Source-linked AI summary
A Scalable Deep Neural Network Architecture for Multi-Building and Multi-Floor Indoor Localization Based on Wi-Fi Fingerprinting
Kyeong Soo Kim, Sanghyuk Lee, Kaizhu Huang
TL;DR
Large-scale indoor services need localization that remains scalable across multi-story building complexes, while existing approaches estimate building, floor, and position hierarchically. This paper proposes a single-DNN Wi-Fi-fingerprinting architecture combining an SAE with a feed-forward multi-label classifier; on UJIIndoorLoc, it demonstrates feasibility and near-state-of-the-art integrated performance.
Problem
Scalable indoor localization is needed for large complexes of multi-story buildings, where existing methods estimate building, floor, and position hierarchically with separate task-specific algorithms.
Method
The proposed architecture combines an SAE for feature-space reduction with a feed-forward multi-label classifier for building, floor, and location estimation.
Results
99.82% building hit rate, 91.27% floor hit rate, 91.18% success rate, and 9.29 m positioning error were achieved at κ=8 and σ=0.2.
Takeaways & Limitations
The system demonstrates feasible multi-building and multi-floor localization using a single DNN in an integrated way, with information retained in DNN weights after training.
Takeaways & Limitations
The reported experiments are preliminary and use non-optimized DNN parameters, leaving performance and complexity trade-offs for further study.
Abstract
from arXiv · showhide
One of the key technologies for future large-scale location-aware services covering a complex of multi-story buildings --- e.g., a big shopping mall and a university campus --- is a scalable indoor localization technique. In this paper, we report the current status of our investigation on the use of deep neural networks (DNNs) for scalable building/floor classification and floor-level position estimation based on Wi-Fi fingerprinting. Exploiting the hierarchical nature of the building/floor estimation and floor-level coordinates estimation of a location, we propose a new DNN architecture consisting of a stacked autoencoder for the reduction of feature space dimension and a feed-forward classifier for multi-label classification of building/floor/location, on which the multi-building and multi-floor indoor localization system based on Wi-Fi fingerprinting is built. Experimental results for the performance of building/floor estimation and floor-level coordinates estimation of a given location demonstrate the feasibility of the proposed DNN-based indoor localization system, which can provide near state-of-the-art performance using a single DNN, for the implementation with lower complexity and energy consumption at mobile devices.
1 Introduction
Indoor Wi-Fi fingerprinting must scale to large multi-story building complexes, where existing methods estimate building, floor, and position sequentially with task-specific algorithms. The paper investigates DNNs as an integrated alternative for scalable building/floor classification and floor-level location estimation.
- Motivation: Large building complexes make the scalability of Wi-Fi fingerprinting an important indoor-localization issue.The motivating settings include shopping malls and university campuses containing many multi-story buildings.
- Prior approach: Existing state-of-the-art methods use a hierarchical sequence that separately estimates building, floor, and position.Different algorithms are tailored to each task in the sequence.
- Prior approach: Traditional Wi-Fi fingerprinting estimates coordinates by selecting similar fingerprints within the previously estimated building and floor, then averaging their coordinates.This procedure follows building and floor estimation based on subsets of fingerprints.
- DNN motivation: DNNs are investigated as attractive Wi-Fi-fingerprinting tools because they require less parameter tuning and adapt to wider conditions than traditional solutions.The cited prior work includes DNNs for coarse positioning followed by HMM refinement and DBNs evaluated on simulated heterogeneous mobile-radio data.
- This paper: The paper proposes an SAE for feature-space reduction and a feed-forward multi-label classifier for building, floor, and location estimation.The proposed system is evaluated on the UJIIndoorLoc dataset.
2 A Scalable DNN Architecture for Multi-Building and Multi-Floor Indoor Localization
The proposed architecture addresses scalability in large multi-building, multi-floor localization by combining hierarchical structure with multi-label DNN classification. It reduces output dimensionality while supporting joint processing of building, floor, and location information.
- Motivation: Large campuses can contain around 20 buildings and thousands of distinct floor-level locations, making scalable indoor localization necessary.The paper motivates scalability using the expanding XJTLU campus and its many buildings, floors, and rooms.
- Limitations of flattened classification: Flattened multi-class classification scales with the total number of locations and treats building, floor, and location errors equally.For UJIIndoorLoc, the flattened architecture represents 933 distinct locations and does not reflect the hierarchy of the classification problem.
- Proposed architecture: The proposed solution uses multi-label classification to preserve the hierarchical building/floor/location structure and address the scalability issue.Building, floor, and location identifiers are independently one-hot encoded and combined into a categorical output vector.
- Proposed architecture: The architecture combines a stacked autoencoder for feature-space reduction with a feed-forward multi-label classifier for building, floor, and location estimation.The design supports joint classification while retaining separate mappings for the three localization levels.
- Scalability: 118 output nodes are required for UJIIndoorLoc with the proposed architecture, versus 905 for flattened multi-class classification.The 118 nodes comprise 3 buildings, up to 5 floors, and up to 110 locations per floor.
- Computational implications: Parallel multi-dimensional DNN outputs can reduce the computational complexity of coordinate estimation compared with selecting nearest fingerprints using Euclidean distances.The output structure also permits separate processing and weighting of building, floor, and location classification errors.
3 Experimental Results
Experiments on UJIIndoorLoc evaluate how κ and σ affect integrated multi-building and multi-floor localization. The best reported setting achieves high building, floor, and success rates with positioning error below 10 m, while the DNN parameters remain unoptimized.
- Experimental setup: κ and σ are varied to evaluate their effects on multi-building and multi-floor localization performance using UJIIndoorLoc.The experiments focus on the number of largest output-location elements and the threshold scaling factor.
- Parameter effects: σ values of 0.1–0.3 perform best for κ≤8, whereas larger κ values favor σ=0.4 for κ=9 and σ=0.5 for κ=10.Larger κ can include reference points relatively far from the target location, making a tighter threshold more effective.
- Results: 99.82% building hit rate, 91.27% floor hit rate, 91.18% success rate, and 9.29 m positioning error are achieved at κ=8 and σ=0.2.These are the collectively best results reported for the proposed system.
- Results: The proposed results compare favorably with baseline results of 89.92% success rate and 7.9 m positioning error from distance-based kNN.Direct comparison with EvAAL/IPIN 2015 results is not possible because public UJIIndoorLoc lacks testing samples and uses a slightly different positioning-error calculation.
- Limitations: The reported results are preliminary because hidden-layer counts and layer widths were not optimized, leaving room for further performance optimization.The study emphasizes feasibility and the effects of κ and σ rather than full DNN parameter tuning.
4 Conclusions
The paper proposes a scalable single-DNN architecture that combines dimensionality reduction with multi-label building, floor, and location classification. Experiments demonstrate feasibility and near state-of-the-art integrated performance, while parameter optimization and training-objective alignment remain open issues.
- 4 Conclusions: The architecture combines an SAE for feature-space reduction with a feed-forward multi-label classifier for building, floor, and location estimation.Systematic label formation exploits localization hierarchy and provides mapping between categorical outputs and building, floor, and location vectors.
- 4 Conclusions: Multi-label reformulation greatly reduces DNN output nodes and better exploits the hierarchical structure than existing multi-class architectures.The stated scalability benefit concerns the number of output nodes and systematic label formation.
- 4 Conclusions: UJIIndoorLoc experiments demonstrate feasibility and near state-of-the-art integrated performance using a single DNN.The conclusion presents this as the principal empirical outcome of the proposed system.
- 4 Conclusions: Once trained, the DNN stores the necessary localization information in its weights rather than requiring the fingerprint database, supporting a possible mobile-only solution without server data exchange.The paper frames secure and energy-efficient mobile operation as a future possibility.
- 4 Conclusions: Further work is needed because experiments used unoptimized DNN parameters and the training or validation cost does not directly match final evaluation performance.The mismatch arises from arg max building/floor decisions and additional processing of the location vector.
Author’s contributions
Kyeong Soo Kim, Sanghyuk Lee, and Kaizhu Huang initiated the seed project underlying the manuscript. Kim proposed the architectures, conducted experiments, and drafted the manuscript, while Lee and Huang checked and clarified it.
- Kyeong Soo Kim, Sanghyuk Lee, and Kaizhu Huang initiated the seed project on scalable indoor localization underlying the manuscript.
- Kyeong Soo Kim proposed the DNN architectures, conducted numerical experiments, and drafted the manuscript.
- Sanghyuk Lee and Kaizhu Huang checked and clarified the manuscript, and all authors approved the final version.
Figures
Figure 1 presents the XJTLU campus at two time points, 2006 and 2017, enabling comparison of the campus across its development.
- Figure 1 compares the XJTLU campus in 2006 and 2017.
Tables
The paper presents tables and figures covering label formation, alternative DNN architectures, parameter settings, coordinate estimation, and comparative competition results.
- Label formation: Table 1 illustrates multi-label formation for building, floor, and location identifiers in two multi-story buildings.
- System architectures: Figures 2–4 contrast hierarchical and integrated architectures and show SAE-based feature reduction with feed-forward classification.
- Experimental parameters: Table 2 lists parameter values used for the scalable DNN-based indoor localization experiments.
- Performance analysis: Table 3 examines how κ, the number of largest location-vector elements, and σ, the threshold scaling factor, affect localization performance.
- Coordinate estimation: Figure 5 describes the procedure for estimating location coordinates, while Figure 6 provides two coordinate-estimation examples.
- Competition results: Table 4 reports the best results from four teams at the EvAAL/IPIN 2015 competition.