Multi-GNSS precise point positioning with next-generation smartphone measurements

ABSTRACT This study evaluates the feasibility of achieving improved positioning accuracy with raw GNSS measurements from recently released smartphones. Using Precise Point Positioning (PPP) as the processing mode, positioning accuracy of selected newly available devices are analyzed. These devices include Xiaomi Mi8, Google Pixel 3, Huawei Mate 20 and Samsung Galaxy S9. The key research questions to be answered are: (1) Given the hardware limitations, what PPP processing changes can be implemented to make use of the raw measurements? (2) What is the best performance that can be achieved with multi-GNSS PPP given the usage of the smartphones by the user?


Introduction
For high-precision geodetic applications, Precise Point Positioning (PPP) is well known to provide a high level of position accuracy as evidenced in various research contributions (Geng and Bock 2013, Li et al. 2013, Tang et al. 2014, Elsobeiey 2014, Laurichesse and Blot 2016, Gayatri et al. 2016).The use of the dual-frequency ionosphere-free linear combination has typically defined conventional GNSS Precise Point Positioning (PPP) processing (Zumberge et al. 1997;Héroux et al. 2001, Chen and Gao 2005, Leandro et al. 2011).However, due to satellite modernisation, the implementation of PPP has changed from the use of dual-frequency legacy measurements to a triple-frequency approach due to the addition of more signals (Hofmann-Wellenhof et al. 2007, Jan 2010, Laurichesse and Blot 2016, Aggrey and Bisnath 2017).The advancement in modernisation of Global Navigation Satellite Systems (GNSS) presents some challenges as well as opportunities with PPP research in solution accuracy, convergence, reliability and integrity (Henkel and Günther 2010, Seepersad and Bisnath 2012, 2014a, 2014b;Elsobeiey 2014).
The proliferation of smartphones over the past few decades has fuelled technological innovation in navigation applications.In contrast to decades-old mainframe computers, the performance of current mobile devices with faster processor chipsets reveals greater and better capabilities.The promise that such advancement ushers in includes lower cost rates in the development, testing and deployment of various applications.From a software perspective, mobile device applications enjoy faster update cycles, while eliminating the need for extra hardware components.There is currently over five billion Global Navigation Satellite System (GNSS) enabled devices worldwide with over 75% of such devices being smartphones (EGSA 2017).The use of location information accounts for more than 50% of applications on smartphones, either through Google Play or Apple stores (Kaplan andHegarty 2017, Sunkevic 2017).This critical need for position and timing capabilities on smartphones makes the case to understand, investigate and improve the accuracy standard.Applications requiring high precision such as car navigation, personnel monitoring and bicycle rental services are currently being hosted on smartphones.The desire for high accuracy location-based services in the mass market only serves as the impetus to advance further with improving GNSS positioning on smartphones as observed in various research contributions (Pesyna et al. 2014, Kirkko-Jaakkola et al. 2015, Banville and Van Diggelen 2016, Yoon et al. 2016, Humphreys et al. 2016, Alsubaie et al. 2017).Considering the rising use of smart wearables and phones, GNSS navigation is bound to expand the possibility of various other applications that currently may not seem feasible due to the constraints of geodetic-grade precision and accuracy.Hence, the primary goal of this paper is to test the bounds of accuracy which would address how far multi-GNSS positioning with PPP can afford the smartphone user.The challenges and benefits of processing multi-GNSS smartphone data in both static and kinematic PPP modes in open-sky and obscured urban environments will be addressed.
The introduction of the world's first dual-frequency GNSS-enabled smartphone, the Xiaomi Mi 8, has brought some excitement to academia and navigation research.The Xiaomi smartphone is equipped with a Broadcom BCM47755 chipset capable of tracking L1/E1 and L5/E5 signals from GPS, Galileo andBeiDou (EGSA 2018, Robustelli et al. 2019).Other releases such as the Huawei Mate 20 and Huawei Mate 20 Pro also have dualfrequency chipsets which track all available GNSSs.With this recent public access to raw GNSS observations on smartphone devices, a new dawn in low-cost positioning is on the horizon.Enabled by Google through its Android API, pseudorange and carrier-phase measurements can be obtained in real-time from off-the-shelf, mass-market devices (Sunkevic 2017).In difficult environments such as urban canyons, the typical expected performance of mobile devices is in the range of a few metres to many tens of metres, considering a single-frequency processing scenario.However, the use of multi-GNSS measurements from either dual-or single-frequency chipsets coupled with extra error modelling promises to enhance the solution accuracy (and initial position solution convergence) to a few decimetres.This paper addresses the advantage of the Android raw measurements in PPP processing, while providing relevant information for a robust and reliable position solution.

Evolution of navigation on smart devices
A pioneering step into the future of car navigation was realised in 1985 when Etak Inc. introduced the Etak Navigator (see supplementary file, Figure 1).This product release represented the first generation of smart device navigation.Though it offered basic features, limited geographical coverage, and few user and route options, it showed the possibilities and potential of chipset receivers.The automobile industry was then revolutionised by navigation technology (Zavoli and Bloch 1990).The dawn of personal digital assistants (PDAs) and Benefons represented the second and third generations of smart navigation devices.In contrast to their first-generation counterparts, these devices offered wireless connections with the acquisition of real-time data, but only for a limited geographical coverage area (Viken 2010, Sullivan 2012, Edwards 2018).However, fourthgeneration devices like Garmin's GPS units offered end users not only global coverage but personalised service-oriented options regulated by cellular service providers.
The academic and industrial progression in enhancing the accuracy of navigation solutions resulted in the modernisation and launching of satellite constellations, as well as refinement and advancement of various navigation algorithms.Furthermore, the rise of more capable chipsets and the transition from geodetic-grade equipment to low-cost changed the tide of navigation.Currently, the fifth generation of smart devices exist with the potential of attaining solution accuracies comparable to geodetic-grade standards.A good example of such devices is the current generation of smartphones, most of which are equipped with single-frequency chipsets.There has been various research contributions showing centimetre-level RTK positioning using single-frequency chipsets in smartphones (Mongredien et al. 2016, Odolinski and Teunissen 2017, 2019).Taking it up a notch, a potential paradigm shift has occurred with the release of dual-frequency chipsets which promises decimetre-to centimetre-level solution accuracy for the end user.Figure 1 shows and summarises the key characteristics of all the generations of smart devices.
Currently, the chipset manufacturing industry has the potential of redefining positioning accuracy with regards to low-cost equipment.2018 saw the emergence of ground-breaking dual-frequency chipsets designed for various applications, with the intention of improving GNSS solution accuracy.The smartphone market was revolutionised in September 2017 with the launch of Broadcom's BCM47755 dual-frequency chipset.This introduction sparked off an innovation and motivation to see more similar chipsets in the market (Murfin 2017).In February 2018, STMicroelectronics and U-blox launched their Teseo and F9 receiver chipsets, respectively, capable of tracking L1/L2 or L1/L5 frequencies and targeting the automotive industry (Cozzens 2018, Markowitz 2018).In the same month at the Mobile World Congress in Barcelona, Qualcomm unveiled their multi-GNSS, multi-frequency Snapdragon X24 LTE receiver with Intel also presenting a dual-frequency receiver prototype (Al Khairy 2018, Leather 2019).
To ascertain the level of accuracy that can be obtained by these recent smartphone chipsets, Trimble and Novatel investigated and achieved sub-metre positioning accuracy with BCM47755 and Tesco V chipsets, respectively (Riley et al. 2018, Abrahams 2018).These initial results demonstrated the possibility of attaining decimetre-level accuracy with smartphone chipsets.It is expected that as positioning techniques continue to improve and hardware limitations are addressed, smartphone GNSS solution accuracy can be significantly improved.There are obvious challenges that threaten the possibility of higher accuracy and these include but not limited to low-cost antenna attenuations, duty cycles and receiver power consumption.However, the recent capability added to smartphones to turn off the duty cycle unearths the optimism among application developers and end-users that such challenges are either being dealt with or will be.It is also worth noting that by turning off duty cycling, it consequentially increases the power consumption of the smartphone.

Raw measurement analysis
Four recently launched smartphones were investigated, namely Xiaomi Mi 8, Huawei Mate 20, Google Pixel 3 and Samsung S9.The characteristics of these smartphones are highlighted in Table 1.All the phones could track at least 3 out of the 4 GNSSs.Xiaomi Mi 8 and Huawei Mate 20 had the capability to track dual-frequency L1/L5 signals.
All the smartphones were assessed against a SwiftNav Piksi and Topcon NET-G3A receivers.It is noteworthy to point out the cost differences between the GNSS chipsets in the phones in comparison to geodetic-grade receiver types.In perspective, these smartphone chips cost in the 10 US dollar range, while the Piksi and Topcon receivers, belonging in the relative low-cost and geodetic categories, costs a few hundreds to thousands of dollars, respectively.With a geodetic-grade antenna, resulting in geodetic-grade measurement quality, the Piksi and Topcon receivers were used to produce reference position solutions throughout the analyses described in the subsequent sections.Unlike the dualfrequency L1/L5 smartphones, the Piksi receiver tracks the L1/L2 legacy signals from all GNSSs.However, Topcon NET-G3A tracked L1/L2 signals for only GPS and GLONASS.
In assessing the measurement quality of the smartphones, the carrier-to-noise density ratio (C/N 0 ) of the received satellite signals was observed.It is well known that one of the determining factors impacting the signal reception quality is C/N 0 , which by  definition, represents the power in the received signal compared to the power spectral density of the receiver noise (Misra and Enge 2006).Four main factors dictate the received signal power at the point antenna centre of mass: (1) power density of the incoming GNSS signal; (2) reception area of the antenna; (3) receiver antenna gain and (4) satellite elevation (Braasch and Van Dierendonck 1999).Low and irregular C/N 0 values can be observed from the smartphones over an observation period, potentially due to the cellphone-grade GNSS receiver chipset and antenna which can impact signal reception (see supplementary file, Figure 2(b-e)).In contrast, the geodetic-grade, lowcost Piksi receiver shows reliable and high C/N 0 values with a consistent average of 45 dB-Hz due to a better antenna (see supplementary file, Figure 2(a)).The delimiting factor of using either Monolithic Microwave Integrated Circuit (MMIC) or monopole antennas in smartphone designs comes at a cost in the quality of GNSS signal reception.However, C/N 0 analysis only tells part of the story in regard to the signal and measurement quality of each sensor.The combined effect of multipath and signal noise was observed to be noticeably significant (supplementary file, Figure 3.Only Xiaomi results, compared to SwiftNav Piksi, are shown because it was representative of the other smartphones analysed).For completeness, multipath observables, representing the effect of multipath on the received signal, were formed for L1/L2/L5 signal bands (supplementary file, Figure 4. Results are shown for 6 h of data for DOY 82 of 2019.The mean of the difference between C/A-code and carrier-phase measurements was removed).Table 2 summarises the RMS of the residuals of the smartphones, as compared the SwiftNav Piksi receiver.The magnitude of the multipath and signal noise level from the SwiftNav receiver was significantly lower in comparison to the smartphones.By differencing the noise levels of the two sensors in Table 2 and representing as a percentage, the multipath effect on the signals for SwiftNav Piksi was approximately 90% less than that of the effect on Xiaomi Mi 8.The signal noise level for Xiaomi Mi 8 told a similar story in comparison to the SwiftNav Piksi receiver.Given the limitations of the antenna of the smartphones, it was expected that the quality of the measurements will deteriorate in comparison to a geodetic-grade receiver and antenna (Gill et al. 2017, Gill 2018).

PPP processing customisation for smartphone measurements
Shown in Table 3 are the processing parameters and settings used in the generation of results discussed in subsequent sections.The YorkU GNSS PPP engine, developed by the GNSS research team in York University, was employed in processing the observations (Aggrey 2015, Seepersad 2012).Precise archived real-time orbits and clocks (CNT products) from Centre National d'Etudes Spatiales (CNES) were used because of the availability of orbits and clocks for all available GNSS constellations.Logging real-time data from the smartphones for PPP processing can be a challenge.It is imperative to note that not all available smartphones have the capability to log the raw GNSS data.A constantly updated list of available smart devices which log the raw GNSS measurements can be found here: https://developer.android.com/guide/topics/sensors/gnss.One prominent drawback is the duty cycling, which is used to regulate battery consumption on smartphone devices.Typically with the embedded GNSS chipsets, smartphones track and log GNSS measurements for about 200 ms before reserving power by shutting down for about 800 ms (Banville and Van Diggelen 2016).This action is critical given the possibility of cycle-slip occurrences which impact positioning techniques such as Real-Time Kinematic (RTK) and PPP.
The interaction with smartphone sensors through application programming interfaces (APIs) makes it possible for developers to extract information.The quest to obtain raw GNSS measurements has been ongoing for years before the introduction of Android's Marshmallow version.Through the APIs, NMEA (National Marine Electronics Association) formatted position, time and velocity (PVT) receiver data could be extracted (Sunkevic 2017).In 2018, Google publicly released their GNSS Analysis Tools as part of their API in the Android Nougat version, enabling the processing and analysis of GNSS measurements.This release enabled users to access the PVT and raw GNSS measurements directly.GNSSlogger, Google's android application, interfaces directly with smartphones and logs GNSS data in the NMEA format (van Diggelen and Khider 2018).Since Google's launch, other loggers have been developed by various organisations for logging smartphone GNSS data in RINEX formats.These logging applications include but are not limited to Geo++ RINEX logger (Geo++ GmbH 2018), RINEX ON (Nottingham Scientific Ltd 2018), GalileoPVT (Crosta and Watterton 2018) and G-RitZ logger (Kubo 2018).Table 4 summarises some of the key details of the current RINEX GNSS android loggers for smartphone GNSS processing.For this research work, Geo++ RINEX Logger and RINEX ON were used.

PPP smartphone data processing
In order to perform uncombined, dual-frequency PPP measurement processing, it was necessary that different sets of standard deviations had to be employed for each set of hardware.The values shown in Table 5 are derived empirically for each set of GNSS receiver hardware (Gill et al. 2017).The standard deviations are used to determine the relative weighting between the pseudorange and carrier-phase measurements.It is important to note that with the smartphones, the noise on the pseudorange measurements increases.As shown in Table 5, the magnitude of the noise on the pseudorange measurements from the smartphones is much higher compared to the other relatively low-cost and geodetic-grade receiver measurements.The higher standard deviation can also be thought in-terms of de-weighting the noisier pseudorange measurements in the estimation process, as the priori standard deviation determines the relative weighting.
In assessing the performance of the smartphones under investigation, the SwiftNav Piksi and Topcon NET-G3A dual-frequency L1/L2 receivers and geodetic-grade antennas were used to determine a reference solution.It was expected that the antenna type differences between the phones and the reference receivers would strike major distinctions in the solution qualities.The smartphone antennas are more susceptible to multipath effects from reflected signals, as shown in the previous section, due to linear polarisation.This phenomenon makes smartphone antennas more sensitive to poor quality GNSS signals, as compared to geodetic-grade antennas (Pathak et al. 2003, Zhang et al. 2018).Static and kinematic user dynamics were investigated and addressed in the subsequent sections.

Static dual-frequency PPP smartphone data processing
The static PPP analysis was performed by setting up an experiment on the roof of the Petrie Science and Engineering building at York University.Figure 2 shows the full setup of the experiment with the smartphone and SwiftNav Piksi receiver and antenna while Figure 3 illustrates the geodetic reference setup using Topcon NET-G3A receiver with a choke-ring  antenna.For clarity, it must be pointed out that these two setups were not done on the same day.The reference setup was however done around the same time as the smartphones.The purpose of using the geodetic setup was to 'ground truth' the point previously occupied by the smartphones and the SwiftNav Piksi receiver and antenna.Data were collected for a period of 6 h for DOY 82 in 2019.The SwiftNav Piksi antenna was mounted on a GNSS carbon fibre pole with the smartphones clamped to it.Though the preferred method to minimise multipath would have been to put the smartphones on the ground level, clamping was chosen to vaguely represent a user holding the phone.It must be noted that the purpose of this study is not compare smartphones against one another, but rather to assess their performance in terms of positioning.Hence, the subsequent results will have the smartphones labelled A to D. Presented in Figure 4 is the horizontal component time series of two of the smartphones (A and B) with the capability to track L1/L5 signals.Solution results were compared to SwiftNav Piksi and Topcon NET-G3A receivers.
The dual-frequency analysis for the smartphones A and B showed some interesting characteristics worth noting.Given both smartphones' ability to track a secondary frequency (L5), it was expected that they would act similarly in terms of PPP solution convergence.However, that was not the case.Defining convergence as the time taken for the solution to reach a horizontal error of 10 cm, Topcon, SwiftNav Piksi and smartphone A solutions took 12, 32 and 38 min, respectively, to converge.Smartphone B, however, did not converge.If the convergence threshold was redefined to be 50 cm in the horizontal component, it took 5 h to converge.The horizontal scatter plot (see supplementary file, Figure 5) told a similar story, showing how smartphone B had tens of metres of error in the North and East components.Given the significance of these deviations from the norm, an investigation ensued and led to some interesting conclusions.
The Geo++ logger was the default logger for all the smartphones.However, due to an Android update on smartphone B, it was not possible to use the Geo++ logger as it could not save the raw measurement logs.Hence, the RINEX ON logger was used only for smartphone B. However, the pseudoranges and carrier-phase measurements were either bad or computed wrongly for most epochs.This behaviour was not the case for the other smartphones using the Geo++ logger.To ascertain that it was a logging issue, RINEX ON logger was used on the other smartphones and the same outcome of bad or wrong measurements was observed.It was thus interesting to note that the choice of logger may potentially affect the measurement quality and ultimately, positioning performance.Given the inability to save data with Geo++ logger on smartphone B and wrongly computed measurement logs from using the RINEX ON logger, smartphone B was not included in the succeeding analysis.Hence, it must be stated that the position results from smartphone B, as presented in Figure 4, does not represent the actual results that can be obtained if there were no issues with data logging.
The exclusion of smartphone B from Figure 4 unearthed some interesting results which were unexpected.With smartphone A now being the only device now capable of L1/L5 signal tracking, it was compared to the SwiftNav Piksi and Topcon receivers.Shown in Figure 5 are the zoomed in version of Figure 4 showing the horizontal results of smartphone A, SwiftNav Piksi and Topcon NET-G3A.Though the SwiftNav Piksi converged in 32 min, as compared to 38 min for smartphone A, both results were closely comparable given a 10 cm horizontal error threshold.Even more telling was the fact there are currently about 13 satellites with L5  capability but only an average of 5 can be seen at any point in time.This fewer number of observed GPS satellites in smartphone A was expected to impact the solution performance in contrast to the reference receivers, even though other GNSSs were similarly tracked and processed.However, this good surprise and outcome from smartphone A does not tell the entire story, especially when viewed from a residual analysis perspective.
Presented in Figure 6 are the pseudorange and carrier-phase residuals for smartphone A and the reference receivers.Though they both follow typical noise-like residual trends, there are numerous jumps and outliers in the smartphone time series, and the residual magnitudes for both sensors are different in the range of tens of metres.The SwiftNav Piksi and Topcon NET-G3A pseudorange and carrier-phase residuals were at the 5 m and sub-centimetre range, respectively.Table 6 summarises the accuracy of positioning as well as the residuals for the sensors.
Though the positioning accuracy for the sensors was similar at the centimetre level, the residuals for the SwiftNav and Topcon were significantly lower than smartphone A, as expected.Residuals rms for the pseudoranges were 30 cm and 6 m for the reference receivers and smartphone A, respectively.The large residuals observed in the smartphone measurement processing can potentially be due to unmodelled hardware bias, high multipath and noise given the hardware limitations of the smartphone (Riley et al. 2018, Robustelli et al. 2019).

Static single-frequency smartphone PPP analysis
As shown in Table 1, the Samsung Galaxy S9 logs only single-frequency measurements.Complimentary to the dual-frequency results shown previously, it was necessary to process all the smartphones' data in single-frequency PPP mode.Shown in Figure 7 is the single-frequency horizontal PPP solutions for smartphones A and C as compared to SwiftNav Piksi.The single-frequency residual analysis was also investigated (supplementary file, Figure 6).
It must be noted however that smartphone D was excluded because it does not track carrier-phase measurements.In the same vein, smartphone B was also excluded due to  the reasons previously highlighted in the dual-frequency PPP analysis.The ionospheric effect was mitigated using Global Ionospheric Maps (GIM) (Gill et al. 2017, Gill 2018).
Defining the convergence threshold as 20 cm horizontal error, solutions of SwiftNav Piksi and smartphone C converged in 54 min and 2 h, respectively.The performance from SwiftNav Piksi was expected given its better hardware but the steadiness of the solution from smartphone C as well as its faster PPP initialisation, was interesting to see.Smartphone A did not converge at the defined convergence threshold.However, similar to smartphone C, its initialisation was better than SwiftNav Piksi.There were also significant deviations in the East and North components for the smartphones and SwiftNav Piksi (see supplementary file, Figure 7).The residual analysis showed very noisy carrier-phase residuals due to the usage of GIM which corrects about 75% of the ionosphere, as well as unmodelled hardware biases and noise.

Static smartphone PPP versus internal smartphone standard positioning performance
Given that the internal solutions from the smartphones, obtained from the logged NMEA logs, were standard positioning results, it was necessary to compare all the smartphones.Shown in Figure 8 are the results from the internally logged standard positioning results from smartphones A, C and D. Smartphone D logged only pseudoranges, hence it was processed in pseudorange-only PPP mode.As evidenced in Figure 9, the PPP solutions were less noisy with less East and North component deviations in contrast to the internal solutions from the smartphones.

Kinematic PPP smartphone data processing
In order to assess the performance of the smartphones in kinematic mode with multi-GNSS PPP processing technique, a test experiment was setup with a pre-determined route with data collected for 45 min.The test route was near York University and covered approximately 17 km.Setting up the smartphones involved stabilising them using car phone mounts on the dashboard of a vehicle.The SwiftNav Piksi was used as a reference solution and had its antenna mounted on the roof of the vehicle with a magnetic mount.Shown in Figure 10 is the entire setup.
From a bird's eye view as shown in Figure 11(a), all the smartphones appear to be on track except for a few deviations from smartphone A during some traffic stops.Given smartphone A was the only device with L1/L5 capability being processed in L1/L5 PPP mode, the number of processed satellites was significantly reduced.High multipath also contributed to the deviations noticed for smartphone A. Again, for reasons previously stated, smartphone B was excluded from this experiment.However, a critical look at the performance of the smartphones, as shown in Figure 11(b) indicates some interesting behaviour.The SwiftNav Piksi geodetic receiver and antenna showed the best results as expected for the obvious reasons of more capable hardware.Smartphone D, being the only pseudorange-only tracking device, showed tens of metres deviation in comparison to the geodetic reference receiver.However, smartphones A and C, L1/L5 and L1/C1 capable, respectively, were relatively comparable, on wide stretches of the road, to the SwiftNav Piksi.However, that was not the case when the car followed sharp bends or roundabouts.Shown in Figure 12 are the performances of the smartphones through a roundabout in comparison to the geodetic reference receiver.
It was interesting to see how well the SwiftNav Piksi performed, given it was a relatively low-cost receiver with geodetic-grade components.The performance of the smartphones, however, generally deteriorated around the bends and roundabouts due to significant drops in satellite number which is a result of the ever-changing satellite geometry.With the smartphones being inside the car, this was expected, along with significant multipath.
The residuals from the SwiftNav Piksi receiver (supplementary file, Figure 8) was well behaved and consistent with what was expected, given the dynamics of the car and process noise employed in the data processing.However, the pseudorange and carrier-phase residuals for all the smartphones were a magnitude of order worse in comparison to the SwiftNav Piksi receiver (supplementary file, Figure 8).As shown in Table 7, the pseudorange residuals for all the smartphones were consistent at the few metres level.The positioning accuracy for the smartphones were all at the metre-level, compared to decimetre range for SwiftNav Piksi.

Conclusions
The positioning results from the dual-frequency chipsets showed improvements in accuracy and reductions in convergence time over a 6-h period and especially over the first few minutes.The smartphones' performance was compared with that of a geodetic and relative low-cost receiver.Experiments were conducted in multiple different scenarios with the aim of testing the smartphones under different multipath profiles.For static PPP results, a static  setup enabled the collection of 6 h of data.Kinematic solutions were also obtained by using a car on a selected route.It was observed that the smartphones raw measurements showed higher multipath profiles and lower C/N 0 compared to the geodetic and low-cost receivers.
Given the quality of the raw measurements, appropriate measurement weighting and software augmentation was made while the convergence threshold in our York PPP engine was set to 10 and 20 cm for dual-and single-frequency modes, respectively.An elevation and C/N 0 threshold of 10°and 15 dB-Hz were also used, respectively, as checks to improve data quality.Single-frequency PPP processing with smartphones showed an average horizontal rms error of 60 cm.However, with dual-frequency multi-GNSS PPP processing, the horizontal rms error was an average of 40 cm.The good news of having duty cycle turned off for the smartphones provided consistent measurements for kinematic solutions.For both static and kinematic scenarios, decimetre-level to metre-level horizontal error was achieved, respectively.In summary, the paper addressed and answered the following research objective questions: (1) What is the typical performance when using the raw measurements from these smartphones in multi-GNSS PPP processing either in static or kinematic scenarios?In static GNSS PPP processing with a smartphone equipped with a dual-frequency chipset, it is possible to obtain decimetre-level accuracy in 38 min, comparable to geodetic-grade receiver and antenna.A kinematic scenario showed an accuracy of a few metres.
(2) Given the limitation that the hardware components of the smartphones present, what PPP processing changes can be implemented to make use of the raw measurements from smartphones?As shown in this research contribution, the signal noise level on the raw measurements from smartphones can be approximately 90% more than that of a geodetic-grade equipment.The C/N 0 of the satellites tracked can be significantly low as 10 dB-Hz.It is thus imperative to implement necessary measurement weighting schemes to accommodate these measurements in PPP processing.
(3) What is the best performance that can be achieved with multi-GNSS PPP given the usage of the smartphones by the user?Through an experimental setup, the GNSS PPP solution converged in 38 min, assuming a horizontal error threshold of 10 cm.The 10 cm error threshold represents a strict tolerance for very precise applications.The convergence is expected significantly improve if the threshold is increased further depending on the user application.

Future work
The advent of dual-frequency chipsets in smartphones enables the capability of improving accuracy in various potential user applications such as location-based services, augmented reality apps, gaming, etc. Future work will attempt the very challenging problem of resolving carrier-phase integer ambiguities, apply external GNSS antennas, and further integration with other sensors.More field tests are also intended to be conducted with different processing tunings in PPP.Real-time PPP processing and its performance will also be investigated with the aim of achieving better solution quality.

Figure 1 .
Figure 1.Characteristics of the generations of smart navigation devices (based on Karimi 2011).

Figure 8 .
Figure 8. Internal standard positioning solutions for smartphones A, C and D.

Figure 9 .
Figure 9. PPP positioning solutions for smartphones A, C and D.
Figure 11.Results for kinematic run: (a) general trajectory (b) detailed zoomed in portion of road showing the performance of the smartphones and SwiftNav Piksi receiver.

Figure 12 .
Figure 12. Results for kinematic run: (a) general trajectory (b) detailed zoomed in portion of road.

Table 1 .
Test smartphones and the supported raw GNSS measurements.

Table 2 .
rms for the signal noise and multipath effect on Xiaomi Mi 8 and SwiftNav Piksi.

Table 3 .
YorkU PPP processing parameters for the smartphones.Phase centre offsets and variations for GPS L5, Galileo and BeiDou satellites were not corrected due to unavailability of corrections in the current ANTEX release.Smartphone and SwiftNav Piksi antenna corrections were also not corrected for the same afore-mentioned reason. a

Table 4 .
Currently existing sample RINEX GNSS loggers for smartphones.

Table 5 .
A-priori standard deviations of pseudorange and carrierphase measurements used in PPP processing for SwiftNav Piksi, Topcon NET-G3A geodetic-grade receiver and the smartphones.

Table 6 .
Accuracy of positioning and pseudorange and carrier-phase post-fit residuals (in metres) for SwiftNav Piksi, Topcon NET-G3A and smartphone A.

Table 7 .
Accuracy of positioning and pseudorange and carrier-phase post-fit residuals (in metres) for SwiftNav Piksi and smartphones in kinematic data processing.