# RELoc: An Enhanced 3D WiFi Fingerprinting Indoor Localization Algorithm with RFECV Feature Selection

## TL;DR

RELoc is a new 3D WiFi-based indoor localization system that improves accuracy by using advanced AI techniques and 3D modeling.

## Contribution

RELoc introduces a 3D indoor localization framework combining RFECV and ERT with Bayesian optimization for better performance.

## Key findings

- RELoc achieves 1.84 m and 4.39 m MAE for 2D localization on SODIndoorLoc and UTSIndoorLoc datasets.
- Incorporating floor information improves performance by 33.15% and 26.88% on the two datasets.
- RELoc outperforms GNN, DNN, and ET by 7.52%, 12.77%, and 40.22%, respectively.

## Abstract

The use of Artificial Intelligence (AI) algorithms has enhanced WiFi fingerprinting-based indoor localization. However, most existing approaches are limited to 2D coordinate estimation, which leads to significant performance declines in multi-floor environments due to vertical ambiguity and inadequate spatial modeling. This limitation reduces reliability in real-world applications where accurate indoor localization is essential. This study proposes RELoc, a new 3D indoor localization framework that integrates Recursive Feature Elimination with Cross-Validation (RFECV) for optimal Access Point (AP) selection and Extremely Randomized Trees (ERT) for precise 2D and 3D coordinate regression. The ERT hyperparameters are optimized using Bayesian optimization with Optuna’s Tree-structured Parzen Estimator (TPE) to ensure robust, stable, and accurate localization. Extensive evaluation on the SODIndoorLoc and UTSIndoorLoc datasets demonstrates that RELoc delivers superior performance in both 2D and 3D indoor localization. Specifically, RELoc achieves Mean Absolute Errors (MAEs) of 1.84 m and 4.39 m for 2D coordinate prediction on SODIndoorLoc and UTSIndoorLoc, respectively. When floor information is incorporated, RELoc improves by 33.15% and 26.88% over the 2D version on these datasets. Furthermore, RELoc outperforms state-of-the-art methods by 7.52% over Graph Neural Network (GNN) and 12.77% over Deep Neural Network (DNN) on SODIndoorLoc and 40.22% over Extra Tree (ET) on UTSIndoorLoc, showing consistent improvements across various indoor environments. This enhancement emphasizes the critical role of 3D modeling in achieving robust and spatially discriminative indoor localization.