Indoor Mapping Using Machine Learning Based Classification of 3D Point Clouds - Research

Aug 4, 2026

Upload by Riley

Indoor maps provide essential spatial information for navigation automation, virtual reality, and robot manipulation. Accurate classification of 3D point clouds from indoor environments is therefore a key step in indoor mapping workflows. Using the S3DIS dataset, this study evaluates Random Forest, XGBoost, MLP, and TabNet for point-cloud class prediction, ranking performance by overall accuracy and F1 scores. Results indicate that machine learning models can effectively support indoor mapping through reliable semantic labeling.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Introduction

Indoor mapping data acquisition

Problem statement and motivation

Methodology

Preprocessing (input and normalization)

Machine learning classification

Indoor Mapping Using Machine Learning Based Classification of 3D Point Clouds Alper Sen*, Atakan Bilgili* * Department of Geomatic Engineering, Faculty of Civil Engineering, Yildiz Technical University, Esenler, Istanbul 34220, Türkiye [alpersen@yildiz.edu.tr](alpersen@yildiz.edu.tr), [atakanb@yildiz.edu.tr](atakanb@yildiz.edu.tr) Abstract. Today, indoor maps remain a valuable source of spatial information for various indoor environments. Classifying 3D point clouds from indoor environments is crucial for indoor mapping. In this study, indoor point clouds from the S3DIS dataset were classified using Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Multi-Layer Perceptron (MLP), and Attentive Interpretable Tabular Learning (TabNet) . The classification performances, based on overall accuracy and F1 scores, can be ranked as RF, MLP, XGBoost, and TabNet. It has been determined that machine learning algorithms can be used to classify indoor point clouds for indoor mapping. Keywords. Indoor mapping, machine learning, point cloud 1. Introduction The automatic generation of high-quality indoor maps for existing buildings poses a significant challenge within navigation automation, virtual reality, and robot object manipulation (Lin et al 2021) . Since the as-built condition of the buildings often deviates from the original plans due to renovations, indoor mapping for existing buildings has garnered extensive research attention.Indoor maps can be considered as the output of indoor measures and serve as the foundation for most indoor-based applications. The complexity of buildings and the increasing use of indoor positioning systems also provides a strong motivation to enhance the cartographic representation of indoor maps (Nossum 2013) . Indoor mapping data acquisition refers to the measurement techniques, sensors, media, and platforms used to obtain raw data from indoor environments. The main components in acquiring indoor mapping data are hardware for data processing and sensor synchronization, typically a mapping sensor such as LiDAR (Light Detection And Ranging) or an RGB-D (Red, Green, Blue-Depth) camera (Otero et al. 2020) . In this study, a backpack- Published in “Proceedings of the 18th International Conference on Location Based Services (LBS 2023)”, edited by Haosheng Huang, Nico Van de Weghe and Georg Gartner, LBS 2023, 20-22 November 2023 Ghent, Belgium. This contribution underwent single-blind peer review based on the paper. [https://doi](https://doi.org/10.34726/5732 |)[.](https://doi.org/10.34726/5732 |)[org/10](https://doi.org/10.34726/5732 |)[.](https://doi.org/10.34726/5732 |)[34726/5732 |](https://doi.org/10.34726/5732 |) © Authors 2023. CC BY 4.0 License. shaped mobile laser scanning system, in collaboration with our university, was used for acquiring indoor point data. The components of this device include GPS, LiDAR, camera, processor, battery, interface, and other connection elements. The classification of 3D point clouds belonging to indoor environments plays a significant role in the generation of indoor models (Lin et al 2021) . Significant progress has been made in the recognition of point clouds belonging to outdoor environments. However, recognizing indoor scenes remains a challenge dueto their confined surroundings, various structural features, and numerous obstacles such as columns and walls (Hangbin et al. 2020) . In recent years, the classification of indoor point clouds using deep learning algorithms has been an active research topic. The classification performances of different deep learning algorithms on the Stanford 3D Indoor Semantics (S3DIS) dataset (Armeni et al. 2016), generated by Stanford University, were provided in the study by Lin et al. (2021) . The performance of machine learning (ML) methods in the classification of indoor point clouds for high-quality indoor mapping is one of the current research topics. In this study, indoor point clouds from the S3DIS d

FAQs

Which machine learning algorithms are evaluated for classifying indoor 3D point clouds?

The study evaluates Random Forest (RF), XGBoost, Multi-Layer Perceptron (MLP), and TabNet.

How are S3DIS classes processed to create the dataset used for training and testing?

To reduce overfitting, related classes are combined into two merged groups, resulting in four total classes: ceiling, floor, merged class-1 (wall/door/window/column/board), and merged class-2 (bookcase/table/chair/clutter).

What inputs and preprocessing are used before classification?

The models use 3D coordinates (x, y, z) and RGB values, with input vectors scaled using min-max normalization.

Aug 4, 2026

Upload by Riley