💻 computer science

Passage-Aware Structural Mapping for RGB-D Visual SLAM

This paper introduces a passage-aware structural mapping approach for RGB-D Visual SLAM that detects and classifies doors and traversable openings by fusing geometric, semantic, and topological cues, thereby enhancing room connectivity modeling within the vS-Graphs framework.

Original authors: Ali Tourani, Miguel Fernandez-Cortizas, Saad Ejaz, David Pérez Saura, Asier Bikandi-Noya, Jose Luis Sanchez-Lopez, Holger Voos

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali Tourani, Miguel Fernandez-Cortizas, Saad Ejaz, David Pérez Saura, Asier Bikandi-Noya, Jose Luis Sanchez-Lopez, Holger Voos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot trying to navigate a house. Most robots today are like people walking through a room with their eyes closed; they can feel the walls and know where the corners are, but they don't really "understand" what those walls are made of or where the doors are. They see a flat surface and assume it's solid all the way through.

This paper introduces a new way for robots to "see" the house more clearly, specifically looking for doorways and passages. Here is how the authors solved this problem, using simple analogies:

The Problem: The "Solid Wall" Illusion

Current robot mapping systems are great at drawing the outline of a room (the walls and the floor). However, they often treat a wall with a door in it the same as a solid wall. To the robot, a closed door and a solid brick wall look identical: just a flat, unpassable surface. This makes it hard for the robot to plan a route from the kitchen to the bedroom because it doesn't know a door exists.

The Solution: A Three-Part Detective Team

The authors created a system that acts like a detective team, using three different types of clues to find doors and openings. They call this "Passage-Aware Structural Mapping."

1. The "Flatness" Clue (Geometric & Semantic)
Think of a door as a picture frame hanging on a wall.

The Clue: The robot first identifies the wall. Then, it looks for a "door" object sitting right on top of that wall.
The Logic: If the robot sees a door object that is perfectly flat and aligned with the wall (like a painting), it knows the door is closed. The robot marks this spot as a "doorway," but notes that it is currently blocked. It's like seeing a closed gate; you know the path exists, but you can't walk through it yet.

2. The "Ghost Walk" Clue (Traversal Evidence)
Imagine you are walking through a house, and suddenly, the wall you were walking next to disappears, and you find yourself in a different room.

The Clue: The robot watches its own movement. If it sees a wall, and then in the very next moment, it sees the other side of that same wall, it knows it must have passed through an opening.
The Logic: If the robot's path crosses from one side of a wall to the other without crashing, it proves there is a hole in the wall. This is strong evidence of a doorway, even if the robot didn't explicitly "see" the door frame.

3. The "Puzzle Gap" Clue (Geometric Validation)
Imagine building a wall out of LEGO bricks. If you leave a gap in the middle, the wall has a hole.

The Clue: The robot looks at the 3D map it has built of the wall. It scans for "gaps" or missing pieces where the wall should be solid but isn't.
The Logic: Finding a gap is a good hint, but it's not perfect. A gap could be a door, a window, or just a poster on the floor. To be sure, the robot checks the size and shape of the gap. If it looks like a door, it marks it as a passage. If it's ambiguous, it waits for more clues (like the "Ghost Walk" clue above) to confirm.

Putting It All Together

The researchers plugged this new "detective team" into an existing robot brain called vS-Graphs.

Before: The robot's map was just a collection of rooms separated by solid lines (walls).
After: The robot's map now includes "gateways." It knows, "This is a wall, but there is a door here, and it is currently closed," or "There is an opening here that leads to the next room."

The Results

The team tested this in real office environments.

Success: The robot successfully found doorways. It could tell the difference between an open path (a doorway you can walk through) and a blocked path (a closed door or a solid wall).
Visuals: In their tests, they used purple markers to show open passages and red clouds to show closed doors, proving the robot could distinguish between the two.

The Future Connection (BIM)

The paper also mentions a future upgrade involving BIM (Building Information Modeling). Think of BIM as the architect's original blueprint of the house.

The authors suggest that in the future, the robot could compare what it sees (the "as-built" reality) with the blueprint (the "as-planned" design).
If the robot thinks it found a door, but the blueprint says there shouldn't be one, the system can double-check. This helps the robot avoid mistakes and understand the building's structure even more reliably.

In short: This paper teaches robots to stop seeing walls as solid barriers and start seeing them as structures that might have doors, windows, or openings, making them much better at navigating indoor spaces.

1. Problem Statement

Modern Visual SLAM (VSLAM) systems have advanced significantly in geometric reconstruction and semantic mapping (e.g., identifying walls, floors, and furniture). However, a critical gap remains in the detection and modeling of passages (doorways, archways, and traversable openings).

Limitations of Current Methods: Existing frameworks often treat walls as continuous surfaces, failing to distinguish between a solid wall and an opening. Some approaches rely on fiducial markers (e.g., AprilTags on doors), which limits scalability and practicality. Others focus on dynamic objects rather than structural connectivity.
The Challenge: Passages are topologically significant landmarks that define connectivity between navigable spaces. Detecting them is difficult because purely geometric cues (gaps in point clouds) are ambiguous; a gap could be a door, a window, a cabinet, or a poster. Furthermore, distinguishing between an open passage and a closed door requires understanding the relationship between the door object and the supporting wall.

2. Methodology

The authors propose a Passage-Aware Structural Mapping approach that fuses geometric, semantic, and topological cues to detect doors and passages. The method is integrated into the vS-Graphs framework.

A. Core Formalism

The system operates on RGB-D KeyFrames ( $K_t$ ) containing point clouds ( $P_t$ ) and images ( $L_t$ ).

Semantic Segmentation: A panoptic segmentation model (e.g., YOSO) extracts semantic labels and instance masks, projecting them onto the 3D point cloud to create semantically segmented subsets (e.g., $P_{wall}$ , $P_{door}$ ).
Planar Fitting: RANSAC is used to fit planes to these subsets, estimating normal vectors ( $n$ ) and offsets ( $d$ ) for walls and doors.

B. Door Modeling & Classification

Doors are modeled as planar entities embedded within walls. Their traversability is determined by their geometric relationship to the supporting wall:

Closed/Non-traversable: If the detected door plane is approximately coplanar with the supporting wall plane (satisfying angular alignment and offset thresholds), it is classified as a closed door.
Open/Traversable: If the coplanarity condition is not met, the door is inferred to be open.

C. Passage Detection Strategies

The system identifies passages using two complementary strategies to handle both explicit door instances and generic openings:

Traversal Evidence (Topological/Geometric):
- Analyzes the camera trajectory relative to mapped wall planes across consecutive KeyFrames.
- A wall traversal event is detected when the signed distance of the camera center to the wall plane changes sign ( $s_t \cdot s_{t+1} < 0$ ) within a bounded distance.
- This indicates the camera physically crossed the wall plane, implying a geometric opening.
- If a door instance is nearby, the passage is classified as a doorway; otherwise, it is a generic opening with default dimensions.
Geometric Opening Validation (Global Analysis):
- Projects the reconstructed wall geometry onto a 2D parameterization.
- Identifies gaps (regions with insufficient point density) as potential openings.
- Contextual Validation: To avoid false positives (e.g., gaps caused by furniture or posters), the system validates these gaps based on geometric consistency (size, aspect ratio) and proximity to detected door instances.

3. Key Contributions

Novel Formulation: Defines passage detection as the identification of traversable openings by jointly fusing geometric (plane fitting, gaps), semantic (door/wall classification), and topological (camera traversal) cues.
System Integration: Successfully integrates this logic into vS-Graphs, a real-time VSLAM pipeline. This enriches the scene graph with "passage-level" abstractions, improving room connectivity modeling without sacrificing real-time performance.
Open Source: The authors have released the source code to facilitate reproducibility and further research.
BIM Compatibility: The framework is designed to be extended with Building Information Modeling (BIM) priors (via ivS-Graphs) to validate detected passages against as-planned architectural models.

4. Results & Evaluation

Dataset: Evaluated on indoor office sequences recorded with the SMapper device.
Performance: The system demonstrated reliable detection of doorways and passages.
- Visual Output: The results (Fig. 4) show distinct markers: light purple for open passages and crimson point clouds for closed/obstructed doorways.
- Robustness: The method successfully distinguished between open and closed states and handled varying spatial layouts and environmental complexities.
Qualitative Analysis: The approach effectively reduced ambiguity in gap detection by combining traversal evidence with semantic context.

5. Significance & Future Work

Enhanced Situational Awareness: By explicitly modeling passages, robots gain a deeper understanding of space connectivity, which is crucial for path planning, navigation, and exploration in complex indoor environments.
BIM Integration: The paper highlights a pathway toward BIM-informed VSLAM. By aligning reconstructed geometry with architectural plans, the system can validate passages, detect structural deviations (e.g., missing walls or misaligned doors), and improve the structural consistency of the map.
Future Directions: The authors plan to incorporate doors and passages directly into the factor graph optimization for tighter coupling between traversability reasoning and pose estimation, and to conduct quantitative benchmarking across diverse environments.

In summary, this paper bridges the gap between low-level geometric mapping and high-level structural understanding, providing a robust framework for robots to navigate indoor spaces by "seeing" and understanding doorways and passages.