Three-dimensional point cloud outlier elimination method based on image segmentation

Abstract

The invention discloses a three-dimensional point cloud outlier elimination method based on image segmentation, which belongs to the field of computer graphics, and comprises the following steps: setting a sampling interval, and performing uniformly sampling in a unit sphere parameter space to generate a projection direction; solving a transformation matrix according to a 313 rotation relationshipbetween the generated projection direction and a Z axis of a world coordinate system where the three-dimensional point cloud is located, and performing attitude transformation on the point cloud by utilizing the transformation matrix; calculating the image resolution of the three-dimensional point cloud after attitude transformation projected to the perspective projection virtual view; obtainingperspective projection virtual views of the three-dimensional point cloud in all projection directions; segmenting a main body part of the image in the obtained perspective projection virtual view byutilizing a main body extraction algorithm based on image segmentation; according to a visible shell technology, forming a convex hull of a three-dimensional point cloud by utilizing a side shadow contour line of a main body part in a perspective projection virtual view, and taking three-dimensional points except the convex hull of the three-dimensional point cloud as outliers to be eliminated.

Description

technical field [0001] The invention relates to the field of computer graphics, in particular to a method for removing outliers in a three-dimensional point cloud based on image segmentation. Background technique [0002] The 3D point cloud obtained from a laser scanner or camera usually contains a large number of outliers due to factors such as illumination, calculation errors, and equipment errors, and the existence of these outliers will greatly affect the accuracy of subsequent 3D point cloud processing. Therefore, before the subsequent 3D point cloud processing, the outlier points in the 3D point cloud need to be eliminated first. [0003] In the field of computer graphics, the preprocessing of 3D point clouds has been receiving extensive attention. The main issues involved in the preprocessing steps of 3D point cloud are the elimination of outliers, the removal of noise, and the preservation of point cloud features. However, there is no uniform definition of outliers...

AI Technical Summary

Problems solved by technology

However, these methods have the following defects: (1) The methods based on geometry and statistics generally count the distance between points and points in their local neighborhood, so as to obtain statistics such as mean and variance, etc., and set A threshold. Once the distance between a point and its neighbors is greater than the threshold, it will be removed. Due to the short distance between dense outliers, it is impossible to use statistical information to distinguish dense outliers from effective 3D point clouds. , so the method based on geometry and statistics cannot deal with dense outliers; (2) The method based on depth map is to use a series of multi-view high-resolution images acquired by the camera as input, estimate the depth map of each view, and Use constraints such as geometric consistency, visibility, color consistency, and illumination consistency in multiple views to eliminate abnormal data points. However, this method needs to obtain real images of the target object in multiple views and has a large amount of calculation. The problem

The method based on information entropy does not consider the semantic information of the feature, so the method cannot generate the optimal viewpoint position for the semantic feature; the method based on the feature has a variety of features to choose from, but different features are applicable to different occasions, which is The selection of features depends on the application scenario; learning-based methods require a large amount of calibration data for training, and these calibration data require a large amount of human interaction

Images