[Paper Review] Precise Indoor Positioning Based on UWB and Deep Learning

Chenyu Wang, Zihuai Lin|arXiv (Cornell University)|Apr 17, 2022

Indoor and Outdoor Localization Technologies4 citations

TL;DR

This paper proposes a UWB-based indoor positioning system enhanced with a BP neural network to improve accuracy by leveraging a fingerprint database derived from measured-to-true distance relationships. The method achieves up to 73% average improvement in positioning accuracy over traditional trilateration by classifying target positions using distance fingerprints trained via deep learning.

ABSTRACT

We examined UWB-based indoor location in conjunction with a fingerprint technique in this work. We built a connection between the measured and real distances for the UWB indoor positioning system. This connection is used to produce a distance database that may be used to generate fringerprints. We created a BP neural network to classify the target node to the relevant fringerpint using the distance database. Our suggested deep learning technology considerably enhances location accuracy when compared to existing trilateration systems.

Motivation & Objective

• To address the limitations of conventional trilateration in UWB-based indoor positioning, particularly due to multipath and NLOS errors.
• To develop a robust fingerprint-based positioning system using measured-to-true distance calibration for improved accuracy.
• To evaluate the performance of a BP neural network in classifying target positions using a distance fingerprint database.
• To investigate optimal reference point selection strategies for training the neural network and minimizing measurement error.
• To explore future enhancements through larger fingerprint databases and alternative deep learning architectures.

Proposed method

• A connection between measured UWB distances and true distances is established to create a calibrated distance database for fingerprinting.
• The system divides the indoor area into a grid, with reference points used to collect RSSI and TOA-based distance measurements for fingerprint generation.
• A BP neural network is trained to map the collected distance fingerprints to specific spatial locations, enabling real-time position classification.
• Three training strategies are evaluated: Two-point Diagonal Training (TDT), Two-point Non-diagonal Training (TNT), and Coalescent Training (CT) combining both.
• The CT method integrates data from both TDT and TNT configurations to improve generalization and reduce error variance.
• Positioning accuracy is evaluated using mean error distance across multiple test points, comparing the BP neural network to conventional trilateration.

Experimental results

Research questions

• RQ1 Can a BP neural network trained on calibrated UWB distance fingerprints significantly improve indoor positioning accuracy over trilateration?
• RQ2 How does the spatial distribution of reference points affect the performance of the fingerprint-based deep learning model?
• RQ3 What is the impact of measurement errors near anchor nodes on neural network generalization and positioning accuracy?
• RQ4 Does combining multiple training strategies (TDT and TNT) through coalescent training yield better accuracy than individual approaches?
• RQ5 How does the size and quality of the fingerprint database influence the performance of the deep learning-based positioning system?

Key findings

• The BP neural network approach reduced mean error distance by up to 73% on average compared to conventional trilateration.
• The Coalescent Training (CT) method delivered the best overall performance, significantly outperforming both TDT and TNT strategies.
• Group 2 in the TDT experiment showed worse performance than Group 1, primarily due to low measurement accuracy when the target was too close to an anchor node.
• The TNT method performed better than TDT in Group 2, indicating that non-diagonal reference point selection reduces error propagation from proximity to anchor nodes.
• The study confirms that reference points should be placed in the central region of the test area rather than near anchor nodes to avoid measurement inaccuracies.
• The results suggest that expanding the fingerprint database with more reference points and optimizing their placement can further enhance system accuracy.