.

Specific steps:
Valid Seed Patches Generation
uses a LxL sliding window to slide on the image, then uses the Linear Least Squares (LLS) method to fit a plane, and finally calculates based on the fitted plane and the information of all points in the sliding window the flatness of the sliding window area,

The Region Growing Process
sorts the sliding window regions obtained in the previous step by flatness. The higher the flatness, it will be used to initialize a plane and gradually grow around it. The growth is based on the distance from the adjacent point to the plane. At the same time Depending on the sensor measurement noise problem, some criteria are used to dynamically adjust this distance threshold (measurement distance, plane size, etc.).

The plane obtained by Refinement of Detected Planes after the above processing will have problems of Over-Growing and Under-Growing. For the former, it is necessary to detect the intersecting lines of the planes, and then separate the merged planes; for the latter, it is necessary to merge two planes

View details

A Fast and Accurate Plane Detection Algorithm for Large Noisy Point Clouds Using Filtered Normals and Voxel Growing

The normal of each point is calculated by filtered weighted plane fitting, and the local flatness score of each point is calculated. The best seed point for a good seed plane is chosen, and that seed plane is diffused by adding all points close to that plane.
Specific steps:

1. Calculate the local density of the point cloud for the next step to filter the normal
2. Compute the normal at each point by fitting a filtered weighted plane, and filter
3. Compute the score for each point's local plane
4. Calculate the octree shape of the point cloud using the octree class template in
5. Voxel-based region growth
   details view

Point-Plane SLAM for Hand-Held 3D Sensors

The specific steps of the plane detection part:

1. Several (20 in experiments) reference points are randomly selected in the 3D point cloud.
2. For each reference point, the best plane is found using nearby points within a small local window (101 × 101 pixels).
3. Find all inliers (threshold 20 mm) that are connected to the reference points of the grid map on the image space.
4. The best plane is determined to be the one with the largest and sufficient (>10000) interior points.
5. If the best plane is found, delete the interior point corresponding to it, and return to stage 1. Otherwise, terminate the algorithm

View details

Indoor Mapping Using Planes Extracted from Noisy RGB-D Sensors

A fast plane extraction method is proposed. Local normal vectors and edges are estimated using a local spherical approximation, and then the depth image is segmented by a flood-fill algorithm. To deal with sensor bias, we propose a new method for estimating the covariance of plane parameters. A fast plane boundary approximation method is proposed to achieve real-time rendering by reducing the amount of data.