Automated Indoor Pedestrian Networks
Reconstruction from As-Built Floorplan
Drawings using LLM and YOLO
Longyong Wu, Fan Xue
Department of Real Estate and Construction Management, The University of Hong Kong, Hong Kong, China

Introduction
Pedestrian networks play a crucial role in pedestrian navigation. However conventional 3D indoor mapping methods for
reconstructing indoor networks are costly, involving intensive labor work and expensive equipment. This paper proposes an
automated pipeline to reconstruct pedestrian networks from 2D as-built floorplan drawings (or 3D BIMs). Pilot tests on the real-world
drawings in Hong Kong proved the pipeline achieved accurate indoor space detection (mIoU =87.24% ) by retraining YOLO.

Methods

STEP 3: Geo-referenced Coordinates Assignment [3]
iB1000
Estimated

2

1. At the Client layer, users interact with the system through
third-party software like Revit, AutoCAD, or ArcGIS Pro.
2. The Server layer handles compute-intensive tasks using
graphics processing units (GPUs).
3. The Data Storage layer stores reconstruction histories, the
resulting 3D indoor maps, and other related information.

argmin𝑅𝑅,𝑡𝑡 ||𝑋𝑋 − 𝑅𝑅𝑅𝑅 − 𝑡𝑡|| , s. t

𝑅𝑅 𝑇𝑇 𝑅𝑅 = 𝐼𝐼

STEP 4: Pedestrian Networks Creation

STEP 1: Drawings Classification

STRUCTURE

SECTION

ELEVATION

1. Streamline the data formats of 2D vector drawings or 3D
BIMs as high-resolution 2D raster images
2. Classify each image into different categories using a Large
Language Model (LLM) , ChatGPT 4o [1].

STEP 2: Structural Elements Detection
Wall

Window

Stair

Door

Lift

Indoor Space

Create indoor pedestrian networks from simplified skeletons of
room spaces against fixed points of structural elements [4].

11111111123
Area of Overlap
IoU =
Results
Area of Union 123
1. Type: Raster (made after Year 2000)
2. Processing time: 25.77s (predict) + 128.57s (post-process) +
3.51s (map + path)
3. Indoor Space IoU (Intersection over Union): 87.24%

Conclusion
The findings of this paper demonstrate that the automated
pipeline, taking advantage of the power of LLM and retraining a
well-known deep learning model, is efficient and cost-effective
for multidimensional information extraction and precision
mapping, supporting smart city development in Hong Kong and
beyond.

Acknowledgment
Conclusion
Thanks
to Miss Vanessa and Mr. Alvin from Survey and Mapping
Office for the data.

References

Retrain the YOLO (V8) [2], a well-known pre-trained deep model
for 2D object detection, to detect key structural elements, such
as walls, doors, openings, staircases, and elevators, from the
images of structural floorplans.

[1] GPT-4o. https://openai.com/index/hello-gpt-4o/
[2] Jocher, G., Qiu, J., & Chaurasia, A. (2023). Ultralytics YOLO
(Version 8.0.0) [Computer software].
https://github.com/ultralytics/ultralytics
[3] Myronenko, A., & Song, X. (2010). Point set registration:
Coherent point drift. IEEE transactions on pattern analysis and
machine intelligence, 32(12), 2262-2275.
[4] CGAL, Computational Geometry Algorithms

Library, https://www.cgal.org

首屆智慧遙感製圖國際研討會

The 1st International Workshop on
Remote Sensing Intelligent Mapping (RSIM)