Device-independent cellular-based indoor location tracking using deep learning

Abstract

The demand for a ubiquitous and accurate indoor localization service is continuously growing. Cellular-based systems are a good candidate to provide such ubiquitous service due to their wide availability worldwide. One of the main barriers to the accuracy of such services is the large number of models of cell phones, which results in variations of the measured received signal strength (RSS), even at the same location and time. In this paper, we propose OmniCells++, a deep learning-based system that leverages cellular measurements from one or more training devices to provide consistent performance across unseen tracking phones. Specifically, OmniCells++ uses a novel approach to multi-task learning based on LSTM encoder–decoder models that allows it to learn a rich and device-invariant RSS representation without any assumptions about the source or target devices. OmniCells++ also incorporates different modules to boost the system’s accuracy with RSS relative difference-based features and improve the deep model’s generalization and robustness.

Evaluation of OmniCells++ in two realistic testbeds using different Android phones with different form factors and cellular radio hardware shows that OmniCells++ can achieve a consistent median localization accuracy when tested on different phones. This is better than the state-of-the-art indoor cellular-based systems by at least 148%.

Introduction

¹Recent years have witnessed an ever-growing demand for accurate and ubiquitous indoor positioning systems in many applications. Towards achieving this goal, WiFi-based indoor localization is widely adopted. WiFi-based indoor localization systems harness the Received Signal Strength (RSS) captured by the user’s phone from WiFi access points as signatures for the location identification, building a WiFi fingerprint database. Fingerprinting is a two-phase technique, consisting of an offline and an online phase. During the offline phase, the received signals are captured at known points (i.e. fingerprints) in the area of interest where their strengths are used to characterize the corresponding locations. Subsequently, the collected fingerprints are used to build a localization model for estimating the user location in the online phase given a signal scan. The major challenge for fingerprint-based approaches is the extraction of robust and discriminative signatures across space and across different devices or different user behaviors.

On the other hand, cellular based techniques have a number advantages that make them more attractive than their WiFi counterparts. First, cellular coverage far exceeds the coverage of WiFi networks as cell towers are dispersed with high density across the inhabited world. Second, all cell phones including low-end ones, by definition support cellular technology. Third, a cellular-based localization system will still work even with a failure in buildings’ electrical infrastructure as cellular base stations are better equipped to tolerate power failures. Finally, network configuration changes rarely occur due to the consequent significant expense and complexity of this process when performed frequently. This leads to a ubiquitous and stable operating environment for localization systems, which does not involve tedious re-calibration.

Current cellular-based localization techniques [2], [3], [4] are designed to capture the fingerprint of the received signal strength from the different cell towers detectable in the area of interest. These fingerprints are then used to train a classifier that differentiates between different reference points in the area of interest. Different types of classifiers are proposed in literature, e.g. Support Vector Machines [3] and K Nearest Neighbors [4]. For better feature representation ability and localization performance, deep learning was adopted in [2].

While these techniques have opened the possibility of leveraging the ubiquitous cellular networks for the indoor localization domain, they face performance degradation when deployed in practice. This degradation is mainly caused as a result of the assumption made by these techniques; i.e. the distribution of the collected fingerprints is independent of the existence of individual scenarios. An individual scenario can be defined as a combination of multiple dependencies, including the considered mobile device (different hardware/software specifications) and its placement (e.g. hold smartphones in a pocket or in a bag), which in turn generates dissimilar cellular distributions; even when the user is stationary at the same location. As such, the quality of the localization models drops significantly when the system is trained in a specific scenario and tested in a different one, which is the typical real world use-case. To solve this problem without compromising on accuracy, a localization model may be trained for every possible consumer phone in every possible placement. However, this solution is intrinsically impractical due to the huge number of phones available on the market and the requirement to collect sufficient amounts of data using each individual phone to train a corresponding model.

To tackle the above problems, we propose OmniCells++: a deep learning-based cellular localization system trained with data collected from as few as only one phone. It is capable of providing stable and accurate positioning even when used by other unseen phones without requiring any information about the considered phones or their placements. This is made possible by extracting device invariant features, i.e. relative features from the data recorded by the calibration phone(s) during the offline phase. Towards this, we leverage LSTM encoder–decoder networks to capture the relationship between the RSSs received from the covering towers in successive scans and transform these raw RSSs to a latent space where different phones’ data, captured at any phone placement, are identically distributed. These features are then harnessed to train a deep neural network to map the transformed RSS data to the user’s location. In the tracking phase, the transformed features are extracted from the cell measurements reported by an unseen phone and then fed to the trained deep model to estimate the user location.

Evaluation is done in two different testbeds using different Android phones. The results confirm that OmniCells++ performs equally when trained and tested using different phones with a consistent median location error of 1.22 m and 1.66 m in a small and a large testbed; respectively. This accuracy surpasses the accuracy of the state-of-the-art cellular-based techniques in the considered testbeds by more than 190.5% and 148.3%; respectively.

This paper extends our earlier work in [1]. Specifically, we propose a new feature extraction model (LSTM encoder–decoder) to provide more robust and device-invariant features. These features are obtained through training the feature extraction model on a sequence of input scans rather than a single scan as in [1]. Therefore, the proposed feature extraction model is designed to learn the underlying relationship between the signals received from cell towers as well as the temporal correlation (i.e. historical changes) between successive scans, leading to better device independent. Furthermore, OmniCells++ compensates for the variation of the signals resulting from the different phone placements. The proposed method enhances the accuracy by 37% and 23.5% compared to our earlier method in [1] in the small and large testbeds, respectively. In addition, it significantly reduces the errors incurred due to varying device placement in run-time.

The rest of this paper is structured as follows. In Section 2, we provide a brief description of the problem and an overview of the proposed system. Section 3 presents in detail the methodology proposed by OmniCells++. In Section 4, we present the detailed evaluation of OmniCells++.

Sections 5 Related work, 6 Conclusions discuss related work and conclude the paper respectively.