Multiclass Permanent Magnets Superstructure for Indoor Localization using Artificial Intelligence

Abstract

Smartphone-based indoor localization methods draw increased interest and are frequently employed for position estimation of users inside enclosures like malls, conference halls, and crowded venues. Magnetic sensing has recently attracted considerable attention due to its pervasiveness and autonomy, introducing two main positioning approaches utilizing magnetic sensors. The first approach is identifying the magnetic field characteristics of the surrounding infrastructure as signatures, thus creating a specific magnetic fingerprint. This scheme does not require additional hardware deployment, while interferences and anomalies from indoor infrastructure and objects greatly improve magnetic field discernibility. Yet, this approach requires knowledge of magnetic maps or computationally expensive calculations for applying simultaneous localization and mapping (SLAM). Furthermore, some environments may cause strong electromagnetic interference that differs from the underline magnetic fingerprint, leading to localization errors. The second approach uses active magnetic flux transmitters as markers, which is energy consuming and requires expensive transmitter units.In our recent work, we introduced the Permanent Magnets Superstructure for Indoor Localization method (PMSIL). This approach allows simple, low-power, and robust smartphone localization in unknown-magnetic-fingerprint locations by deploying permanent magnets in pre-known locations. Practically, we embed small-volume-large-moment permanent magnets in given locations inside the building and arrange them in a single specific geometric configuration. This results in a supervised super-structure magnetic signature pattern, constituting unambiguous magnetic environments. The localization algorithm learned this unique pattern during the training stage using artificial intelligence (AI) methods and detected it in ongoing data streams during real-time localization by applying binary classification. Thus, localization is based on smartphone motion rather than on the static positioning of the magnetometer. In this study, we propose an extended version of the PMSIL method, called Multiclass-PMSIL (M-PMSIL), in which magnets are arranged in a variety of configurations in pre-known locations. These constellations generate a bank of distinct patterns that are learned by the multiclass AI-based algorithms during training and recognized on-the-fly during localization. This extension permits covering a broader area in which the user can be localized. As M-PMSIL is landmark-based, we draw comparisons to three competing state-of-the-art approaches for indoor localization with landmark-based magnetism mechanisms. We also refer to localization methods based either on spatial-temporal sequence matching or on fusion with motion.The database is recorded by a user strolling back and forth with a hand-held smartphone across a 100 m long and 3 m wide in-door corridor. Along this corridor, three neodymium magnets with different magnetic moment norms are positioned in a row, with 3 m separating between them. During the experiment, these magnets are permuted in all 6 possible combinations and every new combination is relocated along the corridor, creating 6 distinct magnetic patterns. For the sake of generalization, recordings are made in different walking paces between 0.5 to 1.5 m/s, and different crossing distances from the magnets of 0.5 m, 1m, and 1.5 m. A snippet of data recorded by the smartphone indoor can be seen in Fig. 1, including 2 of the 6 magnet constellations used in this study. The training set includes 40 minutes of randomly selected recorded data that contains 10 passes across each of 6 different magnetic patterns, while the test set contains 20 minutes of data. The 6 classes, i.e., magnetic permutations to be localized, are balanced for both disjoint sets.In this study, 6 AI-based architectures are used for localization. Their performance is compared in Fig. 2 using the receiving operating characteristic (ROC) curve, allowing analysis of the trade-off between true positive rate (TPR) and false-positive rate (FPR) in various operation points. We may observe that the long short-term memory (LSTM) network produced the best classification results, followed by the gated recurrent unit (GRU) and recurrent neural network (RNN) models. The least performant methods are the fully connected deep neural network (DNN), and support vector machine (SVM), with and without principal component analysis (PCA). To quantify the multiclass classification accuracy of the methods, we derive the maximal sum of TPR and FPR from the ROC curve. LSTM shows 95% accuracy, while GRU and RNN reach 90% and 87%, respectively. DNN shows 82%, SVM-PCA method reaches 80%, and SVM produces 75% accuracy.Our approach achieves mean localization error (MLE) lower than 1 m, supersedes competing landmark-based methods, as UnLoc and MapCraft achieved MLE between 1-2 m. The IODetector was able to obtain classification accuracy as high as 82% indoor, which falls short to the 95% obtained by the LSTM. Additional methods, rooted in spatial-temporal sequence matching or on fusion with motion, also show lower MLE and classification accuracy compared to our study. **