1. Introduction
Smartphone positioning techniques and their navigation capability based on Global Navigation Satellite Systems (GNSS) have seen intensive development in the last decade. The proliferation of GNSS-enabled smartphones and wearables has boosted the evolving industry of Location-Based Services (LBS), around which a growing number of applications such as lane-level navigation, personnel/property monitoring, augmented reality, etc., are thriving. Many of these applications require greater user positioning accuracy, resilience, and/or availability.
It has been a topic for both industry the research communities to innovate within the low-cost domain to address these challenges and improve solutions further. The idea for improving smartphone GNSS positioning is often two-fold: (1) improving GNSS processing techniques to increase solution accuracy; (2) using sensor fusion to enhance accuracy, and more importantly, resilience and availability. Compared to geodetic-grade hardware that is traditionally used in the surveying and mapping community, ultra-low-cost GNSS modules suffer from poor signal reception, low gain, or poor multipath suppression [
1,
2,
3]. On improving GNSS-only solutions, progress has been made in evaluating smartphone signal strength and carrier to noise ratios [
4,
5,
6,
7], observation noise characteristics and optimisation [
8,
9], duty cycling [
10,
11], as well as precise positioning techniques and their further enhancements [
12,
13,
14].
In 2016, Google announced the availability of code and phase raw measurements to smartphone users [
15], enabling more precise positioning techniques such as Precise Point Positioning (PPP), which is capable of delivering centimetre-level positioning in minutes with a standalone geodetic receiver and precise products [
16,
17], without the base station and baseline length constraints [
18,
19]. Most of the earliest smartphone positioning performance assessments focus on single-frequency (SF) processing, ref. [
20] conducted a single-frequency PPP static experiments, and obtained horizontal and vertical rms of 37 cm and 51 cm, respectively, under the open-sky environments. Similar conclusions were also drawn by [
21]. Following these studies, ref. [
22] utilized the ionospheric-constrained, single-frequency PPP strategy to further improve smartphone positioning, and results demonstrated that the level of sub-metre accuracy can be reached with the Mate 30 smartphone in static tests.
Thanks to the emergence of multi-constellation, dual-frequency smartphone GNSS chipsets, it is now possible to utilize more observations and manage the ionospheric delays by means of, e.g., ionospheric-free combination or ionospheric error estimation. In 2018, the first dual-frequency (DF) smartphone MI 8 was released with a BCM47755 chip [
23] and in this context, a host of studies demonstrated that the ionospheric-free dual-frequency MI 8 PPP solutions may achieve decimetre-level accuracy in static environments with real-time [
24] or final products [
25]. Continuing this research with dual-frequency processing, ref. [
26] comprehensively compared the PPP performance with four released smartphones and an average horizontal error of 40 cm can be obtained for dual-frequency MI 8 solutions, which was superior to single-frequency solutions, but the performance degraded to 6 m in a kinematic test. Recently, a subsequent contribution from [
27] demonstrated that ionospheric-constrained dual-frequency PPP is able to benefit smartphone positioning significantly compared to low-cost and geodetic-receivers, in particular in suburban environments. Meanwhile, ref. [
28] conducted a walking experiment and achieve 0.85 m and 1.09 m in horizontal and vertical components, respectively, with the aid of the real-time ionospheric products.
It is often necessary for PPP or other GNSS techniques to fuse solutions with measurements from additional sensors to maintain a similar level of accuracy during GNSS outages. For smartphone navigation, there is a stronger need due to poor GNSS measurement quality, as well as a tendency from users to acquire positioning solutions in obstructed environments such as urban canyons. Traditionally, it has been shown that during GNSS outages, low-cost MEMS (Micro-ElectroMechanical System) IMUs have the potential to achieve decimetre-level accuracy over 60 s when coupled with geodetic-grade GNSS receivers in PPP mode [
29], or at the metre-level for a few seconds with low-cost, single-frequency GNSS receivers [
30]. Fusing more recently available low-cost, dual-frequency GNSS receivers in PPP with MEMS-IMU produces decimetre-to-metre-level accuracy during 30 s of outages with four visible satellites, and dual-frequency processing gives a significant edge of 10 times improvement over single-frequency processing [
31]. The emergence of the latest smartphone-grade, or ultra-low-cost GNSS chipsets has led to studies investigating a native sensor fusion scheme using the onboard inertial sensors of smartphones [
32,
33,
34]. However, in smartphone PPP processing, there is little literature which investigates GNSS-PPP/IMU fusion specific to the ultra-low-cost GNSS receivers and IMUs. An earlier study by the authors, has shown the potential of bridging solution gaps produced from GNSS outages, while maintaining a metre-level solution [
35] using smartphones strapped on top of vehicles. It has yet to been seen if a feasible GNSS-PPP/IMU processing scheme that brings the performance to a similar level in real-world driving environments where complex obstruction and multipath profiles are involved.
In spite of this remarkable progress, the major limitations restricting the use of smartphones for precise navigation applications are their low-quality noisy measurements and positioning degradation during GNSS outages. Thus, utilizing single- and dual-frequency (SFDF) observations, as well as IMU information are vital for smartphone navigation. Ref. [
36] proposed to use the single-frequency ionosphere-corrected code measurements with dual-frequency ionospheric-free code and phase measurements for low-cost GNSS device position determination, and in this context, ref. [
37] proved that this approach would benefit smartphone positioning through walking experiments. While single- and dual-frequency PPP is not a novel concept and there are some studies focusing on SFDF scheme for low-cost and smartphone devices, it should be noted that this paper uniquely explores the benefits of SFDF strategy for smartphone not only in benign environments, but also in realistic (automotive) suburban areas where smartphone GNSS signals tend to be blocked or affected by multipath effects. For this research, this paper also employed the native inertial sensor from smartphones and developed a single- and dual-frequency PPP engine enhanced with ionospheric constraints (PPP-IC). Therefore, the main significant novelties and contributions of this work aim to answer the following research questions:
How does single- and dual-frequency PPP processing improve smartphone GNSS positioning performance and how does it compare with other PPP processing strategies (single-frequency PPP and dual-frequency PPP) in GNSS challenged environments?
How does smartphone IMU dead-reckoning perform compared to other low-cost MEMS IMU? How does the inclusion of the smartphone inertial sensor affect PPP solutions?
What is the “best” positioning performance that smartphones can achieve with multi-GNSS PPP/IMU integration in real-world driving environments?
2. Mathematical Models and Data Processing Strategies
This section introduces the theoretical background behind single- and dual-frequency PPP-IC and IMU tightly-coupled strategy, as well as the York-PPP user processing engine parameter settings.
2.1. Single-Frequency and Dual-Frequency PPP-IC Model
Ionospheric delay is a critical error source for PPP processing. The traditional ionospheric-free (IF) model is capable of eliminating the first-order ionospheric delay through the combination of dual-frequency observations [
38,
39]. Unlike the IF PPP, uncombined PPP is able to use external ionospheric delay information to benefit positioning performance. Recently, refs. [
40,
41] rewrote the uncombined PPP observation equation by decoupling the receiver DCB with estimated slant ionospheric delay. In this context, ref. [
27] suggests that ionospheric constraints (IC) bring significant benefits to low-cost devices, and this PPP-IC equation can be written as:
where
and
are the pseudorange and carrier-phase measurements on frequency
i (
);
is the geometric distance between satellite
s and GNSS receiver
r;
c is speed-of-light in vacuum, and
refer to receiver clock offsets;
and
represents the receiver and satellite code biases in IF combination, respectively; likewise,
and
represent, respectively, the receiver and satellite phase biases;
is the slant troposphere delay;
f refers to the signal frequency, and
is estimated ionospheric delay constrained by external GIM (Global Ionosphere Map) information on first frequency.
is the estimated carrier-phase ambiguity, containing code and phase biases;
is the estimated receiver DCB;
and
are, respectively, pseudorange and carrier-phase unmodelled errors including measurement noise and multipath error.
2.2. Tightly-Coupled PPP/IMU Model
The inertial sensor serves as the dead-reckoning sensor that provides relative displacement information that determines the user’s position based on the previous position. Rather than absolute, position-fixing measurements such as GNSS, an IMU requires initialization from known knowledge to compute a navigation solution. GNSS solution suffers from the inherent disadvantage of a space-based ranging system where challenging environments degrade or disable position solutions. Fusing the measurements from the native IMU within smartphones provides (1) a dead-reckoning solution from mechanization equations and (2) an improved EKF (Extended Kalman Filter) solution that employs sensor fusion. The section briefly reviews the key ideas in mechanization and EKF fusion with PPP.
Typically, in a GNSS/IMU fused system, the IMU has a higher data rate than GNSS receivers. In the case of smartphones, typically the GNSS sensor has a one-second sampling interval and a varying data frequency from IMUs due to internal power-saving or other factors within the operating system. To unify IMU input, the IMU data are pre-processed through interpolation to produce a uniform 100 Hz data stream [
35].
Mechanization enlists a series of deterministic physical equations to compute the updated attitude, velocity, and position based on known information and current IMU accelerometer/gyroscope readings. This solution, without GNSS measurements is considered the IMU-only solution in the subsequent analysis, or the dead-reckoning solution. The IMU measures specific force and angular rate as inputs to the mechanization equations. Here, the subscript b refers to the IMU body frame, and i refers to the inertial frame. In this context, the specific force measured should be interpreted as measurements in the body frame with respect to the inertial frame, resolved in the body frame axes. The inertial solution coasted from the last GNSS observation is carried onto the next available GNSS epoch for the EKF.
A tightly-coupled integration approach is used in this study. In tightly-coupled GNSS/IMU integration, the EKF takes in GNSS measurements and updates both IMU and GNSS states in a centralized approach. The total-state vector is defined as Equation (
2) [
31,
42,
43]:
where
denotes the position states;
denotes the velocity states;
denotes the attitude in local navigation frame;
and
denote the slant ionospheric and tropospheric delay, respectively;
denotes receiver clock errors;
denotes receiver clock drift;
and
denote bias in accelerometer and gyroscope, respectively; and
denotes the float ambiguity terms.
The design of the PPP/TC EKF is based on a conventional closed-loop model demonstrated in
Figure 1. The raw measurements conditioned through pre-processing are fed into the mechanization equations to produce an IMU-only solution. Once the next GNSS observation set becomes available, the GNSS-PPP/IMU EKF runs to produce a solution. Finally, the bias estimates for accelerometers and gyroscopes are fed back to the IMU data reading in the closed-loop feedback approach for the next epoch. The feedback is additive to raw
and
inputs. Since mechanization runs using the state vector inherited from the previous epoch, an accurate dead-reckoning solution requires an accurate fused prior epoch estimate. The EKF employs zero-velocity update (ZUPT) to improve solution quality when the vehicle is stationary.
2.3. York-PPP Engine and Processing Settings
The York-PPP engine is a well-established software that is able to produce PPP/IMU tightly-coupled solutions for geodetic, low-cost, as well as smartphone devices.
Table 1 highlights relevant engine settings for this current processing with corresponding products.
All smartphone collected raw measurements were processed in the aforementioned PPP-IC mode with and without IMU integration. Besides correcting the satellite orbits and clocks with GFZ (Geo-ForschungsZentrum) rapid products, the satellite DCBs are corrected using CAS (Chinese Academy of Sciences) products. The combined final IGSG GIM products served as the external constraints for ionospheric error estimation. Meanwhile, a carrier-to-noise (
) based stochastic weighting scheme was adopted since the smartphone signals contaminated with significant multipath errors tend to have lower
ratios [
5], and corresponding standard deviation of the code and phase observation
can be estimated as Equation (
3) [
48]:
where coefficient
a is 4 m and 6 cm for pseudorange and carrier-phase observations, respectively. These values are empirically derived from the residuals. In addition,
b is the pseudorange chipping length and carrier-phase wavelength for pseudorange and carrier-phase measurements, respectively [
7].
In terms of measurement quality control, the satellite elevation cutoff angle is chosen as 10° and, on average, 0.7 satellites are rejected per epoch for four constellations. This selection ensures that low-elevation satellites will be removed owing to their high noise and multipath. Additionally, the PPP engine is automatically configured to screen out satellites with post-fit residuals tenfold larger than the standard deviations of measurements to mitigate the impacts of outliers.