Transferring Localization Models across Space

Sinno Jialin Pan, Dou Shen, Qiang Yang, James T. KwokAAAI 2008 · OpenReview

Abstract

Machine learning approaches to indoor WiFi localization involve an offline phase and an online phase. In the offline phase, data are collected from an environment to build a localization model, which will be applied to new data collected in the online phase for location estimation. However, collecting the labeled data across an entire building would be too time consuming. In this paper, we present a novel approach to transferring the learning model trained on data from one area of a building to another. We learn a mapping function between the signal space and the location space by solving an optimization problem based on manifold learning techniques. A low-dimensional manifold is shared between data collected in different areas in an environment as a bridge to propagate the knowledge across the whole environment. With the help of the transferred knowledge, we can significantly reduce the amount of labeled data which are required for building the localization model. We test the effectiveness of our proposed solution in a real indoor WiFi environment.

Introduction

Indoor localization is a critical component for smart building navigation and location-based services. Most modern systems rely on WiFi Received Signal Strength (RSS). Traditionally, these systems require an extensive "site survey" where engineers manually record WiFi signals at known coordinates to create a radio map. This process is labor-intensive and fails to scale to large multi-story complexes.

While semi-supervised learning can leverage unlabeled data, most existing methods still require uniform labeled data collection across the entire target space. This research addresses the challenge of spatial transfer: can we train a high-quality model in one well-mapped area (Area A) and adapt it to a new, unmapped area (Area B) with minimal additional effort?

Methodology

The proposed approach, TransMapping, treats localization as a manifold learning problem. It assumes that while WiFi signals exist in a high-dimensional space defined by numerous Access Points (APs), the physical movement of a device occurs on a 2D manifold.

The core innovation lies in using Access Points as "anchors" that bridge different spatial domains. The method follows a two-step optimization process:

  1. Latent Knowledge Extraction: Data from the labeled Area A is used to estimate the physical coordinates of the fixed AP infrastructure. These locations represent domain knowledge that remains constant across the building.
  2. Transfer via Constraints: For Area B, the model formulates a Quadratically Constrained Quadratic Program (QCQP). This optimization uses the previously identified AP locations to constrain the model, ensuring consistency with the physical infrastructure. Graph Laplacians are employed to propagate label information from a few known points in Area B across the entire unlabeled manifold.

Experimental Setup

The researchers evaluated TransMapping in a real-world office environment measuring approximately 35m x 120m. The setup involved 150 detected Access Points, with ground-truth locations known for only 17 of them. The environment was divided into two distinct regions for the transfer task.

Ground Truth Environment showing the spatial separation of Area A and Area B.
  • Area A: Contained 665 labeled examples used for initial training.
  • Area B: Contained 486 examples, the majority of which were treated as unlabeled to test the transfer efficiency.

Key Results

TransMapping significantly outperformed standard semi-supervised baselines, particularly when labeled data in the target area was scarce.

Comparison of average error distance as a function of the number of labeled data points in Area B.

As shown in the results, with only 7 labeled points in Area B, TransMapping achieved an average error of approximately 2 meters. In contrast, the baseline LeMan algorithm exhibited a much higher average error of 6.64m in the same area. The method allows for a drastic reduction in manual labor, requiring only one labeled data point every 15 meters to maintain high accuracy.

Discussion and Future Work

The success of TransMapping hinges on the stability of the WiFi infrastructure; if Access Points are moved, the latent knowledge must be updated. While this study focused on a 2D office plane, the underlying logic of manifold alignment is expected to hold for 3D multi-story environments. Future research will focus on developing "online" versions of this algorithm that can adapt to dynamic environments in real-time as users move through the space.

Summary updated

Abstract

Machine learning approaches to indoor WiFi localization involve an offline phase and an online phase. In the offline phase, data are collected from an environment to build a localization model, which will be applied to new data collected in the online phase for location estimation. However, collecting the labeled data across an entire building would be too time consuming. In this paper, we present a novel approach to transferring the learning model trained on data from one area of a building to another. We learn a mapping function between the signal space and the location space by solving an optimization problem based on manifold learning techniques. A low-dimensional manifold is shared between data collected in different areas in an environment as a bridge to propagate the knowledge across the whole environment. With the help of the transferred knowledge, we can significantly reduce the amount of labeled data which are required for building the localization model. We test the effectiveness of our proposed solution in a real indoor WiFi environment.
BibTeX
@inproceedings{pan2008transferring,
  author = {Sinno Jialin Pan and Dou Shen and Qiang Yang and James T. Kwok},
  title = {{Transferring Localization Models across Space}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
  url = {https://openreview.net/forum?id=rk-xACl_Wr},
}
Figure