Data-driven and machine learning-based approaches for mobile indoor positioning

Abstract

Location is vital for numerous applications driven by uncountable mobile users and developers. The global navigation satellite system (GNSS) has been served for years to provide highprecision localization and relevant applications in outdoor scenarios. However, the low penetration of GNSS signal through walls and obstacles sharply decreases the positioning accuracy in the indoor environment. Consequently, various indoor positioning methods have been proposed in recent years to facilitate location-based services in indoor scenarios.

In recent years, the widespread adoption of smartphones and other mobile devices has led users to generate an enormous amount of data that can be analyzed to extract valuable insights such as movement patterns, user behavior, and environmental conditions. As a result, people now have high expectations for the accuracy and reliability of these services. Advancements in sensor technologies such as Wi-Fi and Bluetooth have made it possible to collect precise data about indoor environments for positioning. Machine learning technologies have allowed for the identification of patterns and relationships that may not be immediately apparent to human analysts from sensor data. Deep learning, a subset of machine learning, can process large amounts of data and can be deployed to identify complex patterns and relationships that would be difficult for traditional machine learning algorithms to detect. Therefore, motivated by the increasing user requirements and advancements in sensor technologies, as well as the emergence of machine learning and deep learning, this thesis investigates the potential of data-driven approaches to deliver dependable and scalable solutions for indoor positioning systems.

This thesis first analyzes and explores the features, challenges, and applications of two popular indoor environmental signals: magnetic field and WiFi. In numerous indoor positioning systems, the magnetometer in smartphones has played a crucial role in providing location information, including orientation, user trajectory construction, and magnetic field-based fingerprinting. However, the magnetometer measurements face challenges due to magnetic disturbance caused by metal infrastructures, electrical equipment, and other electronic devices in complex indoor environments. Consequently, this thesis presents a novel data-driven solution for detecting magnetic disturbance using unsupervised learning. The research focuses on extracting and analyzing statistical features from smartphone magnetometer measurements. Based on extensive experiments and analysis of the covariance and magnitude difference, two unsupervised learning-based methods are developed and evaluated in static and dynamic scenarios to demonstrate their reliability and robustness. Results consistently indicate that the proposed approach outperforms conventional methods across all experimental conditions.

WiFi is also widely used for indoor positioning services on smartphones, but its accuracy is often affected by ranging errors caused by non-line-of-sight (NLOS) conditions. To address this issue, this thesis introduces a novel data-driven approach for real-time NLOS/LOS identification using WiFi Received Signal Strength (RSS) and WiFi Round-trip Time (RTT). Through extensive analysis of the dispersion characteristics of WiFi RSS and WiFi RTT, three machine learning algorithms are selected and developed to differentiate samples corresponding to NLOS/LOS conditions. The experiments are conducted using data collected from commercial smartphones and WiFi access points in actual experimental sites, without any prior infrastructure setup or reconfiguration. The proposed methods exhibit the highest identification accuracy while maintaining the lowest latency compared to contemporary solutions.

Next, the thesis studies the commonly employed WiFi fingerprinting methods for indoor positioning systems. To solve one key issue of lacking pre-collected WiFi fingerprints and to reduce the burden of heavy human labor, a scalable WiFi fingerprint augmentation method is proposed. This method utilizes a multivariate Gaussian process regression (MGPR) model to estimate the collective distribution of all WiFi signals. It then predicts the signals at unsurveyed potential reference points computed by two new schemes, thereby generating additional fingerprints. The effectiveness of the proposed solution is evaluated using an open-source dataset obtained from a multi-floor building. The results demonstrate that the proposed solution significantly enhances positioning accuracy while maintaining lower computational complexity than conventional augmentation methods.

Finally, the thesis explores data fusion for indoor positioning systems that utilize various modalities. A new data-driven method called Multimodal Graph Fingerprinting is proposed. The method constructs a multimodal graph at the location of the user’s smart terminal by integrating radio frequency signals, electromagnetic field (EMF) strength, and inertial sensor measurements. A hierarchical deep graph neural network is developed to learn the correlations between the multimodal graphs and their respective locations by capturing the features of the identities and topology information. The proposed method is evaluated using a real dataset collected from the university campus. The results show that by integrating various modalities, the proposed model achieves a median positioning error of 2.1m.