Point cloud data processing

The research contents of 3D computational vision include:

(1) 3D matching: the matching between two or more frames of point cloud data, because the laser scanning beam is blocked by the object, it is impossible to obtain the 3D point cloud of the whole object through one scanning. Therefore, it is necessary to scan the object from different positions and angles. The purpose of 3-D matching is to splice the adjacent scanning point cloud data together. Three-dimensional matching focuses on matching algorithms, and commonly used algorithms include nearest point iteration algorithm ICP and various global matching algorithms.

(2) Multi-view 3D reconstruction: In computer vision, multi-view generally uses image information and considers some constraints of multi-view geometry, and the related research is very popular at present. Projective geometry and multi-view geometry are the basis of visual methods. In photogrammetry, there are similar linear equations and beam adjustment. Multi-view matching of point clouds is also put here, such as three-dimensional reconstruction of human body. Multi-view reconstruction of point cloud not only emphasizes frame-by-frame matching, but also needs to consider the error accumulation caused by observation from different angles, so there is also an optimization or adjustment process. It is usually realized by observing and forming a closed loop to make overall adjustment, and multi-view reconstruction emphasizes overall optimization. You can use only images or point clouds, or both (depth images). The result of reconstruction is usually a mesh.

(3)3D SLAM: point cloud matching (nearest point iterative algorithm ICP, normal distribution transformation method NDT)+ attitude map optimization (g2o, LUM, ELCH, Toro, SPA); Real-time 3D SLAM algorithm (LOAM); Kalman filtering method. 3D SLAM usually produces 3D point clouds, or octree diagrams. SLAM is based on vision (monocular, binocular, fisheye camera, depth camera) methods, such as orbSLAM and lsdSLAM. ...

(4) Target recognition: pedestrians, cars, bicycles, roads and road ancillary facilities (street trees, street lamps, zebra crossings, etc.). ) is based on laser data detection.

(5) Shape detection and classification: Point cloud technology is widely used in reverse engineering. After constructing a large number of geometric models, how to manage and retrieve them effectively is a very difficult problem. Point cloud (grid) model needs feature description and classification. According to the feature information of the model, the model is retrieved. At the same time, it includes how to retrieve a specific object from the scene, which focuses on the model.

(6) Semantic classification: After obtaining the point cloud of the scene, it is necessary to classify the point cloud by how to effectively use the point cloud information and how to understand the content of the point cloud scene, and each point cloud needs to be labeled. It can be divided into point-based method and segmentation-based classification method. From the method, it can be divided into supervised classification-based technology or unsupervised classification technology, and deep learning is also a promising technology.

(7) Stereo vision and stereo matching? ZNCC

(8)SFM (Motion Recovery Structure)

1, point cloud filtering method (data preprocessing):

Bilateral filtering, Gaussian filtering, conditional filtering, straight-through filtering, random sampling consistency filtering.

Voxel mesh

2. Main points

ISS3D, Harris 3D, NARF

SIFT3D、

3. Function and function description

Normal and curvature calculation, normal estimation, eigenvalue analysis, EGI.

PFH, FPFH, 3D shape background, spin image.

4. Point cloud matching

International comparison scheme, robust international comparison scheme, point-to-point international comparison scheme, point-to-line international comparison scheme, MBICP, GICP.

Three-dimensional nondestructive testing, multi-layer nondestructive testing

FPCS、KFPCS、SAC-IA

Line segment matching, ICL

5. Point cloud segmentation and classification

Segmentation: region growth, Ransac line and surface extraction, NDT-RANSAC,

K-Means, normalized cut (based on context)

Keywords 3D Hough transform (line and surface extraction), connectivity analysis,

Classification: point-based classification, segmentation-based classification; Supervised classification and unsupervised classification

6.SLAM graph optimization

g2o、LUM、ELCH、Toro、SPA

SLAM method: ICP, MBICP, IDC, likehood Field, Cross? Correlation and nondestructive testing

7. Target recognition and retrieval

Hausdorff distance calculation (face recognition)

8. Change detection

Change detection based on octree

9. 3D reconstruction

Poisson reconstruction, Delaunay triangulation

Surface reconstruction, human reconstruction, building reconstruction, tree reconstruction.

Real-time reconstruction: reconstructing the 4D(3D+ time) growth trend of vegetation or crops; Human posture recognition; Expression recognition;

10. point cloud data management

Point cloud compression, point cloud index (KD, octree), point cloud LOD (pyramid), massive point cloud rendering.

The Main Research and Application of Computer Graphics Driven by Point Cloud

/research/20 15/Points/