Estimating Temporally and Spatially Dependent GNSS Errors Across Roadway Networks using Crowdsourced Data
Nigel Williams, Chao Wang, and Matthew Barth, University of California, Riverside
Alternate Number 3
With the advent of Connected and Automated Vehicles (CAVs), Global Navigation Satellite Systems (GNSS) play a key role in vehicle positioning and navigation. While vehicles with some level of automation are generally equipped with positioning sensors beyond a GNSS receiver, Connected Vehicles (CVs) in current deployments across the U.S. primarily use Differential GNSS (DGNSS) receivers for positioning. DGNSS is capable of lane-level accuracy under open-sky conditions, which fulfills the positioning requirements of many CV applications. However, the positioning accuracy may drop below lane-level (in fact, errors may be on the order of tens of meters) in urban canyons due to blockage of GNSS signals and multipath effects. This poses a major problem for many CV applications, particularly those that rely on lane-level positioning accuracy.
There are ways to mitigate this to some extent, such as integration with an Inertial Measurement Unit (IMU). However, an IMU accumulates error over time and may not fulfill the positioning requirements during the entire period of degraded GNSS accuracy.
At the same time, the amount and variety of crowdsourced vehicle activity data are growing rapidly. An example of crowdsourced data from CVs is the University of Michigan’s Safety Pilot dataset, containing the GNSS data from more than 100 Connected Vehicles driving in the vicinity of Ann Arbor, Michigan. Since a fundamental part of Connected Vehicles is broadcasting position (and other) data, this enables collection of a vast amount of GNSS data from vehicles. Crowdsourced GNSS data have been used to, for example, estimate the number of lanes in a section of roadway. If the number of lanes is already known (from a source such as OpenStreetMap), we propose using the GNSS data to estimate the variation of GNSS position error throughout a roadway network. An assumption is that the data come from receivers of similar quality, which is true for Connected Vehicles in current deployments.
Given the limitations of GNSS and the rise of crowdsourced data, a key question that arises is: How can we improve the performance of CAV applications (with regard to positioning) without substantially increasing system costs? This paper proposes a solution: estimating which areas of a roadway network are more prone to GNSS error (e.g., due to deeper urban canyons or increased shading from foliage). Next, a strategy such as one of the below may be adopted with the goal of improving CAV application performance: 1) Vehicles minimize their operations in those areas, choosing a route which goes through less error-prone areas; 2) Infrastructure components that can supplement the positioning solution (e.g., using traffic cameras or pseudolites) in higher-error areas.
In order to estimate which areas of a roadway network are more prone to GNSS error, we fuse the following information: 1) measures of GNSS accuracy obtainable from GNSS receivers, such as Dilution of Precision; and 2) the lateral (i.e., perpendicular to road) spread of GNSS positions. The idea is that a higher spread indicates greater position error, in the lateral direction at least. A Gaussian mixture model (GMM), oriented perpendicular to the road, can be fit to the position data. The spread of GNSS positions is obtained from the variance of each GMM component. The number of components in the GMM (number of lanes) is obtained from OpenStreetMap.