Empowered by deep learning, data-driven smartphone inertial tracking approaches have attracted much attention due to high accuracy and robustness. However, training deep models requires significant amounts of IMU data with high-precision labels which incurs high annotation costs. In this work, we propose a practical data-driven IMU tracking solution without tedious annotations in a systematic way. We firstly design a simple but efficient method to annotate IMU data with phone labels based on the location service provided by smartphones. However, the phone labels are too noisy to train an accurate deep model. In-depth experimental studies reveal the noises are dependent along an IMU sequence; the noisy phone labels have knowledge useful for learning and make sense. The unique characteristics of the phone labels make existing learning-with-noisy-label models (LNLs) fail to work. We then propose EasyTrack, a noise-resistant data-driven IMU tracking system, which adopts a dual-model framework to enable LNLs. To handle the noisy labels, we design a series of effectively noise-resistant techniques. Extensive experiment results demonstrate without tedious annotations, EasyTrack achieves high accuracy by learning with noisy phone labels, outperforming existing LNLs and data-driven IMU tracking approaches.

Vehicle positioning using raw global navigation satellite system (GNSS) observations from smartphones is crucial for intelligent transportation and smart driving. However, it is challenging to achieve high accuracy due to GNSS signal blockage and reflection in urban areas. In this work, we propose a data-driven smartphone positioning system PhoeNix-RTK empowered by deep transformer model. PhoeNix-RTK corrects positions estimated by the real-time kinematic (RTK) approach by learning from GNSS features with PhoeNix model. To make latent features independent of satellites, PhoeNix encodes the GNSS features based on set transformer. To exploit the temporal dependence in the GNSS features, PhoeNix predicts correction vectors based on transformer in a sequence-to-sequence manner. Extensive experimental results on a public vehicle dataset demonstrate PhoeNix-RTK achieves high accuracy in urban areas, outperforming its counterparts.

Wi-Fi signals vary over time due to multipath fading and dynamic indoor environment. Hence in the long-run deployment of Wi-Fi fingerprinting localization, to retain high accuracy the fingerprint database has to be updated regularly, which is usually labor-intensive and time-consuming. In this paper, we propose a novel unsupervised domain adaptation model TransLoc for Wi-Fi fingerprint update, to keep high accuracy yet at a low cost. TransLoc consists of a feature extractor, a generator, a discriminator, and a location predictor. The feature extractor learns domain-invariant features by cooperating with other components. To further guarantee localization accuracy, the location predictor is designed as a semi-supervised regressor with three parallel sub-modules. We carry out extensive experiments in two typical real-world indoor environments with a total area of over 8,200$m^{2}$across three months. Experimental results show that with only an initial fingerprint database and current unlabeled fingerprints, TransLoc maintains high localization accuracy at a low cost in the long run.

In Wi-Fi fingerprinting indoor localization, since signals of Wi-Fi APs vary over time, to retain high localization accuracy, the fingerprint database has to be updated periodically, which is labor-intensive and time-consuming. In this paper, we consider how to detect Wi-Fi fingerprint alteration accurately on line so as to update fingerprint database efficiently. To this end, we propose a deep learning model, AReAE (Alteration Reducing AutoEncoder), to reconstruct fingerprints from altered ones by learning alteration distribution. Based on AReAE, we further develop a FADet (Fingerprint Alteration Detection) system, which constructs a fingerprint ASmap (Alteration Strength map) with crowd-sourced Wi-Fi RSS (Received Signal Strength) measurements. The ASmap indicates how strong fingerprints change anywhere in the area of interest. FADet is verified through extensive experiments in real-world indoor scenarios. Results show that FADet yields accurate ASmaps for all scenarios tested, which can facilitate fingerprint database updating in time.

Geomagnetic field and WiFi networks have attracted much attention in indoor localization, since they are pervasive, and their signals are easy to collect by off-the-shelf smartphones. However, there are still challenges in practice: 1) GFI (Geomagnetic Field intensity) read by magnetometer in smartphones is hard to be used directly, since it is in phone coordinate system and varies with phone attitude. On the other hand, the magnitude of GFI lacks location diversity; 2) WiFi RSS (Received Signal Strength) suffers from fluctuations due to multipath fading effect indoors. To address the above issues, we first propose Mag2D, a high quality feature of GFI. Mag2D is robust to phone attitude and has better spatial discrimination than the magnitude. We then study the complementary properties of Mag2D and RSS, and design MagWi, a practical fingerprinting indoor localization system fusing Mag2D and RSS. MagWi adjusts fusion weights of Mag2D and RSS according to their capabilities of discriminating locations. To do so, we propose LCD (Local Clustering Degree) to quantify the location-discriminating capability and build a LCD-Error model. The fusion weights are thus assigned dynamically based on the LCD-Error model during localization. We have implemented MagWi to provide location service on an Android phone in three typical indoor environments, covering a total area size over 600m2. Our experiment results show that MagWi achieves high accuracy and is easy to deploy in practice.

Thanks to the highly-dense lighting infrastructure in public areas, visible light emerges as a promising means to indoor localization and navigation. State-of-the-art techniques generally require customized hardware (sensing boards), and mainly work with one single light source (e.g., customized LEDs). This greatly limits their application scope. In this paper, we propose NaviLight, a generic indoor localization and navigation framework based on existing lighting infrastructure with any unmodified light sources (e.g., LED, fluorescent, and incandescent lights). NaviLight simply adopts commercial off-the-shelf mobile phones as receivers, and light intensity values as location signatures. Unlike existing WiFi systems, a single light intensity value is not discriminative enough over space though the light intensity field does vary, which makes our design more challenging. We thus propose a LightPrint as a location signature using a vector of multiple light intensity values obtained during user's walks. Such LightPrints are created by leveraging any user movement (of varying distance and direction) in order to minimize user efforts. A set of techniques are proposed to achieve quick LightPrint matching, which includes a coarse-grained classification and a fine-grained matching over dynamic time warping. We have implemented NaviLight to provide real-time service on Android phones in three typical indoor environments, covering a total area size over 1000m2. Our experiments show that NaviLight can achieve sub-meter localization accuracy to meet practical engineering requirements.