University of Otago – Ōtākou Whakaihu Waka researchers have developed algorithms that improve the precision of location tracking in smartwatches.
Led by Associate Professor Robert Odolinski, a visiting researcher with Google from Otago’s School of Surveying, the research team demonstrated that a smartwatch determined its location with centimeter-level precision over four hours with a stationary setup. The result was achieved by using the Google GnssLogger app and combining precise signals from several GNSSs.
The research was done in collaboration with Google’s Android Context group and the Chinese Academy of Sciences. Results are published in the scientific journal GPS Solutions.
For decades, achieving centimeter-level positioning has required industries such as surveying, construction and engineering to invest in expensive GPS equipment.
“While the use of the so-called carrier-phase signals has long been known to improve the positioning performance, the specialized antenna and receivers needed for this have traditionally come at a cost far beyond the reach of many who would benefit from the technology. This is just the beginning of what wearable high-precision positioning can potentially achieve.”
GPS was introduced in a wearable watch in 1999, but hardware and power consumption limitations prevented it from tracking the carrier-phase signals needed for high-precision results. Recent advances in smartwatches now make this possible.
Precise centimeter-level positioning on a smartwatch during 4 hours of data in Dunedin, New Zealand. The dots show the repeatability of one second of data in comparison to precise benchmark coordinates. The repeatability of the positioning is about 8 cm, at most twice as large as the smartwatch diameter of 4 cm (displayed to scale).
NTNU researchers have built SmartNav, a system that overcomes urban GPS errors using satellite corrections and Google’s 3D data. It achieves near-centimeter precision, paving the way for safer, more reliable self-driving cars.
Researchers at the Norwegian University of Science and Technology (NTNU) have created SmartNav, combining satellite corrections, wave analysis, and Google’s 3D building data for remarkable precision. Their method achieved accuracy within 10 centimeters during testing, and could make reliable urban navigation accessible and affordable worldwide, including autonomous vehicles.
“Cities are brutal for satellite navigation,” explained Ardeshir Mohamadi. “In cities, glass and concrete make satellite signals bounce back and forth. Tall buildings block the view, and what works perfectly on an open motorway is not so good when you enter a built-up area.”
Mohamadi, a doctoral fellow at NTNU, is researching how to make affordable GPS receivers much more precise without depending on expensive external correction services. “For autonomous vehicles, this makes the difference between confident, safe behavior and hesitant, unreliable driving. That is why we developed SmartNav, a type of positioning technology designed for urban canyons,” Mohamadi said.
To solve this problem, the researchers combined several technologies to correct GPS signals, resulting in a computer program that can be integrated into the navigation system of autonomous vehicles. The software developed by the researches uses PPP-RTK (precise point positioning – real-time kinematic), which combines precise corrections with satellite signals. The European Galileo system now supports this by broadcasting its corrections free of charge.
An assist from Google
Meanwhile, Google launched a new service for its Android customers that provides 3D models of buildings in almost 4,000 cities around the world. The company is using these models to predict how satellite signals will be reflected between the buildings, allowing users to see if they are walking on the correct side of he street.
The researchers were able to combine all these different correction systems with algorithms they had developed. When they tested it in the streets of Trondheim, they achieved an accuracy better than 10 centimeters 90 percent of the time.
The use of PPP-RTK will also make the technology accessible to the general public because it is a relatively affordable service.
“PPP-RTK reduces the need for dense networks of local base stations and expensive subscriptions, enabling cheap, large-scale implementation on mass-market receivers,” Mohamadi said.
“Seen & Heard” is a monthly feature of GPS World magazine, traveling the world to capture interesting and unusual news stories involving the GNSS/PNT industry.
Free navigation history course
Photo: Harvard University
Harvard University offers PredictionX: Lost Without Longitude, a free online course that examines the evolution of navigation from ancient methods to modern technologies. The program explores the science and history of navigation, focusing on the challenges of determining longitude before GPS existed. It highlights key advancements, such as John Harrison’s marine chronometer and the Longitude Prize. Through multimedia content — including videos, infographics and Worldwide Telescope tours — the course is designed to demonstrate how centuries of advancement in navigation enabled humanity to achieve milestones such as landing on the moon.
Self-driving cars collect geospatial data
Photo: Finnish Geospatial Research Institute
In Finland, self-driving cars are being used to collect geospatial data to address urban challenges. The ARVO autonomous vehicle from the Finnish Geospatial Research Institute is equipped with high-precision sensors that map its environment in real-time, collecting information on road conditions, urban vegetation as carbon sinks and factors influencing flood risks. In partnership with Aalto University and funded by the European Regional Development Fund, this initiative seeks to explore various uses of this data, such as city planning, environmental monitoring and infrastructure management.
Stopping scammers
Photo: Carlos Alvarez / iStock Editorial / Getty Images Plus / Getty Image
Google has taken legal action against a network of scammers responsible for creating more than 10,000 fake business listings on Google Maps. The scammers fabricated profiles targeting urgent service industries and bolstered them with fake reviews to appear credible. Victims were misled into contacting these fake businesses, which then sold their personal information as “leads” to legitimate service providers without consent. Google has removed the fake listings and is suing individuals involved in the scheme, CBS News reported.
Mapping Uganda’s disappearing tropical glaciers
Photo: guenterguni / E+ / Getty Image
Project Pressure, in collaboration with UNESCO and the Uganda Wildlife Authority, conducted an expedition to the Rwenzori Mountains to map the region’s disappearing tropical glaciers. The team created the first 3D model of Mt. Stanley’s glaciers and installed monitoring equipment, revealing that Mt. Speke and Mt. Baker have lost their glaciers entirely, while the Stanley Plateau Glacier has shrunk by 29.5 percent since 2020 and is heavily fragmented. The project aims to continue monitoring the glacial retreat, develop mitigation strategies and engage the local community in ongoing research.
Google has released three Google Maps application programming interfaces (APIs) for developers to map solar potential, air quality and pollen levels. The three APIs apply artificial intelligence (AI) and machine learning, along with aerial imagery and environmental data, to provide up-to-date information about these three variables, enabling developers, businesses, and organizations to build tools that map and mitigate environmental impact.
The Solar API utilizes mapping and computing resources to design detailed rooftop solar potential data available for more than 320 million buildings across 40 countries including the United States, France and Japan. To obtain this data, the AI model extracts 3D information about roof geometry from aerial imagery, while considering past weather patterns and energy costs, enabling quicker installation of solar panels.
The Air Quality API shows air quality data, pollution heatmaps, and pollutant details for more than 100 countries around the world. The API validates and organizes several terabytes of data an hour from multiple data sources — including government monitoring stations, meteorological data, sensors and satellites — to provide a local and universal index.
Google Maps uses machine learning and live traffic information to predict different pollutants in an area at a given time. The Air Quality API offers companies in healthcare, the automotive market and other forms of transportation the ability to provide accurate and timely air quality information to their users.
The Pollen API shows current pollen information for common allergens in more than 65 countries. The API provides localized pollen count data, heatmap visualizations, detailed plant allergen information, and actionable tips for allergy-sufferers to limit exposure. To obtain this information, Google Maps uses machine learning to determine where specific pollen-producing plants are located.
The Google Smartphone Decimeter Challenge (SDC) competition, co-sponsored by the Institute of Navigation (ION), took place this summer. For the competition, teams developed high-precision GNSS positioning using a pool of smartphone GNSS + inertial measurement unit (IMU) datasets accompanied by high-accuracy ground truth. Teams competed to achieve the best location accuracy with the datasets provided. Winners received cash prizes and sponsored attendance at the ION GNSS+ 2022 conference in Denver, Sept. 19-23, to present their results.
Origins
The SDC has its origins in the Android Operating System, which is an open-source platform. In 2016, Google made GNSS raw measurements available as a public application programming interface (API) on all Android phones. Since then, the available measurements have become more sophisticated and more accurate. For example, dual-frequency carrier-phase data is now available on many Android phones. This enables new areas of research.
Goals
The competition had two goals:
• Stimulate the research and development of high-accuracy algorithms that can produce submeter position accuracy on phones.
• Establish a publicly accessible repository of labeled data so that all future research on location algorithms can be judged in a consistent way against a standard set of data.
The first goal was met beyond our expectations. A total of 1,381 teams participated in the two competitions of 2021 and 2022. Discussion among competitors on the competition platform (kaggle.com) was wide-ranging, incredibly collegial, and beneficial to the entire community.
Competitors have written and shared detailed descriptions, and these have been reviewed and commented on by other competitors. Moreover, winners have written formally peer-reviewed papers and made presentations at the ION GNSS+ conferences, which are available from ion.org.
The second goal is a work-in-progress and is intended to be the legacy of the events.
Legacy
Disciplines such as machine learning have established benchmarks that make it possible to compare new approaches to previous ones in a proper quantitative way. In the GNSS community, this convention has been missing — a glance across papers at conferences will show that different algorithms tend to be presented with different test data and different metrics. Usually, the authors collect this data, and it is often fairly sparse (one or two drive tests, for example). Also, the reader never knows whether the data was cherry-picked (were bad results not mentioned?).
The SDC data provides:
• 206 different drive tests
•86 total hours of dual-frequency (L1, L5) data with code and carrier-phase measurements
•All labeled with ground-truth positions and velocities collected using NovAtel SPAN ISA-100C, with precise lever-arm compensation and validated with Google’s analysis tools.
The Kaggle site allowed users to submit their results, then automatically scored them against the ground-truth data. We advocate that all GNSS researchers use this resource to measure their location algorithm improvements in a standard way. This creates trust in published results, accelerating the recognition and adoption of truly great improvements for the benefit of the entire industry and GNSS users worldwide.
The top three winners of this year’s Smartphone Decimeter Challenge described their projects to Matteo Luccio, GPS World editor-in-chief.
Suzuki
Taro Suzuki, Chiba Institute of Technology
1st Place Winner: Two-Step Optimization of Velocity and Position using Smartphone’s Carrier Phase Observations
What is your research focus and how does it relate to the contest?
My current research focuses on the accurate positioning of vehicles and mobile robots in urban environments where GNSS multipath occurs. I usually use commercial GNSS receivers for my research. This competition is very relevant to my current research, except that the smartphone is replacing a receiver.
How long have you been developing the technology or approach you used to win the contest?
The competition was held for three months, but I concentrated my efforts on the past three weeks. However, I used technologies and resources developed in my previous research (for example, source code developed in last year’s competition).
Have you participated in previous editions of this contest?
Yes, I participated in the last competition and won. The approach used in this year’s competition is based on the method used to win last year’s competition, with additional innovations and improvements.
Where, in what GNSS signal conditions, and at what speeds were the test data collected?
The competition provides a training dataset, which contains raw GNSS observations from a smartphone installed on a vehicle as it travels on real roads. In addition to GNSS observations, the training dataset contains the ground truth of the smartphone’s position. The training dataset includes a wide range of GNSS signal conditions, such as driving on highways around San Francisco and Los Angeles, driving on tree-lined urban streets, and driving in tunnels and under overpasses. I have developed an algorithm that uses a training dataset containing ground truth to accurately estimate the location of smartphones in a variety of GNSS signal reception environments.
What accuracies were you able to obtain?
The competition metric was “average of 50th and 95th percentile horizontal errors.” The metrics are computed for each of the 36 runs in the test dataset, which are divided into public and private groups, then the metrics are averaged in each group to compute the final score. My final score was 1.382 m for public and 1.229 m for private. The best score given after the competition was 1.372 m for public and 1.197 m for private. The final result achieved sub-meter accuracy in the median (50th percentile).
What are the key features of your approach?
The key point of my method is global optimization using graph optimization, unlike a conventional Kalman filter or least-squares-based positioning methods. In addition, highly accurate relative position estimation using the time difference of carrier wave phases of smartphones contributed to the accuracy. Because the competition dataset included environments such as tunnels and elevated structures in which GNSS cannot be received at all, I devised an algorithm with two optimization steps (first velocity optimization, then position optimization) and applied it to the competition. This method enables highly accurate position estimation for vehicle driving data in various GNSS signal reception environments using only smartphone GNSS observation data.
What end-user applications are you expecting your approach to enable?
Decimeter-accurate location estimation could lead to lane-level navigation for vehicles, pedestrian navigation, and advanced location-based smartphone games.
Dai
Shubin Dai, Kaggle Community
2nd Place Winner: Improving Smartphone GNSS positioning using Gradient Descent Method
What is your research focus and how does it relate to the contest?
I am a data scientist and one of the top competition grandmasters on Kaggle. My research interests include computer vision, natural language processing, autonomous driving, and reinforcement learning. I placed in the top three in 14 related competitions (13 of which were solo). So, despite my lack of background knowledge in the GNSS field, these methods, skills and experiences helped me find a solution.
How long have you been developing the technology or approach you used to win the contest?
I spent about 50 days on this competition, including learning principles of GNSS and understanding all kinds of algorithms by reading books, papers and source codes. The Kaggle platform is very helpful when we want to get started in a new field.
Have you participated in previous editions of this contest?
I did not participate in the competition held last year, but I learned a lot from solutions of recent years, particularly the third-place solution.
Where, in what GNSS signal conditions, and at what speeds were the test data collected?
The benchmark datasets include raw GNSS measurement and raw readings from inertial sensors, using smartphones (Xiaomi Mi 8, Google Pixel 4, etc.) enabled with dual-frequency and ADR (accumulated delta range) in driving scenarios, collected in the San Francisco Bay area.
In the GSDC2021 dataset, there are 29 drives with 73 phone GNSS logs in the training set and 19 drives with 48 phone logs in the test set. Compared to 2021’s competition, in the GSDC2022 dataset we can see more data overall and a wider variety of routes: 62 drives with 170 phone logs are provided in the training set and 36 drives with only one phone per drive are provided in the test set.
The drives in the training set took 15 to 60 minutes at an average speed of 18 m/s.
What accuracies were you able to obtain?
According to the metric of this competition, the score is calculated as the mean of the 50th and 95th percentile distance errors. The score on my local validation set is 1.929 m, the score on the public test set is 1.608 m, and the score on private test set is 1.499 m. When we calculate the mean error, the score is 1.401 m on a validated set, the mean error of 40% of the trips are under 1 m. I think the competition metric is more reliable as the 95th percentile distance error is also important.
By the way, my local validation set is more difficult to optimize than the test set, so the mean error on the test set is expected to be lower than 1.401 m.
What are the key features of your approach?
The competition data is noisy due to multipath effects, non-line-of-sight receptions, receiver noise and missing data, therefore it’s quite challenging. I found that the optimal estimation for each point locally is not stable and can be affected by noise at that point on the track. If we can find a solution to a whole track globally, the noise can be reduced as the model must follow all kinds of constraints, such as geometry constraints, speed constrains, and global acceleration constraints.
Although we could extend the WLS and Kalman filter solution to take more points on a track into consideration, it’s not so easy to model all kinds of constrains. On the other hand, if we use a global optimization method, such as factor graph optimization and neural networks, we can add the constrains easily, which makes it more efficient to conduct experiments.
Following the solution of the third-place winner in last year’s competition, I used the global optimization method by taking into account gradient descent, pseudorange, pseudorange rate, accumulated carrier phase (ADR), phone speed and acceleration constraints of every time epoch on a track. When optimizing the track using gradient descent, the losses are designed to filter out abnormal data and reduce the noise by a series of physical and geometrical rules. I spent much time searching for the constraints, proving them and turning them into losses that can be used to update the coordinates iteratively during the competition.
What end-user applications are you expecting your approach to enable?
According to the setting of this competition, we can post-process data collected using Android phones, which is easily obtained. The track obtained can then be optimized using the solutions from this competition. The solutions from the first and the second place can both be considered as a framework that can be extended by adding more constrains to it to improve accuracy.
Everett
Tim Everett, RTK Consultants LLC
3rd Place Winner: An RTKLIB Open-Source-Based Solution
What is your research focus and how does it relate to the contest?
I develop and maintain the demo5 fork of the popular RTKLIB open-source GPS/GNSS software tool. I have optimized this software for low-cost precision GNSS solutions, so it is very closely related to the goals of this competition. My background is in control system theory and I worked in product and technology development for servo systems in the disk drive industry for 25 years before switching to the GNSS field. The mathematics turns out to be quite similar between the two as both are problems in precision positioning, just different in scale. In disk drives, it is nanometers over centimeters and in precision GNSS, it is centimeters over kilometers.
How long have you been developing the technology or approach you used to win the contest?
I have been developing and maintaining low-cost precision GNSS solutions in the RTKLIB software for about six years but have only worked with smartphone solutions in the last year or two.
Have you participated in previous editions of this contest?
I did not participate in last year’s competition but I did work with the data after the contest was over and shared a solution using RTKLIB that would have placed fifth in the competition.
What accuracies were you able to obtain?
I achieved a score of 1.648 m on the private leaderboard. This represents the average of the 50th percentile and the 95th percentile of the errors as scored by Kaggle. Kaggle does not provide any further breakdown of this number but, based on the training data for which ground truths were provided, this corresponded to a 50th percentile error of roughly 0.9 m and a 95th percentile error of roughly 2.3 m. With a small tweak to my solution after the competition was over, I was able to improve my private leaderboard score to 1.593 m, which would have been within 1 cm of the third-place solution.
What are the key features of your approach?
My approach was to use the existing post-processing kinematic (PPK) solution algorithm in RTKLIB but to reoptimize it for the unique characteristics of the smartphone observation data. A PPK solution is the post-processing equivalent of a real-time kinematic (RTK) solution and is a differential solution that relies on differencing the receiver observations with observations from a nearby base station to cancel out most of the largest error sources — including atmospheric, orbital and clock errors — since these errors are common between the two sets of proximate observations.
Because smartphones have very poor GNSS antennas and they were mounted inside vehicles, the signal quality is much lower and the multipath much greater than those for which the RTKLIB algorithm was optimized. In addition, the smartphones were using the L5 frequency band, whereas RTKLIB was optimized for the more commonly used L2 frequency band. One of the main goals of my optimization process was to include many low-quality observations in the solution that would normally be discarded, but to de-weight them appropriately.
What end-user applications are you expecting your approach to enable?
RTKLIB software is currently used to provide precision solutions for many end-user applications such as surveying, drone photogrammetry, sports tracking, precision agriculture, utility location, marine navigation and ground subsistence monitoring. Although smartphones won’t replace dedicated low-cost GNSS receivers, the challenging nature of the smartphone data severely stresses the RTKLIB algorithms and exposes numerous opportunities for improvement that are much less obvious with more typical, higher quality data. I have pulled these improvements into the main branch of the demo5 version of RTKLIB, and hence this work should immediately improve the quality of all these applications and extend their use into more challenging environments.
Photo: Google
Acknowledgements: Thanks to the Institute of Navigation (ION) for co-sponsoring the 2022 Smartphone Decimeter Challenge. Thanks to Luke Walcher and Tolu Ojelade for their contributions to the photos used in this article.
Qualcomm Technologies has unveiled new wearable platforms, the Snapdragon W5+ Gen 1 and Snapdragon W5 Gen 1.
The platforms are designed to advance ultra-low power and breakthrough performance for next-generation connected wearables with a focus on extended battery life and premium user experiences. They incorporate a series of innovations including low power islands for GNSS, Wi-Fi and audio; ultra-low power Bluetooth 5.3 architecture; and low power states such as Deep Sleep and Hibernate.
New enhancements to the flagship Snapdragon W5+ platform offer 50% lower power, 2X higher performance, 2X richer features, and 30% smaller size, compared to our previous generation, enabling wearable manufacturers to deliver the differentiated experiences consumers demand. Based on the hybrid architecture, the purpose-built platform is comprised of a 4 nm-based system-on-chip and 22 nm-based highly integrated always-on co-processor.
By using these platforms, manufacturers can scale, differentiate and develop products faster in the continuously growing and segmenting wearables industry, Qualcomm said.
Qualcomm also announced two reference designs from Compal and Pegatron, which showcase the capabilities of the platform and the company’s collaboration with ecosystem partners, helping customers develop products faster.
Bird uses the ARCore Geospatial API to enable a scooter parking app. (Image: Bird)
Google has launched the ARCore Geospatial API in ARCore software development kits (SDKs) for Android and iOS across all compatible ARCore-enabled devices.
The application programming interface (API) is available at no cost to download and opens up nearly 15 years of Google Maps data to help developers build more useful and immersive augmented reality (AR) experiences.
“The Geospatial API provides access to global localization — the same technology that has been powering Live View in Google Maps since 2019, providing people with helpful AR-powered arrows and turn-by-turn directions,” explains a Google blog.
“Based on the Visual Positioning Service (VPS) with tens of billions of images in Street View, developers can now anchor content by latitude, longitude and altitude in more than 87 countries, without being there or having to scan the physical space, saving significant time and resources.
“For end users, discovering and interacting with AR is faster and more accurate as images from the scanned environment are instantaneously matched against our model of the world,” the blog states. “This model is built using advanced machine-learning techniques, which extract trillions of 3D points from Street View images that are then used to compute a device’s position and orientation in less than a second.
“In other words, users can be anywhere Street View is available, and just by pointing their camera, their device understands exactly where it is, which way it is pointed and where the AR content should appear, almost immediately.”
Early-access partners include the NBA, Snap and Lyft, who are exploring and building applications in areas such as education, entertainment and utilities. For example, micromobility companies Bird, Lime and WeMo are using the API to remove friction from parking e-scooters and e-bikes, adding pinpoint accuracy so that riders know exactly when their vehicle is in a valid parking spot. Lime has been piloting its app in London, Paris, Tel Aviv, Bordeaux, Madrid and San Diego.
Originally posted in the Android Developers Blog, the following is reprinted with permission from authors Frank van Diggelen, principal engineer, and Jennifer Wang, product manager, Google.
At Android, we want to make it as easy as possible for developers to create the most helpful apps for their users. That’s why we aim to provide the best location experience with our APIs like the Fused Location Provider API (FLP). However, we’ve heard from many of you that the biggest location issue is inaccuracy in dense urban areas, such as wrong-side-of-the-street and even wrong-city-block errors.
This is particularly critical for the most-used location apps, such as rideshare and navigation. For instance, when users request a rideshare vehicle in a city, apps cannot easily locate them because of the GPS errors.
The last great unsolved GPS problem
This wrong-side-of-the-street position error is caused by reflected GPS signals in cities, and we embarked on an ambitious project to help solve this great problem in GPS. Our solution uses 3D mapping aided corrections, and is only feasible to be done at scale by Google because it comprises 3D building models, raw GPS measurements, and machine learning.
The December Pixel Feature Drop adds 3D mapping aided GPS corrections to Pixel 5 and Pixel 4a (5G). With a system API that provides feedback to the Qualcomm Snapdragon 5G Mobile Platform that powers Pixel, the accuracy in cities (urban canyons) improves spectacularly.
Image: Frank van Diggelen
Image: Frank van Diggelen
Pictures above show a pedestrian test, with Pixel 5 phone, walking along one side of the street, then the other. Yellow = Path followed, Red = without 3D mapping aided corrections, Blue = with 3D mapping aided corrections.
Why hasn’t this been solved before?
The problem is that GPS constructively locates you in the wrong place when you are in a city. This is because all GPS systems are based on line-of-sight operation from satellites. But in big cities, most or all signals reach you through non line-of-sight reflections, because the direct signals are blocked by the buildings.
Diagram of the 3D mapping aided corrections module in Google Play services, with corrections feeding into the FLP API. 3D mapping aided corrections are also fed into the GNSS chip and software, which in turn provides GNSS measurements, position, and velocity back to the module. (Image: Frank van Diggelen)
Image: Frank van Diggelen
The GPS chip assumes that the signal is line-of-sight and therefore introduces error when it calculates the excess path length that the signals traveled. The most common side effect is that your position appears on the wrong side of the street, although your position can also appear on the wrong city block, especially in very large cities with many skyscrapers.
There have been attempts to address this problem for more than a decade. But no solution existed at scale, until 3D mapping aided corrections were launched on Android.
How 3D mapping aided corrections work
Image: Frank van Diggelen
The 3D mapping aided corrections module, in Google Play services, includes tiles of 3D building models that Google has for more than 3,850 cities around the world. Google Play services 3D mapping aided corrections currently supports pedestrian use-cases only. When you use your device’s GPS while walking, Android’s Activity Recognition API will recognize that you are a pedestrian, and if you are in one of the 3,850+ cities, tiles with 3D models will be downloaded and cached on the phone for that city. Cache size is approximately 20MB, which is about the same size as 6 photographs.
Inside the module, the 3D mapping aided corrections algorithms solve the chicken-and-egg problem, which is: if the GPS position is not in the right place, then how do you know which buildings are blocking or reflecting the signals? Having solved this problem, 3D mapping aided corrections provide a set of corrected positions to the FLP. A system API then provides this information to the GPS chip to help the chip improve the accuracy of the next GPS fix.
With this December Pixel feature drop, we are releasing version 2 of 3D mapping aided corrections on Pixel 5 and Pixel 4a (5G). This reduces wrong-side-of-street occurrences by approximately 75%. Other Android phones, using Android 8 or later, have version 1 implemented in the FLP, which reduces wrong-side-of-street occurrences by approximately 50%. Version 2 will be available to the entire Android ecosystem (Android 8 or later) in early 2021.
Android’s 3D mapping aided corrections work with signals from the USA’s GPS as well as other GNSS: GLONASS, Galileo, BeiDou, and QZSS.
Our GPS chip partners shared the importance of this work for their technologies.
Francesco Grilli, vice president of product management at Qualcomm Technologies Inc.:
“Consumers rely on the accuracy of the positioning and navigation capabilities of their mobile phones. Location technology is at the heart of ensuring you find your favorite restaurant and you get your rideshare service in a timely manner. Qualcomm Technologies is leading the charge to improve consumer experiences with its newest Qualcomm Location Suite technology featuring integration with Google’s 3D mapping aided corrections. This collaboration with Google is an important milestone toward sidewalk-level location accuracy.”
Charles Abraham, senior director of engineering, Broadcom Inc.:
“Broadcom has integrated Google’s 3D mapping aided corrections into the navigation engine of the BCM47765 dual-frequency GNSS chip. The combination of dual frequency L1 and L5 signals plus 3D mapping aided corrections provides unprecedented accuracy in urban canyons. L5 plus Google’s corrections are a game-changer for GNSS use in cities.”
Yenchi Lee, deputy general manager of MediaTek’s Wireless Communications Business Unit:
“Google’s 3D mapping aided corrections is a major advancement in personal location accuracy for smartphone users when walking in urban environments. MediaTek’s Dimensity 5G family enables 3D mapping aided corrections in addition to its highly accurate dual-band GNSS and industry-leading dead reckoning performance to give the most accurate global positioning ever for 5G smartphone users.”
How to access 3D mapping aided corrections
Android’s 3D mapping aided corrections automatically works when the GPS is being used by a pedestrian in any of the 3850+ cities, on any phone that runs Android 8 or later. The best way for developers to take advantage of the improvement is to use FLP to get location information. The further 3D mapping aided corrections in the GPS chip are available to Pixel 5 and Pixel 4a (5G) today, and will be rolled out to the rest of the Android ecosystem (Android 8 or later) in the next several weeks. We will also soon support more modes including driving.
Android’s 3D mapping aided corrections cover more than 3850 cities, including:
North America: All major cities in USA, Canada, Mexico.
Europe: All major cities. (100%, except Russia & Ukraine)
Asia: All major cities in Japan and Taiwan.
Rest of the world: All major cities in Brazil, Argentina, Australia, New Zealand, and South Africa.
As our Google Earth 3D models expand, so will 3D mapping aided corrections coverage.
Google Maps is also getting updates that will provide more street level detail for pedestrians in select cities, such as sidewalks, crosswalks, and pedestrian islands. In 2021, you can get these updates for your app using the Google Maps Platform. Along with the improved location accuracy from 3D mapping aided corrections, we hope we can help developers like you better support use cases for the world’s 2B pedestrians that use Android.
Continuously making location better
In addition to 3D mapping aided corrections, we continue to work hard to make location as accurate and useful as possible. Below are the latest improvements to the Fused Location Provider API (FLP):
Developers wanted an easier way to retrieve the current location. With the new getCurrentLocation() API, developers can get the current location in a single request, rather than having to subscribe to ongoing location changes. By allowing developers to request location only when needed (and automatically timing out and closing open location requests), this new API also improves battery life. Check out our latest Kotlin sample.
Android 11’s Data Access Auditing API provides more transparency into how your app and its dependencies access private data (like location) from users. With the new support for the API’s attribution tags in the FusedLocationProviderClient, developers can more easily audit their apps’ location subscriptions in addition to regular location requests. Check out this Kotlin sample to learn more.
Qualcomm and Snapdragon are trademarks or registered trademarks of Qualcomm Incorporated. Qualcomm Snapdragon and Qualcomm Location Suite are products of Qualcomm Technologies Inc. and/or its subsidiaries.
Movement, closeness, privacy — many things we took for granted a few months ago have become luxuries after the onset of COVID-19. To get an understanding of the scale and impact of the virus, we can look at global movement trends of people and merchandise using GNSS technology.
Marine
Before the coronavirus pandemic, globalization seemed to be increasing endlessly. Now, we face new trade restrictions, protectionist policies, and a global economic downturn that threatens to stunt growth for years to come.
In April, the World Trade Organization (WTO) forecast that global trade would fall by between 13% and 32% in 2020, surpassing the “great trade collapse” of 2009 spurred by the global financial crisis.
However, the situation isn’t completely bleak. According to a recent Tradeshift report, global trade decreased by 14.8% in the second quarter of 2020, putting us on the optimistic end of the WTO estimate. June saw a rise in trade activity, suggesting that we may be recovering from the initial effects of the pandemic.
Cargo vessels, tankers, tugs, and other kinds of commercial ships are equipped with satellite navigation devices that can receive information from GNSS satellites to compute precise location and time. Maritime tracking insights obtained via GNSS/GPS signals are a great method for measuring the impact of the coronavirus on trade.
Photo: shaunl/E+/Getty Images
GPS data from MarineTraffic shows that ship arrivals decreased in nearly all of China’s ports from January to March.
Source: MarineTraffic
Just as trade shipping began to pick up in China, the United States and Europe were hit hard by the pandemic. However, shipments have now begun to climb worldwide to compensate for cancellations earlier this year.
The cruise industry, on the other hand, shows little indication of recovery. Cruises were the fastest growing segment of the travel industry over the past five years — until the pandemic hit.
On June 19, the Cruise Line International Association (CLIA) and Centers for Disease Control and Prevention (CDC) announced a “no sail order” for cruise ships. The order has been extended through Sept. 30. Furthermore, the cruise industry’s reputation has been damaged by multiple outbreaks on ships, most recently the MS Roald Amundsen of Norway. At least 43 people were infected on the MS Roald Amundsen, and Norway has now banned cruise ships with more than 100 passengers from disembarking at Norwegian ports.
Maritime intelligence company VesselsValue is using AIS data to map cruise ship activity throughout the pandemic. They have noted that while cruise ships typically sail at 13.5 knots to 15 knots, average speed has dropped to 11 knots in 2020 as ships attempt to lower fuel costs. Port-to-port sailings have declined for the 10 most popular cruise routes.
Source: VesselsValue
Source: VesselsValue
Most of the port-to-port sailings that constitute the 2020 columns in the above graph actually represent ships being repositioned for a break in service, also called a “layup.” Layups can cost cruise companies millions of dollars per month, but with no sail orders and port closures, they are a necessary expense.
According to the Oxford COVID-19 Government Response Tracker, a real-time monitoring system that evaluates government policies, April marked the strictest lockdown measures across all 133 available coastal countries. The tracker uses a stringency index between 0 and 100 based on national containment and closure policies.
Source: Oxford COVID-19 Government Response Tracker
Air
Commercial air traffic has decreased as well. According to GPS flight tracking service Flightradar24, the number of global daily flights was slashed by nearly two thirds between March and April. While there were 15,012 flights in the air at 15:00 UTC on March 7, there were only 5,275 at the same time on April 7.
Global air traffic March 7, 2020. (Source: Flightradar24)
Global air traffic April 7, 2020. (Source: Flightradar24)
There were 55% fewer flights in the last week of March 2020 than in the last week of March 2019. While all types of air traffic have been reduced for fears of infection, the coronavirus has especially decimated demand for passenger flights. Passenger airlines across the world have canceled flights and cut capacity for the foreseeable future. Some passenger airlines have even switched to transporting cargo in a desperate attempt to avoid bankruptcy.
Looking at regional tracking data can further illuminate the impact of COVID-19 on air travel. Travel restrictions and border closures were enacted en masse in late March, though some nations adopted more stringent policies than others.
Europe
The number of flights in Europe has plummeted since the onset of the coronavirus pandemic. There were 2,400 fewer flights in Europe on April 7 than March 7.
Air traffic over Europe March 7, 2020. (Source: Flightradar24)
Air traffic over Europe April 7, 2020. (Source: Flightradar24)
North America
Though air travel has been disrupted across North America, the United States remains significantly more busy than its neighbors. The March 7 image shows 8,400 flights while the April 7 image shows 2,950 flights, most of them concentrated over the United States.
Air traffic over North America March 7, 2020. (Source: Flightradar24)
Air traffic over North America April 7, 2020. (Source: Flightradar24)
East Asia
The decrease in air traffic over East Asia has been severe and persistent. Since China and its neighbors began to experience a trade downturn as early as the third week of January, we can compare January 7 to April 7 to capture the effects of the coronavirus. As of late July, air traffic at China’s busiest airports was still down approximately 60% from normal levels.
Air traffic over East Asia January 7. (Source: Flightradar24)
Air traffic over East Asia April 7. (Source: Flightradar24)
Ground
Google has released — and continues to update — a series of community mobility reports that chart movement trends in public spaces. The reports are a compilation of GPS data for Google users across the world.
Mobility changes are particularly stark for regions hit hard by the coronavirus. New York state showed a 46% reduction in visits to transit stations and a 42% reduction in visits to workplaces in June and July compared to pre-pandemic baseline levels. New Yorkers are also visiting parks 74% more often.
Screenshot: Google
Screenshot: Google
Trends in the United States as a whole are also dramatic.
Screenshot: Google
Screenshot: Google
Israeli journey-planning app Moovit is using mobile phone data to document trends in public transit ridership. Many of the world’s largest metropolitan cities experienced a steep decline in ridership between mid-January and late March. Millions of people that rely on mass transit have had to cope with cumbersome rules and the danger of catching the virus itself. Efforts to reduce overcrowding on trains and buses have translated into reduced capacity requirements and therefore, lengthy wait times. The rise of remote work has also lowered public transit ridership.
Screenshot: Moovit
Italy was hit particularly hard – and early – by the pandemic. Two large outbreaks occurred in Northern Italy in late February, prompting widespread closures and government-mandated quarantines in Lombardy and 14 neighboring provinces. Public transit ridership plummeted when the quarantine took effect in early March.
Screenshot: Moovit
Usage of GPS-reliant ride-hailing apps has also dropped severely. Daily installs of China’s three biggest ride-hailing apps were down 75% the week of February 10 compared to the same week in 2019.
Source: Sensor Tower
American ride-hailing giants Uber and Lyft have seen similar losses. However, business for food delivery apps like Uber Eats and Grubhub are on the rise as more people stay home instead of grocery shopping.
The impacts of COVID-19 have been less severe — but still significant — for the trucking industry. The U.S. trucking industry is an economic powerhouse, typically generating over 700 billion in annual revenue and transporting 72.5% of American freight. The American Transportation Research Institute generated a truck activity index based on GPS data across six states from the week of February 9 to the week of February 12. The data shows an initial spike in trucking operations due to increased demand for consumer goods and medical supplies. However, as stay-at-home orders and restrictions ceased business operations across the country, truck activity declined.
Source: American Transportation Research Institute
Coronavirus safety restrictions adopted by countries across the world have generally begun to loosen up, for better or for worse. It will be interesting to see how the coming months unfold through the lens of GNSS data.
Roi Mitt works for Regulus Cyber, a company researching GPS cybersecurity and offering various software products to protect the integrity, reliability and security of GNSS devices. The company’s products are designed for multiple industries using GPS-based time and location, in order to ensure a future in which GNSS technology is safe and reliable to use.
Contact tracing can help stem the spread of the COVID-19 pandemic. It involves tracking the movement and interactions of infected individuals to identify others at risk.
National and regional responses to the COVID-19 pandemic have included containment through quarantine and restriction of movement. When properly implemented, these solutions limit spread of the contagion to prevent it from overwhelming healthcare and emergency management systems.
According to the World Health Organization, the Centers for Disease Control and Prevention, and virtually all medical professionals, any effective strategy to return the world to normal requires three components: testing, contact tracing and isolation.
While testing to find the people who are infected is the absolute top priority, contact tracing is vital for stopping a disease from spreading out of control. It involves tracking the movement and interactions of infected individuals to identify others at risk. Any positive test without contact tracing is bad public health — it misses an opportunity to reduce the spread of the virus.
While the concept of contact tracing has just entered popular consciousness, it has been a standard public health tool for a century. For example, in the 1930s, Great Britain used it to contain the incidence of sexually transmitted infections. In the 1960s and 1970s, South American, African and Asian countries used it to eradicate smallpox. Additional diseases for which contact tracing is performed include tuberculosis, measles, HIV, Ebola, bloodborne infections, serious bacterial infections and novel infections.
What Is Contact Tracing?
The World Health Organization describes three basic steps:
Contact identification. Those who have been in contact with someone who has been confirmed to be infected are identified, by asking about their activities and those of the people around them.
Contact listing. All persons who have had contact with the infected person are informed of their status and told to receive early care if they develop symptoms.
Follow-up. Contacts are monitored for symptoms and tested for signs of infection.
In some cases, quarantine or isolation is required for high-risk contacts.
The enormous dimensions of the current pandemic, however, challenge traditional models of contact tracing, which are very resource intensive. In search of a technological assist, several Asian countries already have been taking advantage of the functionalities of smartphones to scale up contact tracing to match the pandemic’s rate of growth, such as the Trace Together app built by the Singapore government. Companies and organizations around the world are following suit, including Britain’s National Health Service, a pan-European initiative, and an unprecedented joint venture by Apple and Google.
Automating Contact Tracing
A study published on March 31 in Science concluded that “viral spread is too fast to be contained by manual contact tracing but could be controlled if this process [were] faster, more efficient and happened at scale.” A contact-tracing app that memorizes close contacts and immediately notifies users if they have had contact with infected individuals, prompting them to self-isolate, could control the pandemic without need for mass quarantines if enough people used it, the study argues.
Privacy versus Protection. A similar app has been deployed in China, where people are required to use it to be allowed to move beyond their neighborhood, enter public spaces, or use public transport. A central database collects data on each user’s movement and coronavirus diagnosis, artificial intelligence analyzes these data, then the app displays a red, amber or green code that determines the user’s freedom of movement. This app has been credited with significantly helping China suppress the pandemic, but has been criticized for its disregard for data protection and privacy.
Relying on fundamental epidemiological principles and common smartphone functionality, the Science study authors designed a simple algorithm to replace manual contact tracing. “Coronavirus diagnoses are communicated to the server, enabling recommendation of risk-stratified quarantine and physical distancing measures in those now known to be possible contacts, while preserving the anonymity of the infected individual.” Symptomatic individuals could use the app to request testing, and everyone could use it to access COVID-19-related health services, information and instructions, or even to request deliveries of food or medicine during self-isolation.
Public trust in the app and how the gathered data are used would be critical to its success. The study’s authors lay out a series of requirements for its ethical implementation, then point out that “the algorithmic approach we propose avoids the need for coercive surveillance, since the system can have very large impacts and achieve sustained epidemic suppression, even with partial uptake.”
The authors of a similar article in the journal JMIR mHealth and uHealth write that a contact-tracing system can limit any central coordination to notifying users who have been in contact with an infected person. Their core idea is that it does not matter where someone contacts an infected person, only that they were in close enough contact to risk infection. Particularly sensitive location data, such as GPS or phone cell data, “is actually neither necessary nor useful.” No one learns who the user is because the app is not linked to an identity, and it neither records nor stores location data.
The authors argue their proposed app is the most effective epidemiologically because it would determine which people were in close proximity, and it would receive user cooperation. “Only if people trust a system — because it does not spy on them — will the system find broad support in the population.”
GPS, Bluetooth or Both?
Technologically, the concepts of location and proximity are embodied in two standard smartphone components: GPS receivers and Bluetooth transceivers. GPS-derived location data makes it possible to map and analyze the movements of individuals and of large numbers of people; for example, it is how we know that many in this country have begun relaxing social distancing rules ahead of the lifting of legal mandates. Bluetooth’s use of low energy, generally a drawback, becomes an advantage here because it can tell us whether two or more people have been within 1.5 to 2 meters of an infected person for at least 10 to 15 minutes — when the risk of infection is highest. This knowledge can enable newly infected, pre-symptomatic people to self-isolate and not infect others.
Numerous companies are developing and proposing smartphone-based contact tracing apps.
Vesedia Mobile Technologies proposes that people who test positive to COVID-19 be asked to provide information about public places they visited in the preceding days and at what times, using their phone location history for verification. The information would be anonymized by healthcare officials, and entered into a database that would be publicly accessible via a website and mobile app.
Ramesh Raskar. (Photo: MIT/John Werner)
Intersections. The COVID Safe Paths phone app and the Safe Places browser tool for contact tracers were created by Ramesh Raskar and other researchers at the MIT Media Lab. If a user tests positive and consents, his or her data is uploaded, redacted by healthcare authorities to remove any personally identifiable information, and downloaded by the app.
The app then performs “intersections” — it identifies and notifies people with whom the infected person has crossed paths. By clicking on intersections, users can display their timeline for the past 14 days, in a calendar view, which tells them how many intersections have occurred each day.
The app also provides news reports from authorized local news channels, based on each user’s position or if they tap the URL for their local healthcare authority, said Abhishek Singh, the program’s tech lead, who is helping with the app’s development.
“We are also building an interoperable architecture,” Singh said. “Because there are many contact tracing apps already in the wild, we want to make sure that they have some common standards and guidelines that enable them to utilize data from other apps securely and through consent.” More than 1,200 people are voluntarily contributing to the project. “It is being led by the open-source community, and all our source code is out there and anybody can contribute,” Singh said.
Safe Places is helping health authorities by making the data and insights visible, enabling them to make the right decisions such as targeting resources to areas that need them most, helping them impose restrictions such as lockdowns, or reopen the economy. “The economy will not reopen in a single burst, but step by step.” Singh said. “A dashboard that allows them to monitor where the infection is spreading and where it has been contained helps them decide where to take which steps.”
The GPS Advantage. The uptake required for GPS-enabled contact tracing to be successful is generally lower than for Bluetooth-based contact tracing, Singh argues, citing an Oxford University simulation. “With GPS, you do not need people to have the app already downloaded for it to be effective,” he points out. A person who tests positive for the virus can use the Safe Place web tool to manually create a GPS trail and help healthy people. This is one of the biggest advantages of GPS compared to Bluetooth, because the latter requires exchanging information directly through the hardware, which cannot be done after the fact.
Because the app is open source, any government can deploy it using its own IT infrastructure. However, a government that wants to adopt Safe Paths must sign a letter that commits it to complying with privacy and ethical guidelines. Preventing authoritarian governments and nosy employers from requiring people to use this app and reveal their data requires stringent guidelines as to how it is deployed and who can access the data, Singh said.
Apple and Google Join Forces
Apple and Google have joined in an unprecedented alliance to develop a system for notifying people who have been near others who have tested positive for COVID-19. Eight out of 10 people in the United States own smartphones, and the two companies’ operating systems run more than 99% of them. Apps built directly into iOS and Android, especially if interoperable, could dramatically increase the reach of public health authorities (the only organizations that would receive the data). To avoid fragmentation and encourage wider adoption, Apple and Google will allow only one app per country to use their system, but will allow U.S. states to use it and support countries that opt for a state or regional approach.
The system will use Bluetooth signals from phones to detect encounters rather than GPS location data. It will not run ads, will require users to opt-in, be decentralized, and use randomized IDs not tied to a user’s actual identity to communicate potential contacts with individuals with a confirmed positive COVID-19 diagnosis.
With GPS, you do not need people to have the app already downloaded for it to be effective.
API Coming. On May 20, Apple and Google released an API to developers. Next, they will issue a system update to build in contact tracing at the OS level. Should a user’s phone notify them of a possible contact, they will be prompted to download and install a public health app from their local health authority to obtain trusted instructions.
Developers of coronavirus-related apps for several U.S. states have argued that GPS location data is vital to identify infection hotspots and track outbreaks. However, for various technical reasons, workarounds designed to bypass the decision by Apple and Google and collect GPS data in connection with their contact tracing system would work poorly.
Ethical and Equity Concerns
“The work that we are doing for COVID-19 is pretty similar to work that we do on a routine basis with other reportable communicable diseases,” said Lisa Ferguson, nursing supervisor for Communicable Disease Investigations and Case Management for Multnomah County, Oregon, which includes the city of Portland. Most commonly, her unit is notified of illnesses by the state database, which receives electronic lab reports. “We assign that as a case to somebody on our team, and they call the person, interview them, ask some questions about their illness, their symptoms and where they could have possibly been exposed,” Ferguson explained. “Then, they talk about who that person might have exposed and where they were from two days before they became sick up until the time of the interview or the time that their symptoms were resolved.”
The Multnomah County, Oregon, Disease Detection Team. (Photo: Multnomah County, Oregon)
How could technology — such as smartphone location data — best help Ferguson’s team conduct contact tracing for COVID-19? “In the public health world, we are not used to having access to technology in that way,” she said. “We need to have some ethical discussions before we are prepared to utilize something like a technology that can track people.” Also, unlike tracking measles, which requires knowing whether someone was in an airspace and who was there after them, “We do not automatically consider someone to have been exposed if they were in the same airspace as someone who tested positive.”
If the privacy concerns could be adequately addressed, receiving a list of all the people who were less than six feet away for at least 10 minutes from someone who had tested positive could help her team scale up, Ferguson said. Her team would then reach out to those people, using such language as “You may have been exposed,” and “Please watch yourself closely.”
Ferguson’s team always has “equity concerns,” fearing they might under-identify groups that do not have access to the technology. “It is a supplemental tool, but it certainly would not replace the work that we are doing,” she said.
Help Wanted
Safely reopening the United States will require a new workforce of at least 100,000 contact tracers, according to a report from the Johns Hopkins Center for Health Security and other experts. Any technological assist to contact tracing does not diminish this need. For example, smartphone alerts can help filter out those at low or no risk so that human tracers can focus on genuine cases, people at higher risk, or those who are harder to contact.
Two out of 10 people in the United States do not own a smartphone, and only 42% of those above the age of 65 — who suffer 80% of the deaths from COVID-19 — do, according to a 2017 Pew Research Center poll. Hardly any homeless people own a smartphone. Among those most vulnerable to the pandemic are immigrants who do not speak English and are fearful of efforts to collect their personal information, strengthening the need for this to be done in person by trusted community members.
Finally, even if Google and Apple’s automated service is widely adopted and works well, it will require many thousands of health workers to conduct tests and follow-ups.
Creating detailed street maps and keeping them updated is an expensive and time-consuming task performed mostly by large companies. They ignore the many parts of the world where this task is not profitable, even though the need is high due to rapid growth and change in the street network, such as in Thailand.
To automate the process and make accurate digital maps available in any country, researchers at the Massachusetts Institute of Technology (MIT) and the Qatar Computing Research Institute have developed an artificial intelligence (AI) model called RoadTagger. It uses satellite imagery to tag road features in digital maps, such as lane counts, which are essential for reliable navigation.
Satellite imagery companies are constantly expanding their coverage and increasing their refresh rate, so this source of mapping data is more readily available and up to date than the data collected on the ground, such as by Google’s fleet of mapping cars. However, satellite imagery often suffers from occlusion from trees, buildings, overpasses and other obstacles.
RoadTagger gets around this problem by using a combination of neural network architectures to predict hidden features. Testing of the model with digital maps of 20 U.S. cities showed that it predicted the number of lanes with 77% accuracy and the road type with 93% accuracy.
An AI model developed at MIT and Qatar Computing Research Institute that uses only satellite imagery to automatically tag road features in digital maps could improve GPS navigation, especially in countries with limited map data. (Map data: Google/MIT News)
RoadTagger, which combines a convolutional neural network (CNN) and a graph neural network (GNN) is fed only raw data and automatically produces output, without human intervention. The CNN, commonly used for image-processing tasks, takes as input raw satellite images of target roads. The GNN — widely used to model relationships between connected nodes in a graph — breaks the road into roughly 20-meter “tiles,” each of which is a separate graph node.
For each node, the CNN extracts road features and shares that information with its immediate neighbors, thereby propagating road information along the whole graph. For example, if only two lanes of a four-lane road are visible in an image, the model uses information from nearby tiles, such as road width, to conclude that the road has four lanes.
The researchers trained and tested RoadTagger using the OpenStreetMap data set. First, they collected confirmed road attributes from 688 square kilometers of maps of 20 U.S. cities, then they gathered the corresponding satellite images from a Google Maps dataset. The training taught the model what weight to assign to various features and node connections, and it now automatically learns which image features are useful and how to propagate those features along the graph.
The researchers hope that RoadTagger will help humans validate the constant stream of changes in OpenStreetMap and similar datasets as well as enrich them with details that they do not already contain, such as whether a road is paved.
Citation. He, S., Bastani, F., Jagwani, S., Park, E., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S., & Sadeghi, M. A. (Dec. 28, 2019). “RoadTagger: Robust Road Attribute Inference with Graph Neural Networks.” arXiv:1912.12408v1.
