A New Statistical Approach to Calibration-Free Localization Using Unlabeled Crowdsourced Data

2025-04-04

AI Summary

Traditional fingerprint-based indoor localization suffers from poor scalability due to high-cost site surveys and frequent recalibration. This paper proposes an unsupervised indoor localization method that requires only unlabeled crowdsourced RSS measurements. It constructs a distance estimation model based on the cumulative distribution function (CDF), replacing empirical path-loss models to mitigate shadowing and multipath effects. Localization is then achieved via a three-step unsupervised geometric framework. To our knowledge, this is the first unsupervised approach achieving positioning accuracy comparable to supervised k-NN—without requiring fingerprint collection or periodic re-surveying. Extensive evaluation on ray-tracing simulations and real-world measurements demonstrates significantly higher accuracy than conventional path-loss models, validating the effectiveness of statistical modeling and CDF-based mapping for unsupervised localization.

Technology Category

• Intelligent Robots: Localization, Mapping, and Navigation
• Planning, Routing, and Scheduling: Scheduling under Uncertainty
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Security and Privacy: Large-scale security measurements
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs

Abstract

Fingerprinting-based indoor localization methods typically require labor-intensive site surveys to collect signal measurements at known reference locations and frequent recalibration, which limits their scalability. This paper addresses these challenges by presenting a novel approach for indoor localization that utilizes crowdsourced data without location labels. We leverage the statistical information of crowdsourced data and propose a cumulative distribution function (CDF) based distance estimation method that maps received signal strength (RSS) to distances from access points. This approach overcomes the limitations of conventional distance estimation based on the empirical path loss model by efficiently capturing the impacts of shadow fading and multipath. Compared to fingerprinting, our unsupervised statistical approach eliminates the need for signal measurements at known reference locations. The estimated distances are then integrated into a three-step framework to determine the target location. The localization performance of our proposed method is evaluated using RSS data generated from ray-tracing simulations. Our results demonstrate significant improvements in localization accuracy compared to methods based on the empirical path loss model. Furthermore, our statistical approach, which relies on unlabeled data, achieves localization accuracy comparable to that of the supervised approach, the $k$-Nearest Neighbor ($k$NN) algorithm, which requires fingerprints with location labels. For reproducibility and future research, we make the ray-tracing dataset publicly available at [2].

Problem

• Eliminates need for labeled data in indoor localization
• Overcomes limitations of empirical path loss models
• Achieves accuracy comparable to supervised methods

Innovation

• Uses unlabeled crowdsourced data for localization
• Proposes CDF-based RSS to distance mapping
• Eliminates need for known reference locations