Comparative study of indoor positioning datasets focusing on localization accuracy success rate and floor classification

Abstract

Indoor positioning systems (IPS) are crucial for enabling location-aware services in GPS-denied environments such as buildings, hospitals, and industrial facilities. However, the lack of standardized, high-quality datasets remains a significant barrier to the development and fair evaluation of localization algorithms. This paper presents a comprehensive comparative analysis of four widely used indoor positioning datasets—BLE Indoor, SODIndoorLoc, TUJI1, and UJIIndoorLoc—evaluating their localization accuracy, success rate, and floor classification performance. Through exploratory data analysis and systematic experimentation using an XGBoost model with consistent hyperparameters, we identify how dataset-specific characteristics such as signal sparsity, access point (AP) density, device heterogeneity, and spatial layout affect localization outcomes. Results show that while high AP density can enhance accuracy, other factors like environmental complexity, weak RSSI signals, and multi-device variability significantly influence model performance. Based on our findings, we offer practical guidelines for dataset selection, feature engineering, and design improvements to support robust and generalizable IPS development. This study provides critical insights for both researchers and practitioners aiming to advance reliable indoor localization solutions.