Laser & Optoelectronics Progress, Volume. 56, Issue 4, 040002(2019)
3D Point Cloud Scene Data Acquisition and Its Key Technologies for Scene Understanding
Figures & Tables(12)
Fig. 1. Three-dimensional data acquisition systems. (a) Vehicle system; (b) fixed scanner; (c) trolley type acquisition system; (d) backpack type acquisition system
Fig. 2. Examples of scene understanding. (a) Semantic segmentation of image; (b) target detection of point cloud; (c) semantic segmentation of point cloud
Fig. 3. Examples of 2D images affected by environmental factors. (a) Severe occlusion of target; (b) semantic segmentation of (a); (c) target affected by light; (d) semantic segmentation of (c)
Fig. 8. Research status of point cloud feature extraction and point cloud segmentation
Table 1. Comparison of advantages and disadvantages of different data in scenario
View tableTable 1. Comparison of advantages and disadvantages of different data in scenario
Name Instance Advantage characteristic Disadvantage 2D RGB image 2D information, color textures, etc. Image is susceptible to illumination, lacks depth information, cannot directly acquire geometric information. Point cloud data based on image 3D reconstruction 3D scenes, color textures and other sparse point clouds, with depth information, distance. Point cloud is susceptible to light and environment, and reconstruction information is lost. Kinect-based RGB-D point cloud data 3D indoor small scene, dense point cloud, only for close range, with depth. Point cloud is susceptible to light, only for indoor scenes, small field of view. Point cloud data based on vehicle lidar High-precision, dense point clouds, and three-dimensional spatial information and intensity information. Less affected by environmental factors. Limited by platform, only point cloud scenes with linear road trajectories can be generated. Unable to extract scene color. information Point cloud data based on static lidar 3D outdoor scene, high-precision, dense point cloud, 3D spatial information and intensity. Collection device cannot be moved, and point cloud scene is incomplete. Point cloud data based on aerial lidar 3D outdoor scene, high-precision information, sparse point cloud, suitable for large-scale rough modeling. Unable to reproduce ground details, laser radar cannot extract scene color information. Collision-based panoramic image and laser point cloud fusion data 3D outdoor scenes, high-precision information, dense point clouds, mobile modeling, full-frame 3D space, color and intensity. Scene color information cannot be extracted by the laser point cloud alone.
Table 2. Comparison of noise filtering methods
View tableTable 2. Comparison of noise filtering methods
Name Principle Characteristic Through filter According to point cloud, threshold should be set in range of X ,Y , andZ axes, and then threshold isfiltered to remove points that are out of range. Fast, but not accurate enough, mostly used for rough processing in the first step. Statistical filtering Noise point is removed according to point density. The denser the points at certain area, the larger the amount of information. The noise information is useless, and amount of information is small. By calculating average distance of each point to its nearest k points,Gaussian distribution of distances of all points in point cloud is obtained to eliminate noise. The effect is better than the pass-through filter, which can accurately filter out noise points inside bounding box. Radius filtering Given a radius threshold, calculate the number of points in its radius for each point in the point cloud. When the number is greater than a given number of thresholds, point is retained, otherwise point is taken as a noise point and rejected. Filter out noise points inside bounding box at a faster rate.
Table 3. Comparison of ground filter methods
View tableTable 3. Comparison of ground filter methods
Name Principle Characteristic Elevation based filtering method According to point distribution filtering in point cloud, manually set or adaptively find z -direction threshold,and filter out point where z -value is less than thresholdin point cloud as ground. Fast, low robustness. Model based filtering method Select a model to fit ground (such as RANSAC-based planar model, CSF cloth simulation) and use fitted inner point as ground point. The algorithm is suitable for specific environments, and the robustness is poor, but the filtering effect is relatively good. Filtering method based on region growing Taking normal vector direction of point as criterion for regional growth, first adaptively find the point that is most likely to be the ground. Based on this, based on angular difference between normal vector direction of its neighborhood point and its normal vector direction, it is judged whether it grows or not. Continuous iteration to find all ground points. When the ground is not undulating, the algorithm can separate the ground well, but the time and space cost are relatively large. Method based on window movement Points distributed on ground should be of a continuous nature. Set a suitable window size to find the lowest point in current window. Then set threshold by the lowest point calculation model and filter out all points where elevation difference exceeds threshold. Faster, but the window size is too dependent on manual settings, and only local features are considered. Triangulation based filtering method The discrete points are connected according to a certain rule into a plurality of triangles covering entire area without overlapping each other to form an irregular triangular network. The sparse TIN is generated by the seed point, and slope of model is analyzed to perform initial segmentation, and slope is eliminated. Large triangular regions, and then through connectivity analysis to obtain features such as elevation differences for each segment. It avoids data redundancy when terrain is flat, but data structure is complex and space complexity is high.
Table 4. Common methods and comparison of point cloud feature descriptions
View tableTable 4. Common methods and comparison of point cloud feature descriptions
Name Type Principle Characteristic Spin image Local feature Count projection coordinates of all vertices to base plane to obtain descriptor. Resistance to rigid transformation and background interference; sensitive to density changes. 3DSC Local feature Counting number of points in different grids of spherical neighborhoods to obtain descriptors. Strong discrimination, anti-noise. PFH Local feature Parameterize the spatial difference between point and neighborhood and form a multi-bit histogram. Highly robust to point cloud density changes; high computational complexity. FPFH Local feature Recalculate K neighborhood by calculatingtuple of the query point and its neighbors compared to the PFH. Retains most of recognition capabilities of PFH, and reduces computational complexity compared to PFH. SHOT Local feature Spherical neighborhood rasterization, construct a histogram according to the angle of normal vector, and then concatenate histogram. Descriptive, anti-noise; large amount of calculation, sensitive to density changes. ESF Global feature Describe angle, distance, and triangle area of three random points in a point cloud. No need for pre-processing, strong feature description. VFH Global feature Add pilot direction to relative normal calculation after extending FPFH. Strongly discernible.
Tools
Get Citation
Copy Citation Text
Yong Li, Guofeng Tong, Jingchao Yang, Liqiang Zhang, Hao Peng, Huashuai Gao. 3D Point Cloud Scene Data Acquisition and Its Key Technologies for Scene Understanding[J]. Laser & Optoelectronics Progress, 2019, 56(4): 040002
Paper Information
Category: Reviews
Received: Aug. 6, 2018
Accepted: Sep. 6, 2018
Published Online: Jul. 31, 2019
The Author Email: Tong Guofeng (tongguofeng@ise.neu.edu.cn), Zhang Liqiang (zhanglq@bnu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20