Bluetooth low energy indoor positioning: A fingerprinting neural network approach

Abstract

This study explores the application of neural networks in indoor positioning using BLE (Bluetooth Low Energy) and the Fingerprinting location technique. The methodology involves two main phases: the capture and filtering process, where received BLE signals are smoothed and combined into fingerprint vectors, and the subsequent location prediction phase, which compares the position estimation from eight neural network designs and the classical trilateration method. We conduct a performance comparative analysis of each prediction method and study the optimal parameter values for the capturing and filtering processes. The research underscores the limitations of training metrics in reflecting real-world performance, emphasizing the importance of testing models on actual trajectories. Results indicate that regression neural networks outperform classification ones, and a complex dense neural network model proves most versatile and stable across testing scenarios. Our approach achieves a mean error of 1.9 meters, surpassing existing accuracies of 3.7 meters for trilateration and 3.1 meters for state-of-the-art neural network designs, thus holding promise for significantly improving indoor positioning accuracy with practical implications across various domains.