A Semi-Supervised Indoor Localization Algorithm Based on Probabilistic Distribution Modeling

doi:10.1109/TMC.2025.3608276

Return

A Semi-Supervised Indoor Localization Algorithm Based on Probabilistic Distribution Modeling

2025-09-10

PRE

Ruofei Gao

Xiaotao Li

Wai Chen

DOI：10.1109/TMC.2025.3608276

Original

Original request for help

Save

Abstract

En 中文

Recent advancements in wireless technology have spurred the growth of indoor localization applications based on wireless signals. Most existing methods model the relationship between input data and the target’s location using the $l_{2}$ loss function, which assumes that the residual between the predicted location and the ground truth follows a Gaussian distribution with a fixed standard deviation. Nevertheless, this approach may not conform to the actual distribution and lacks confidence measures for individual predictions, potentially compromising accuracy. Furthermore, in the realm of regression, Semi-Supervised Learning (SSL) remains relatively unexplored due to the absence of a reliable method to quantify prediction uncertainty, especially when labeled data is scarce. To address these challenges, we developed a novel indoor localization algorithm that employs probabilistic distribution modeling. Our approach focuses on indoor localization with Channel Impulse Response (CIR) as the input. Crucially, it leverages Maximum Likelihood Estimation (MLE) to model the localization error as a Gaussian distribution, and utilizes Residual Log-likelihood Estimation (RLE) to capture arbitrary error distributions. This enables us to extract the confidence of each prediction and utilize the most reliable ones as pseudo-labels for the unlabeled data. By employing probabilistic distribution modeling, we observed a significant improvement in localization accuracy over the $l_{2}$ loss function. Additionally, by integrating pseudo-labeled data for model retraining, our algorithm achieves superior performance compared to existing state-of-the-art machine learning-based and SSL-based methods. This is demonstrated by our evaluation on two public datasets, showcasing the efficiency of the proposed method.

Keywords:

Indoor localization

semi-supervised learning

probabilistic distribution modeling

channel impulse response

maximum likelihood estimation

pseudo-labels