LabellessFace

Abstract

In environments where GPS is impractical for indoor positioning, magnetic information has emerged as a promising alternative. This study proposes a novel location fingerprinting method called "MagSerea ((Magnetic Field of Specified aREA))," which utilizes time-series data of three-axis magnetic information obtained by a wearable device attached to the arm when opening a door. Experimental results indicate that the door-opening motion restricts the deviations in the device's position and angle, thereby maintaining high identification accuracy despite aging and changes in user's possessions. These findings suggest that MagSerea demonstrates high identification accuracy in real-world conditions, offering lower installation costs and fewer constraints on reference points compared to existing methods, and thus holds potential for a wide range of applications.

Proposed Method

Obtaining Sensor Information

Magnetic and angular velocity data are collected along three axes. The magnetic data is used as the location fingerprint, while the angular velocity data assists in preprocessing the magnetic information.

Preprocessing

Preprocessing involves several steps to prepare the data for analysis. First, high-frequency noise in the magnetic and angular velocity data is filtered using an RC low-pass filter to retain important low-frequency information:

$$ S_{\text{filtered}} = \alpha \cdot S_{\text{current}} + (1 - \alpha) \cdot S_{\text{previous}} $$

Here, $ S_{\text{filtered}} $ is the filtered sensor information, $ S_{\text{current}} $ is the newly acquired data, and $ S_{\text{previous}} $ represents the previously recorded data. The coefficient $ \alpha $ is set to 0.8.

After filtering, the system extracts the specific action section related to door-opening by detecting peaks in the angular velocity data. This isolates the portion where the wrist rotates during doorknob manipulation. Next, normalization is applied to address internal magnetism fluctuations from the device, which can affect the magnetic data. The magnetic data is standardized using:

$$ x' = \frac{x - \bar{x}}{s} $$

Here, $ x' $ is the normalized data, $ x $ is the original data point, $ \bar{x} $ is the mean, and $ s $ is the standard deviation. This ensures data from different sessions can be consistently compared.

To ensure uniform data lengths across all location fingerprints, resampling is performed using 3D spline interpolation. This technique smooths the time-series data and adjusts it to the required input size for the classifier:

$$ S_i(x) = a_i + b_i(x - x_i) + c_i(x - x_i)^2 + d_i(x - x_i)^3 $$

Here, $ x_i $ represents the position of a data point in the time-series, and $ S_i(x) $ is the interpolated value between data points $ x_i $ and $ x_{i+1} $. This ensures that the location fingerprints have uniform data lengths before being input into the 1D-CNN model for feature extraction.

Machine Learning Model

The machine learning model used is a one-dimensional convolutional neural network (1D-CNN) combined with ArcFace, a distance metric learning method. The lightweight 1D-CNN, with 2.07 MB and 12,842 parameters, consists of three convolutional layers with ReLU activation and pooling layers to reduce feature dimensionality.

ArcFace enhances location feature distinction by calculating similarity scores. It computes the inner product of input features and a class matrix, applying a margin to the correct class score. The 1D-CNN produces a 64-dimensional feature vector, which is used to determine whether two location fingerprints match based on the similarity score.

Proposed Method Diagram

Ageing Robustness Evaluation

The evaluation used data from three sessions to assess the system's performance over time. Training was based on 80 data points from the first session, with testing on the remaining 20. Data pairs from different positions were undersampled for consistency. The evaluation metrics, including EER, AUC, and F1 score, are shown in the table below.

Session	EER (%)	AUC (%)	F1 (%)
Session 1-1	0.075	99.0	99.0
Session 1-2	0.89	99.4	98.2
Session 1-3	2.39	99.6	97.6

User’s Property Change Robustness Evaluationn

This evaluation tested the robustness of location fingerprints when the user's possessions changed. Using 400 data pairs per case, accuracy was the primary metric. The results show the method is robust to changes in the user’s items.

Case	Description	Accuracy
Case 0	Smartphone in trouser pocket near doorknob	1.00
Case 1	Wearing wireless earphones	0.9725
Case 2	Wearing wireless earphones in the ears	1.00
Case 3	Holding a PC under arm (opposite side of doorknob)	1.00

Evaluation of Location Fingerprints Obtained by Different Users

The system’s performance was evaluated across four users using 400 data pairs each, with accuracy as the primary metric. Users 1 and 2 showed high accuracy, while users 1 and 3 did not, suggesting the system may be affected by individual differences in door-opening actions.

User	Accuracy
User 0	1.00
User 1	0.585
User 2	1.00
User 3	0.54

Citation


        @inproceedings{aiba2024mobilehci,
          title={MagSerea: Fingerprinting Magnetic Field of Specified Area with Wearable Sensors},
          author={Tomoya Aiba and Takahiko Arakawa and Tatsuya Mori and Masakatsu Nishigaki and Tetsushi Ohki},
          booktitle={The ACM International Conference on Mobile Human-Computer Interaction (MobileHCI 2024)},
          year={2024}
        }

Acknowledgement

This work was supported in part by JSPS KAKENHI JP22K19782, and JST Moonshot R&D Grant Number JPMJMS2215.