Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Line-Based Room Segmentation and Continuous Euclidean Distance Fields

Erik Warberg, Adam Miksits, Fernando S. Barbosa

TL;DR

This work presents a real-time, line-based approach to indoor mapping that combines per-room Gaussian Process Euclidean Distance Fields with line-segment priors. By constructing a visibility graph from depth-derived line segments and applying spectral clustering, the method identifies rooms and maintains lightweight, room-specific GP-EDFs that update online and incorporate residual measurements via adaptive inducing points. The key contributions are the line-based room segmentation pipeline, the use of line segments as distance priors in GP-EDFs, and the demonstrated scalability and real-time performance compared to global GP baselines. The approach enables scalable, continuous maps suitable for larger indoor environments and provides open-source code for broader adoption and extension to 3D and dynamic settings.

Abstract

Continuous maps representations, as opposed to traditional discrete ones such as grid maps, have been gaining traction in the research community. However, current approaches still suffer from high computation costs, making them unable to be used in large environments without sacrificing precision. In this paper, a scalable method building upon Gaussian Process-based Euclidean Distance Fields (GP-EDFs) is proposed. By leveraging structure inherent to indoor environments, namely walls and rooms, we achieve an accurate continuous map representation that is fast enough to be updated and used in real-time. This is possible thanks to a novel line-based room segmentation algorithm, enabling the creation of smaller local GP-EDFs for each room, which in turn also use line segments as its shape priors, thus representing the map more efficiently with fewer data points. We evaluate this method in simulation experiments, and make the code available open-source.

Real-Time Line-Based Room Segmentation and Continuous Euclidean Distance Fields

TL;DR

This work presents a real-time, line-based approach to indoor mapping that combines per-room Gaussian Process Euclidean Distance Fields with line-segment priors. By constructing a visibility graph from depth-derived line segments and applying spectral clustering, the method identifies rooms and maintains lightweight, room-specific GP-EDFs that update online and incorporate residual measurements via adaptive inducing points. The key contributions are the line-based room segmentation pipeline, the use of line segments as distance priors in GP-EDFs, and the demonstrated scalability and real-time performance compared to global GP baselines. The approach enables scalable, continuous maps suitable for larger indoor environments and provides open-source code for broader adoption and extension to 3D and dynamic settings.

Abstract

Continuous maps representations, as opposed to traditional discrete ones such as grid maps, have been gaining traction in the research community. However, current approaches still suffer from high computation costs, making them unable to be used in large environments without sacrificing precision. In this paper, a scalable method building upon Gaussian Process-based Euclidean Distance Fields (GP-EDFs) is proposed. By leveraging structure inherent to indoor environments, namely walls and rooms, we achieve an accurate continuous map representation that is fast enough to be updated and used in real-time. This is possible thanks to a novel line-based room segmentation algorithm, enabling the creation of smaller local GP-EDFs for each room, which in turn also use line segments as its shape priors, thus representing the map more efficiently with fewer data points. We evaluate this method in simulation experiments, and make the code available open-source.
Paper Structure (18 sections, 5 equations, 4 figures)

This paper contains 18 sections, 5 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Contribution
  4. Outline
  5. Room Segmentation
  6. Line Segment Processing
  7. Connecting Line Segments to Create Corners
  8. Splitting Line Segments to Create Corners
  9. Splitting Line Segments at Doorways
  10. Visibility Graph Construction
  11. Graph Clustering
  12. Room Identification
  13. Room-Based GP-EDF
  14. Line Segments as Distance Priors
  15. Incorporating Residual Measurement Points
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Result of the proposed real-time line-based room segmentation algorithm applied to a selected use case. Each room is depicted in a different color, with its line segments presented in the same color. Black dots represent obstacles detected using a 2D Lidar sensor. Dark red dots represent each room in the connectivity graph, with red lines as edges connecting neighboring rooms.
  • Figure 2: Visualization of the visibility graph, with nodes represented as blue line segments and edges as red line segments.
  • Figure 3: This figure shows a few snapshots of the room segmentation and the room-based GP-EDF being created simultaneously in real time. At the top-left corner is the very first frame after the initial sensor reading, and progress is shown clockwise. The bright red dot represents the position of the robot in the given frame, and black dots are obstacles sensed with a 2D Lidar sensor. The color of each line segment overlapping obstacle regions is represents the room to which they belong to. Colorful level curves depict the GP-EDF of the room the robot is inside of.
  • Figure 4: Graphs depicting changes in computation time for all GP-EDF models.