Device-independent cellular-based indoor location tracking using deep learning
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.
Section snippets
Problem definition and system overview
In this section, we start by describing the problem in question. We then provide an overview of the different modules of the system and how they inter-operate to achieve device-independent localization. The details of OmniCells++ are given in Section 3.
The OmniCells++ system
This section presents the details of the different modules of OmniCells++ including the Pre-processor, the Features Extractor and Localization Model Constructor modules as well as the processing done in online phase. Table 1 summarizes the notations used in this paper.
Evaluation
In this section, we evaluate the OmniCells++ system in two real-world indoor testbeds as summarized in Table 2. The first one (denoted as Apartment) is an apartment of 132 m area (Fig. 8(a)). The second one (denoted as Campus), shown in Fig. 8(b), is a floor in our university campus (in a different city) with a 629 m area containing several labs with different sizes and furniture placements, meeting room, offices as well as corridors.
We start by describing the data collection setup and
Related work
In this section, we discuss the most relevant literature to our OmniCells++ system.
Conclusions
We proposed OmniCells++, a deep learning system for indoor localization designed to combat the hardware diversity problem in end-user devices. We presented the details of the system and its ability to extract the core features, in different calibration scenarios (e.g. multiple phone and single phone), to mitigate the device heterogeneity effect. To achieve that, OmniCells++ leverages a combination of relative differences and a learned, device-invariant representation using stacked LSTM
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (59)
- et al.
Gsm indoor localization
Pervasive Mob. Comput.
(2007) Powertutor-a power monitor for android-based mobile platforms
- et al.
Deep neural networks for wireless localization in indoor and outdoor environments
Neurocomputing
(2016) - et al.
Autloc: Deep autoencoder for indoor localization with rss fingerprinting
- et al.
Linear regression algorithm against device diversity for indoor wlan localization system
- et al.
Omnicells: Cross-device cellular-based indoor location tracking using deep neural networks
- et al.
Cellindeep: Robust and accurate cellular-based indoor localization via deep learning
IEEE Sens. J.
(2019) - et al.
Robust indoor localization and tracking using gsm fingerprints
EURASIP J. Wireless Commun. Networking
(2015) - et al.
Computer Networks: A Top Down Approach Featuring the Internet
(2010) - et al.
Effectiveness of data augmentation in cellular-based localization using deep learning
Hyperbolic location fingerprinting: A calibration-free solution for handling differences in signal strength (concise contribution)
Unsupervised learning of video representations using lstms
Ofcod: On the fly clustering based outlier detection framework
Data
Efficient backprop
Adam: A method for stochastic optimization
Dropout: A simple way to prevent neural networks from overfitting
J. Mach. Learn. Res.
Practical recommendations for gradient-based training of deep architectures
A scalable deep neural network architecture for multi-building and multi-floor indoor localization based on wi-fi fingerprinting
Big Data Anal.
Automatic device-transparent rss-based indoor localization
Radar: An in-building rf-based user location and tracking system
Topics in probabilistic location estimation in wireless networks
The horus wlan location determination system
Dynamicslam: Leveraging human anchors for ubiquitous low-overhead indoor localization
IEEE Trans. Mob. Comput.
The tale of two localization technologies: Enabling accurate low-overhead wifi-based localization for low-end phones
Learn to see: A microwave-based object recognition system using learning techniques
No need to war-drive: Unsupervised indoor localization
Semanticslam: Using environment landmarks for unsupervised indoor localization
IEEE Trans. Mob. Comput.
A survey of selected indoor positioning methods for smartphones
IEEE Commun. Surv. Tutor.
Csi-based fingerprinting for indoor localization: A deep learning approach
IEEE Trans. Veh. Technol.
Cited by (27)
Prediction of California bearing ratio and modified proctor parameters using deep neural networks and multiple linear regression: A case study of granular soils
2024, Case Studies in Construction MaterialsRecent advances in floor positioning based on smartphone
2023, Measurement: Journal of the International Measurement ConfederationOn the feature extraction process in machine learning. An experimental study about guided versus non-guided process in falling detection systems
2022, Engineering Applications of Artificial IntelligenceCitation Excerpt :If the dataset used does not contain enough characteristics and varied data that broadly represents the problem itself (in this study, the identification of falls and risk events), the tendency of this class of models to identify very particular characteristics results in overtraining. This is a common drawback of machine learning models that are delegated automated feature extraction (Mazurowski et al., 2019; Rizk et al., 2021). Another factor that influences the training of RNN models is the need to use a weighted loss function to avoid imbalance in accuracy between fall and risk events.
A Comprehensive Review of Indoor/Outdoor Localization Solutions in IoT era: Research Challenges and Future Perspectives
2022, Computer NetworksCitation Excerpt :As a result, according to experimental outcomes, the dataset created by GAN+ may reach an average location error that is more than ten times lower than that obtained from the original dataset. OmniCells++, the suggested system in [193], produces synthetic data from short-term data observations that reflect regular RSS fluctuation observations. The augmented data is beneficial in preventing overfitting, which can occur when training with a bit of quantity of data.
Location-Based Services for Indoor Environments: A User Experience Test
2024, BIO Web of ConferencesNeural networks implementation for the environmental optimisation of the recycled concrete aggregate inclusion in warm mix asphalt
2024, Road Materials and Pavement Design