Software Blog

Analysis of positioning with GNSS and RTK with the Geographic Info Toolbox

by Elvira Thonhofer

In many of our developments and applications at ANDATA, precise, trustworthy but also cost-effective localization is an important component. A technical variant of this is Real-Time Kinematic Positioning (RTK). This blog includes a quick examination of RTK compared to other GNSS variants using the Geographic Info Toolbox.

Use cases and requirements

The following use cases provide the functional requirements for localization accuracy.

  • Collision avoidance of any vehicle with other vehicles as well as voulnerable road users (two-wheelers and pedestrians) requires a prediction of the movements of these road users. This can sometimes be done or improved by utilizing GNSS localization data.
  • Many direct vehicle network and communication services - ITS-G5 as well as 5GAA - include position information. The exemplary application of Collective Perception is explained in more detail in Project COPE, for example.
  • A digital twin of the traffic system (see https://ieeexplore.ieee.org/document/10102410 ) requires simple, pragmatic and fast precise positioning and localization of a wide variety of static infrastructure objects and of moving road users.
  • Behavioral models of road users from naturalistic driving or floating car data need position data in order to be able to determine the appropriate environmental context from maps and/or digital twins.

The ANDATA RTK base station

We have been operating our own RTK base station for research and development purposes since summer 2024 in order to be able to use RTK in the area of Hallein. Our base is listed on rtk2go (rtk2go NTRIP Caster Table), the data can be obtained from interested third parties.

Name HalleinANDATA
Protocol RTCM 3.2
Messages 1005(1), 1077(1), 1087(1), 1230(1)
GNSS GPS+GLO

The use of the RTK correction data is subject to the rtk2go Terms of Use. We do not assume any liability for this and for data provided by our base station.

Test Configurations

To compare performance, 3 different configurations were tested:

ID

Description

Features

Logger System

Cost

conv. GNSS

Navilock NL-82002U u-blox NEO-M8U

GPS, Galileo, Beidou, Glonass, IMU

Raspberry Pi 3B+

ca. €200,-

2D

2D GPS Antenna with IMU

GPS, Galileo, Beidou, Glonass, IMU, Virtual Sensor for Motorbike Dynamics

2D Sticklogger

ca. €2000,-

GPS+RTK

c94-m8p u-blox RTK application board

GPS + RTK

Raspberry Pi 3B+

ca. €500,-

 

We have been using conventional GNSS on our test vehicles for many years. We had replaced the GNSS antenna at the beginning of the last measurement season, which now also has an internal IMU and can use it to improve the positioning result. We have also been using the 2D sensor from Debus & Diebold on a motorcycle test vehicle for several years. The use of RTK is new and has only been available to us since the summer of 2024.

All 3 systems are mounted together on the luggage rack at the rear of a motorcycle with a clear view upwards, so that their relative difference in position to each other is less than 10cm. GNSS and RTK are logged via the same Raspberry Pi.

Test track

The Rossfeld-Panorama-Straße serves as a test track, which is approached from the ANDATA office via the Dürnberg-Landesstraße. The test track is passed in both directions. The RTK base station is set up at the ANDATA office, which is also the start and end point of the test drives. The following figure shows the test track as well as the location of the RTK base station and an 8km radius within which the entire test track is located.

This test track contains some challenges regarding positioning using GNSS, in particular various sections with reduced visibility of satellites due to alpine terrain formation, forestation and house walls (city of Hallein):

  • At the beginning of the Dürnberg state road there is an avalanche gallery from position 47°40'52.9"N 13°05'25.8"E, followed by winding passages with limited satellite visibility. (Details 7 and 8 in the following overview)
  • At the toll station on Rossfeld-Straße, the satellite view is usually covered for more than a minute during the payment process (detail 10).
  • On the west side of Rossfeldstraße there are some forest passages in dense vegetation, as well as on the approach to the toll station (details 14 and 18).
  • The return journey ends with the passage through Hallein's historic city center with narrow urban canyons of Bräuerstraße and Wichtelhuberstraße as well as Ruprechtgasse (details 5 and 6).

The above-mentioned detail sections are shown in the following two graphics. Several test drives with slightly different routes through the old town or along the Rossfeld circuit were undertaken, therefore the sections shown can be associated with individual test drives.

Comparison of the different configurations

Two test drives were undertaken, in which all 3 configurations were in use at the same time. These two trips are shown and discussed in detail below.

Note: The map background shown is not suitable for judging absolute deviation of the position data with respect to the road or even lane. The quantification of absolute accuracy within the lane will be discussed later and uses highly accurate, digitally available map information.

Detail 1 - Hallein, start and end point

The start and end points in Hallein show anomalies in both test runs. The GNSS antennas first have to detect the satellites and usually need a few seconds for cold or warm start. The startup time until the first trustworthy positioning is available can usually be found in the antenna/chip specifications.

Detail 2 - Hallein, Pernerinsel South

Before turning on the "Galsterer intersection" in the direction of the city bridge and Pernerinsel, there is a traffic light where the motorcycle comes to a standstill in this test drive. There, all 3 signals drift to the left (west). The measured positions at the intersection are therefore not trustworthy and inaccurate. On Pernersinsel, the conventional GNSS signal is much further to the left than the other two signals.

Passing through the Pernerinsel intersection is no problem here on the second test drive. Nevertheless, there is a greater lateral difference of the positioning signals between the configurations on the way back at the exit from the historic center (bottom left) and when entering the intersection approaching from the south.

Detail 3 - Hallein, Pernerinsel North

Passing through the roundabout on Pernerinsel is no problem for GPS+RTK or for the 2D sensors, the conventional GNSS shows deviation. Approaching the roundabout (on the right in the picture), the signal of the 2D sensor shows a deviation. The explanation for this is currently being investigated, will be provided in a future blog post.

Detail 5 - Hallein, historic city center

The entrance to the historic city center of Hallein is generally a difficult passage for GNSS-based sensors. The view of the satellites is largely obscured by tall buildings and the terrain behind them (Großer and Kleiner Barmstein). In addition, signal reflections can occur on the walls of buildings, which cannot be detected automatically when visibility of the satellites is reduced. Hence, the three positioning results show significant deviation from each other.

The signals also deviate very strongly during test drive 2. The GPS+RTK system jumps significantly to the south compared to the other sensors, indicating satellite signal reflection. The blue trajectory in the upper right of the picture is an implausibly large deviation from the conventional GNSS in the northern area of Pernerinsel on the outward journey and is to be ignored in this view.

Detail 7 - Dürnberg State Road Avalanche Protection Gallery

The first section of the Dürnberg state road leads into mountainous and wooded terrain on the one hand, and on the other hand the road runs under avalanche protection galleries in places. The visibility of the satellites in the gallery is almost completely obstructed. All three systems have problems in these sections that result in inaccurate positioning signals. The 2D sensor technology works somewhat more reliably in these sections, as additional information is used as virtual sensors with the other sensor signals from the IMU that are still available.

During the second test drive, all systems have problems in the gallery. 2D provides the best results. Both conventional GNSS and GPS+RTK show large local deviations and need about 100m distance to provide more reliable position data again.

Detail 8 - Dürnberg-Landesstraße

Immediately further down the road, conventional GNSS loses connection to the satellites and no longer outputs data points (indicated by the straight line connecting the only available data points). Shortly before the last curve shown, the GNSS+RTK system also loses the signal and no longer outputs position data. However, the outage is significantly shorter than with conventional GNSS.

Detail 9 - Rossfeld wooded section

The RTK signal has an offset along the long straight in a densely forested section. Both conventional GNSS and 2D sensor technology are close to each other.

In the second test drive, the signal from the conventional GNSS drifts significantly in the densely forested section. This is to be expected. The GPS+RTK signal shows only minimal deviations compared to the 2D sensor system and both follow the road well.

Detail 10 - Rossfeld toll booth

All three signals drift briefly under the roof of the toll booth when the motorcycle is at a standstill. As expected, both standstill and roofing pose problems for GNSS-based positioning.

In the second test drive, the conventional GNSS signal drifts significantly under the roof of the toll booth. GPS+RTK and 2D sensor signals remain stable with minimal drift. After the toll booth, the signals move away from each other, with conventional GNSS running between the other two signals.

Detail 11 - Rossfeld hairpin before the plateau

All three signals are close together during the first test drive. The map background suggests that all three are clearly too far inside the curve. However, this statement cannot be deduced, as the map material does not have a guaranteed minimum accuracy. It is just a visually appealing background, not a precise orthophoto with guaranteed, global accuracy.

In the second test drive, the trajectories are more plausible in relation to the same map as above. However, the conventional GNSS signal moves away from the other two signals after driving through the hairpin bend. This is not to be expected, as there is a good view of the sky there. It is unclear what the cause of this deviation is.

Detail 14 - Rossfeld 3 hairpin curves

There is an interesting peculiarity at the upper hairpin bend. Conventional GNSS jumps and is implausible. GPS+RTK and 2D sensor have an offset relative to each other. The site is densely wooded and therefore probably difficult for GNSS-based positioning. The combination of a forest passage with a narrow hairpin bend is particularly difficult here.

Detail 15 - Rossfeld long curve

The conventional GNSS signal shows unexpectedly large deviations at this point. The positioning data does not even run along the road, sometimes data points are missing (which are recognizable by straight lines between data points). It is (for now) unclear why this deviation occurs here.

Analysis of GNSS error dimensions in detail

It is now necessary to investigate in detail whether the quality of the positioning can be derived from the additionally available GNSS data (error ellipses, number of satellites, etc.), or whether the loss of quality in particular has other, unknown (to us) causes.

It is known, for example, that the number and distribution of satellites used has an influence on quality. It is therefore to be expected that if the visibility of satellites is obscured (forest, canopy, mountain flanks, urban canyons, etc.), the GNSS-supported positioning result will be worse. We can investigate these aspects with the available data. The analysis of selected passages is presented in the next section.

Other influencing factors are probably also the altitude profile of the track, as well as the other sensor signals that are used for sensor fusion in the 2D sensor. Since the algorithm used is unknown to us, the only thing left here is the systematic investigation of the signals and the formulation of the hypotheses.

When using conventional GNSS systems (GPS, GLONASS, Galileo and BeiDou), both the number of satellites (at least 4 for a valid position calculation) and their distribution in the sky are important. Depending on the view of the sky and elevation mask, the number of visible satellites can vary greatly. This is even more true for sensors that are permanently mounted on a motorcycle that leans significantly in curves.

The use of RTK requires an RTK base station computing and transmitting correction data for a subset of well-distributed satellites available to it. A GNSS+RTK sensor in the field can therefore only use a maximum of those satellites that are also available at the base station. Because a "change" of the satellites used is algorithmically unfavorable, the number of satellites used in the use of RTK is largely constant.

For the analysis of data, our software tool Stipulator is a suitable tool at our disposal to load raw data of the measurement runs, to process data in any complex way and to display data in a variety of flexible plots and customizable figures.

Signal Quality Analysis for Details 14 and 15

In the Rossfeld section shown in Detail 14-15, the conventional GNSS shows large deviations from the actual road in one of the test runs. The hypothesis that this is due to the visibility of satellites can now be analyzed. In addition to the number of satellites used for each data point, we can also analyze the Geometric Dilution of Precision (GDOP).

In those sections where the deviation is particularly large (hairpin in the north and curve in the southwest), the number of satellites used is particularly low (red data points).

It turns out that the GDOP values of conventional GNNS are also particularly high in the problem areas. In addition, it is noticeable that if fewer than 4 satellites are visible, the GDOP value is no longer logged as an output (missing data points, in the GDOP plot, where there are still data points in the plot of the number of visible satellites).

The following correlations can be noted:

  • If fewer than 4 satellites are visible, no GDOP value is calculated. The positioning result is not trustworthy.
    • see figure, in 2 places: hairpin bend north and curve southwest
  • The GDOP values are not very sensitive to how many satellites are used, as long as more than 6 satellites are well distributed in the sky
    • see figure in the east: GDOP in the range 1.5 to 3, with the same number of satellites used in the range 4 to 16+

Visible satellites along the entire Rossfeld Circuit

In the overview maps of the trajectories shown above, only one direction of travel is shown in the Rossfeld area at a time, so that the positioning results are easier to interpret. Even during two test runs, different results are achieved. To better understand this behavior, we looked at the number of satellites used for one test run in both directions and compared them directly.

Used satellites and GDOP for conventional GNSS

In conventional GNSS, the number of used satellites vary and differ for the respective directions of travel. The problem area shown in detail above clearly stands out as an exception in which fewer than 4 satellites are used. Otherwise, the number of satellites used varies between 6 and over 16.

Used satellites and GDOP for GPS+RTK

The use of RTK stabilizes the number of satellites used. As far as visible, all satellites that are also used at the base station and for which RTK correction data are therefore available are used. As expected, the number of satellites used is therefore largely constant and is around 10-12 satellites. Thus, the systematic advantage of RTK is noticeable in practice (compared to the partially reduced satellite visibility in conventional GNSS without RTK above).

Reference data for the quantitative assessment of accuracy

In order to be able to quantitatively assess the accuracy of our measurement data automatically, we need digitally available "ground truth" data. Our test track runs in the border area between Bavaria (Germany) and the province of Salzburg (Austria). The data basis must therefore be available for both countries.

For the assessment of horizontal accuracy, for example the accuracy within the road or even lane, we have several potential sources at our disposal:

Orthophotos are suitable in combination with QGIS for visual verification of accuracy. However, the road centerlines are not available digitally and are therefore not readily suitable for automated evaluation.

Maps from OSM, on the other hand, provide digital road data. These can be queried via an API in an area. At the same time, OSM does not promise absolute accuracy of map data.

So we took the following pre-processing steps:

  • OSM data in the area of the Rossfeld Round automatically queried with our tools
  • Section-by-section manual validation of the position of the centerlines via ortho photos (QGIS)
  • Use of isolated centerline data in our horizontal positioning accuracy calculation tools

The satellite images used as backgrounds in the figures shown are not orthophotos and are therefore not suitable for assessing quantitative accuracy.

We also have several sources at our disposal to assess vertical accuracy:

The terrain models are essentially point clouds, with each point having as coordinates the geographical longitude, latitude and height above sea level (note: possibly different reference points). It should be noted that the terrain model - as the name suggests - depicts the shape of the terrain. In sections where the road is run as a tunnel or bridge, it is not visible in the terrain model. The terrain model also contains no (explicit) information about the location of the road, i.e. there are no labels that identify road points. It is still possible to isolate those parts of the terrain model that belong to the roads in our test track, as we already have the road centerline digitally available. This enables us to subsequently work with the digitally available altitude information.

Statistical evaluation of accuracy within a lane

For the subsequent use of measurement data to train behavioral models, it is necessary to ensure that the data basis is of sufficient quality. This also includes that  accuracy can be quantified and, based on suitable criteria, measurement data can be used or discarded section by section.

With the tool Stipulator and the Geographic Info Toolbox, we have functions at our disposal that allow us to evaluate and analyze many test runs collectively. Thus, systematic errors can be better detected and suitable criteria for calculating quality characteristics can be developed and evaluated.

To evaluate horizontal accuracy, the distance to the centerline of the road (ground truth) is used. The map material is made available as open government data from Bavaria and the province of Salzburg. After the centerlines have been (automatically) digitized, they can be used to calculate the horizontal accuracy of the positioning result.

Overall, depending on the system, we only have a few test runs available that we can evaluate collectively.

Statistical evaluation of horizontal accuracy for conventional GNSS

With conventional GNSS (and the latest GNSS antenna) we have performed 11 test runs. The following figures can be read as follows: The average distance of the position data to the true centerline of the road is shown in color. A distance between 0 and 3.5 m (positive, yellow to dark red) means that the position data is measured on average in the correct lane. Negative distances (green to dark blue) mean that the positioning on the opposing-traffic lane has been measured.

In this diagram, sections with mean unsatisfactory positioning accuracy can be detected (curve in the west of the Rossfeld Circuit). Our goal is to reliably identify these sections and identify (measurable) causes. In addition, the average positioning accuracy is between light green and yellow, which corresponds to the center of the road. This is implausible, because even with a motorcycle the test drivers remain roughly in the middle of the lane.

Statistical evaluation of horizontal accuracy for GPS+RTK

With GPS+RKT, we have 2 test runs in out database. The RTK base station has been online since the end of summer 2024. In the next motorcycle season, we plan to significantly increase the number of test runs in order to be able to make a fair statistical evaluation for this system as well.

The dispersion of the average accuracy is remarkably high, which is due to the low number of test runs. The average accuracy is in the orange range, which corresponds to the middle of the lane. Passages with poor positioning accuracy are visible, but it takes significantly more trips to identify systematic problem areas.

Statistical evaluation of horizontal accuracy for 2D sensors

With 2D sensors, we have 17 test runs in our database. The positioning results are overall very good and homogeneous with this sensor configuration. There are hardly any sections that are noticeably worse. From this it can be concluded that the 2D sensors that use sensor fusion deliver very stable and good positioning results overall.

Conventional GNSS shows in some sections

  • large differences between individual test runs (not shown),
  • very large deviations to the left (rarely) or to the right (often) beyond the driveable area of the road,
  • average positioning in the middle of the road, which is implausible,
  • signal breaks or completely implausible trajectories.

GNSS + RTK positioning results are

  • mostly on the correct directional lane,
  • in sections large deviation to the right beyond the driveable area of the road.
  • The "statistical evaluation" is the evaluation of a single trip over the majority of the test track.

2D sensor technology shows

  • very good positioning within the correct lane of the directional carriageway,
  • no systematically conspicuous weaknesses.

Vertical accuracy of positioning

The vertical accuracy of positioning usually plays only a subordinate role in our research problems. Nevertheless, with good data, it is possible to use the gradient of roads as information in behavioral models and thus also in forecasting.

To be able to make a quantitative statement about the vertical accuracy, we need a "ground truth" - i.e. the true height of the road(s) in which the test vehicle is driving. In the area of our test track, we have two sources at our disposal:

The digital terrain model serves as the basis and "ground truth" of the elevation information. The position data can now be projected onto the terrain model and compared to the corresponding true altitude above sea level.

For a single test run, an evaluation of vertical accuracy is shown here.

Vertical accuracy of conventional GNSS

The following two figures represent the same information: the deviation of the measured elevation above sea level from the true elevation. The terrain model is shown in grayscale. In color, the absolute, vertical error is shown over the course of the Rossfeld circuit. Large deviations are shown in red, small deviations in dark blue to black.

Conventional GNSS shows very large deviations in both directions in several sections.

This view represents the isolated trajectory (only one way) without the terrain model. The shade of gray indicates the height above sea level.

Vertical accuracy of GPS+RTK

The GPS+RTK configuration shows good results in terms of vertical accuracy, i.e. small deviations and few outliers.

The isolated trajectory (one direction of travel) shows one bad spot (bottom right). There, the terrain model does not correspond to the height of the road, as it leads across a bridge over a depression in the terrain.

Vertical accuracy of the 2D sensor

The 2D sensor configuration also does not show any major deviations vertically. Compared to the GPS+RTK configuration, the deviations over long distances are a bit larger, although still significantly smaller than those of conventional GNSS.

The isolated trajectory (of one direction of travel) does not show much deviation at the point where the terrain model does not match the course of the road. In this individual case, this is due to an additional deviating horizontal position that does not follow the course of the road.

Statistical evaluation of the vertical accuracy of the three sensor configurations

Over one test run, the calculated heights match the true altitude above sea level better or worse in sections. The statistical evaluation of each test run provides the following overall impression:

Conventional GNSS has high deviations:

  • Median deviation > 5 m There are many outliers, which are also related to the poor positional accuracy (horizontal).
  • The outliers (red markers) range from -400 m to +40 m.

GNSS + RTK has significantly narrower deviation distribution:

  • The median accuracy is about 3.5 m
  • There are significantly fewer outliers, and these are in a narrower range

2D sensor technology also has narrow deviation distribution:

  • The accuracy is around 4 m, so it's slightly worse than RTK.
  • The deviation range is slightly wider than with RTK, but significantly better than with conventional GNSS.

The use of RTK correction data enables improved vertical positioning quality. The vertical deviations from the terrain model that occur are much narrower than when conventional GNSS is used.

Summary

  • We can equip our measurement vehicles with different GNSS antennas and use RTK correction data.
  • Each of our sensor systems has advantages and disadvantages, which must be evaluated depending on the goal of the measurement run and in the context of the analysis.
  • Depending on the research or development goal, there are different requirements for the accuracy of positioning. With our tools, we can automatically quantify the positioning accuracy achieved. Thus, insufficiently accurate data can be identified and, for example, eliminated from a model training dataset.
  • The GNSS devices used are relatively inexpensive and will allow us to equip different road users in the future. Especially when locating VRUs (pedestrians, cyclists) or building digital twins of the traffic system, we can collect targeted measurement data to develop and consistently improve prediction models.
  • The operation of our own base station allows not only us, but also other interested parties in the Hallein area to use RTK data. Any user can access our base station data.
  • We do not achieve the theoretical accuracy of < 10 cm in mountainous and wooded areas with RTK. Nevertheless, the GNSS+RTK combination is a more reliable configuration than conventional GNSS. In terms of vertical accuracy, even 2D is not better.
  • The 2D sensor technology is very accurate over many journeys and is particularly robust where GNSS-based systems have problems (occlusion, concrete jungle, forest).
  • The development and use of virtual sensors and other advanced methods for data fusion is an effective method for improving sensor data-based information, as is demonstrated by 2D.

The goal for the next measurement period is therefore to generate a sufficiently large pool of RTK measurement data with which a statistical investigation also has sufficient confidence. Then, the systems can be compared fairly. In addition, further correlations between positioning accuracy and available quality and error measurements of the GNSS antennas can be investigated to develop suitable parameters for the identification of insufficiently accurate data. This significantly improves our developments in the various fields of application mentioned above.

Further information and in-depth information will follow in future blog posts or can be requested from info@andata.at.

 

Go back

Comments

Add a comment

What is the sum of 9 and 9?