Search

CN-121982315-A - Edge segmentation method and device for visual semantic map of intelligent robot with body

CN121982315ACN 121982315 ACN121982315 ACN 121982315ACN-121982315-A

Abstract

The invention discloses an edge segmentation method and device of visual semantic map of a body-equipped intelligent robot, wherein the method comprises the following steps of using 3D points of the current space Constructing a bounding box for the center, and judging the 3D point based on the bounding box Whether the voxels belong to points at the semantic boundary, if the voxels belong to the semantic boundary, carrying out voxelization on the bounding box to decompose the bounding box into N voxels, determining the semantics corresponding to each voxel, clustering based on the spatial information corresponding to each voxel obtained by the voxelization and the semantics, wherein the voxels which are adjacent in the clustering space and have the same semantic category are classified into a set, and judging the 3D points Whether the number of elements in the voxel set to which the voxel belongs meets a preset threshold value or not, and if so, the 3D point is obtained Semantics corresponding to belonging voxel set as 3D point Is defined by the meaning of (1). The semantic classification accuracy can be improved by detecting whether the 3D points in the space are in the semantic boundary range or not and then determining the semantics.

Inventors

  • YOU QINGZHEN

Assignees

  • 红象科技(北京)有限公司

Dates

Publication Date
20260505
Application Date
20260126

Claims (10)

  1. 1. An edge segmentation method of a visual semantic map of an intelligent robot with a body is characterized by comprising the following steps: in the current space 3D point Constructing a bounding box for the center, and judging the 3D point based on the bounding box Whether it belongs to a point at a semantic boundary; Determining the semantics corresponding to each voxel and clustering based on the space information corresponding to each voxel obtained by the voxel processing and the semantics, wherein voxels which are adjacent in the clustering space and have the same semantic category are classified into a set; Judging 3D point Whether the number of elements in the voxel set to which the voxel belongs meets a preset threshold value or not, and if so, the 3D point is obtained Semantics corresponding to belonging voxel set as 3D point Is defined by the meaning of (1).
  2. 2. The method for edge segmentation of a visual semantic map of an intelligent robot with body according to claim 1, further comprising: if 3D points If the number of elements in the voxel set to which the voxel belongs does not meet the preset threshold value, searching for a 3D point All non-3D points adjacent to the voxel at which they are located A clustering set of a voxel set to which the located voxel belongs; If the number of the cluster set elements meets a preset threshold, taking the semantics corresponding to the cluster combination as 3D points If not, continue to find the satisfied set.
  3. 3. The method for edge segmentation of a visual semantic map of an intelligent robot with body according to claim 1, wherein determining the semantics corresponding to each voxel comprises: and counting the semantic categories of all 3D points in each voxel and the number of points corresponding to the same semantic category, and taking the semantic category with the largest number of points as the semantic of the corresponding voxel.
  4. 4. The edge segmentation method of visual semantic map of intelligent robot with body according to claim 1, wherein 3D points are used as current space Building a bounding box for a center includes: Setting an initial height of the bounding box In 3D points A bounding box is constructed for the center, wherein, Representation of The mean distance of nearest neighbor m 3D points, cube_scale, represents the magnification of bounding box.
  5. 5. The edge segmentation method of a visual semantic map of an intelligent robot with body according to claim 4, wherein the 3D point is determined based on the bounding box Whether or not a point at a semantic boundary comprises Counting the number of points corresponding to each semantic category in the current bounding box, and based on the total number of points in the bounding box Number of points corresponding to each semantic category Judging whether to enlarge the bounding box; if the bounding box is judged to need to be enlarged, enlarging according to a preset rule, and continuously judging whether the bounding box needs to be enlarged again; If the dominant semantic category exists in the bounding box space and the expansion times reach K times, judging the 3D point Points that do not belong to semantic boundaries.
  6. 6. The edge segmentation method of a visual semantic map of an intelligent robot with body according to claim 1, wherein determining whether to enlarge the bounding box based on the total number of points in the bounding box and the number of points corresponding to each semantic category comprises: Setting a dominant semantic threshold thmain to determine whether to / ≥thmain; If it is greater than it is determined that it is necessary to enlarge the bounding box.
  7. 7. The edge segmentation method of a visual semantic map of an intelligent robot with body according to claim 1, wherein when it is required to enlarge a bounding box, the method comprises: Based on Is enlarged, wherein Is the height step size of each expansion of the bounding box space.
  8. 8. An edge segmentation device of a visual semantic map of an intelligent robot with a body is characterized by comprising: Semantic range judging unit for 3D point in current space Constructing a bounding box for the center, and judging the 3D point based on the bounding box Whether it belongs to a point at a semantic boundary; A clustering unit for processing the points belonging to the semantic boundary, and carrying out voxelization processing on the bounding box to decompose the bounding box into N voxels, determining the corresponding semantics of each voxel, and clustering based on the corresponding spatial information of each voxel obtained by the voxelization processing and the corresponding semantics, wherein voxels with the same semantic category and adjacent space in the clustering are classified into a set; Semantic determining unit for judging 3D point Whether the number of elements in the voxel set to which the voxel belongs meets a preset threshold value or not, and if so, the 3D point is obtained Semantics corresponding to belonging voxel set as 3D point Is defined by the meaning of (1).
  9. 9. A computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-7.
  10. 10. An electronic device comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores a computer program executable by the at least one processor to cause the at least one processor to perform the method of any of claims 1-7.

Description

Edge segmentation method and device for visual semantic map of intelligent robot with body Technical Field The invention relates to the technical field of semantic information processing, in particular to an edge segmentation method and device for a visual semantic map of an intelligent robot with a body. Background In the process of semantic segmentation, in particular to the situation of similar semantic scenes, the semantic discrimination usually has higher error rate, so that a visual semantic map generates a large number of noise points at the boundary, 3D points marked with the error semantics in the space cause the incorrect semantic map, directly influence the use of the subsequent semantic map, and therefore, a large amount of manual work is needed for checking, and the cost is high and the efficiency is low. Disclosure of Invention The invention mainly aims to provide an edge segmentation method and device for visual semantic map of an intelligent robot with a body, so as to solve the defects in the related art. To achieve the above object, according to a first aspect of the present invention, there is provided an edge segmentation method of a visual semantic map of an intelligent robot with body, including 3D points in a current spaceConstructing a bounding box for the center, and judging the 3D point based on the bounding boxWhether the voxels belong to points at the semantic boundary, if the voxels belong to the semantic boundary, carrying out voxelization on the bounding box to decompose the bounding box into N voxels, determining the semantics corresponding to each voxel, clustering based on the spatial information corresponding to each voxel obtained by the voxelization and the semantics, wherein the voxels which are adjacent in the clustering space and have the same semantic category are classified into a set, and judging the 3D pointsWhether the number of elements in the voxel set to which the voxel belongs meets a preset threshold value or not, and if so, the 3D point is obtainedSemantics corresponding to belonging voxel set as 3D pointIs defined by the meaning of (1). Optionally, the method further comprises, if the 3D pointIf the number of elements in the voxel set to which the voxel belongs does not meet the preset threshold value, searching for a 3D pointAll non-3D points adjacent to the voxel at which they are locatedIf the number of the cluster set elements meets a preset threshold value, the semantics corresponding to the cluster combination are used as 3D pointsIf not, continue to find the satisfied set. Optionally, determining the semantics corresponding to each voxel comprises counting the semantic categories of all 3D points in each voxel and the number of points corresponding to the same semantic category, and taking the semantic category with the largest number of points as the semantics of the corresponding voxel. Optionally, in the current spatial 3D pointBuilding a bounding box for a center includes setting an initial height of the bounding boxIn 3D pointsA bounding box is constructed for the center, wherein,Representation ofThe mean distance of nearest neighbor m 3D points, cube_scale, represents the magnification of bounding box. Optionally, the 3D point is determined based on the bounding boxWhether points belonging to the semantic boundary include counting the number of points corresponding to each semantic category in the current bounding box, based on the total number of points in the bounding boxNumber of points corresponding to each semantic categoryJudging whether to enlarge the bounding box, if so, enlarging the bounding box according to a preset rule, continuously judging whether to enlarge the bounding box again, and if the dominant semantic category exists in the bounding box space every time, judging the 3D point after the number of times of enlarging reaches K timesPoints that do not belong to semantic boundaries. Optionally, determining whether to expand the bounding box based on the total number of points in the bounding box and the number of points corresponding to each semantic category includes setting a dominant semantic threshold thmain, determining whether to/And if the size is larger than thmain, judging that the bounding box needs to be enlarged. Optionally, when it is desired to expand the bounding box, the method includes, based onIs enlarged, whereinIs the height step size of each expansion of the bounding box space. According to a second aspect of the present invention, there is provided an edge segmentation apparatus for a visual semantic map of an intelligent robot with a body, comprising a semantic range judging unit for 3D points in a current spaceConstructing a bounding box for the center, and judging the 3D point based on the bounding boxWhether the points belong to the semantic boundary, a clustering unit for processing the points belonging to the semantic boundary, carrying out voxelization processing on the bounding box to decompose th