Graph-based 3D point cloud object instance segmentation method in out-of-order scene

Abstract

The invention relates to a graph-based 3D point cloud object instance segmentation method in an out-of-order scene. The method comprises the following steps: step 1) carrying out downsampling and denoising on original scene point cloud data; 2) extracting key points of the scene point cloud, and forming a key point graph according to the neighbor relation between the key points; 3) completing coarse segmentation of the point cloud based on the key point graph; 4) constructing a direction bounding box based on the roughly segmented region, judging whether the region needs to be finely segmented according to the size of the direction bounding box, and extracting a complete target point cloud from the scene according to the direction bounding box; and in a fine segmentation stage, through combination of graph nodes and elimination of cut points of the graph, after mutually adhered part point clouds are segmented, extracting a complete target point cloud from a scene according to a direction bounding box. The method has the beneficial effect that the adhered parts can be separated more effectively.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method for segmenting 3D point cloud object instances in a graph-based out-of-order scene. Background technique [0002] With the continuous development of 3D data acquisition technology, machine vision and other technologies, 3D point cloud data has been widely used in industrial inspection, medical image processing, 3D reconstruction, robot intelligent vision operations, cultural relics protection and other fields. Point cloud object instance segmentation is an important part of point cloud data processing, that is, the point cloud dataset is divided into multiple regions, so that the points in each region only belong to the same object, and the extracted objects after segmentation can be used for object recognition, semantic understanding, targeting, etc. Due to the uneven sampling density and lack of clear structure of point cloud data, point cloud object instance...

AI Technical Summary

Problems solved by technology

The edge-based segmentation method is fast, but it is easily affected by noise and point cloud density, resulting in low segmentation accuracy; the point cloud segmentation algorithm based on model fitting is only suitable for processing objects with regular shapes, and its applicability in disordered scenes is relatively low. Poor; the region growing algorithm is sensitive to the selection of seed points, and the segmentation effect is greatly affected by the threshold of the region growing criterion; the Euclidean clustering algorithm is the most commonly used point cloud segmentation method in engineering. The distance is classified, but the distance threshold needs to be set in advance. If the threshold is too large or too small, it will lead to different degrees of under-segmentation and over-segmentation, which will affect the final segmentation effect.

Images