VI-SLAM2tag: Low-Effort Labeled Dataset Collection for Fingerprinting-Based Indoor Localization

TL;DR

VI-SLAM2tag automates the labeling of indoor fingerprint data using visual-inertial SLAM, significantly reducing manual effort and enabling the creation of accurate datasets for fingerprinting-based localization.

Contribution

The paper introduces VI-SLAM2tag, a novel system that auto-labels indoor fingerprints with high accuracy using local trajectory transformations, facilitating easier dataset collection.

Findings

Achieved approximately 50 cm labeling error for 15-minute trajectories.

Collected a multi-floor dataset with WLAN and IMU data.

Neural network models trained on this data reach about 2 m positioning accuracy.

Abstract

Fingerprinting-based approaches are particularly suitable for deploying indoor positioning systems for pedestrians with minimal infrastructure costs. The accuracy of the method, however, strongly depends on the quality of collected labeled fingerprints within the calibration phase, which is a tedious process when done manually in a static fashion. We present VI-SLAM2tag, a system for auto-labeling of dynamically collected fingerprints using the visual-inertial simultaneous localization and mapping (VI-SLAM) module of ARCore. ARCore occasionally updates its internal coordinate system. Mapping the entire trajectory to a target coordinate system via a single transformation thus results in large drift effects. To solve this, we propose a strategy for determining locally optimal sub-trajectory transformations. Our system is evaluated with respect to the accuracy of the generated position labels using a geodetic tracking system. We achieve an average labeling error of roughly 50 cm for trajectories of up to 15 minutes, which is sufficient for fingerprinting-based localization. We demonstrate this by collecting a multi-floor dataset including WLAN and IMU data and show how it can be used to train neural network based models that achieve a positioning accuracy of roughly 2 m. VI-SLAM2tag and the dataset are made publicly available.