100 times more accurate than leading GNSS-based positioning system
Dr Anthony Ashbrook
The project
Between 21 September 2020 and 27 October 2020, our trainborne sensor travelled 13,500km on a national railway network, acquiring data to validate the absolute positional accuracy and run-on-run repeatability of the our sensor and compare that to an industry standard GNSS-based solution. What we found was “mindblowing”, or so said a senior contact at the network operator instigating the project.
Background
A unique benefit of RailLoc, Machine’s With Vision’s rail positioning technology, is that it uses track-feature matching to ensure very high run-on-run repeatability. Our target performance is to ensure that any measurements taken at the same location will not differ by more than 100mm.
A method for automatically quantifying the repeatability of RailLoc and the incumbent GNSS-based positioning system has been developed and is used here. The method works as follows. Every track-feature measured by RailLoc has an associated timestamp. This is the time that this feature was measured by the DVS camera sensor and is accurate to within 200ns. When RailLoc identifies a match between a detected track-feature and a track-feature stored in the map it means that the train at these two times must have been at the same location. RailLoc measurements at these two times should agree and any disagreement is a measurable repeatability error. Similarly measurements from the existing positioning system (or any other positioning system) should also agree at these times and the repeatability of the other positioning system can be quantified.
In the following sections this method is used to quantify the repeatability of RailLoc and the existing positioning system using two sections of the 13,500km run.
Short section (1.3km) used to quantify the repeatability of run-on-run position measurements
To provide some context, Figure 1 presents the vision-based track-feature matching over a period of around 200ms. The top image displays a segment of data recorded on the 21st of September 2020 and this data is used as the map. The bottom image displayed live data being recorded on the 19th of October. The coloured lines indicate track-features detected in each of the images and found to correspond.
Figure 1: Visualisation of vision-based track-feature matching. The top image displays data recorded on the 21st of September and is used as the map. The bottom image displays data recorded on the 19th of October and is used to match to the map, enabling high repeatability.
Figure 2: In this test, RailLoc has a systematic error of 7mm and a statistical error of 3.6cm. The existing positioning system (NPS) has a systematic error of 2.6m and a statistical error of 29cm.
Figure 2 presents the histogram of differences between RailLoc and the existing system’s position measurements as the measurement train travels along the short section (1.3km), quantifying the repeatability of these systems. The key observations are as follows:
RailLoc position measurements from one run to the next agree extremely well. There is only a very small systematic difference of 7mm in this test and the standard deviation of errors is 27mm. This means that run-on-run 99.9% of measurements will be within +/- 100mm.
The existing positioning system exhibits a much larger variance in measurement repeatability than the claimed 1m. In this test there is a mean systematic difference in positions reported by the existing positioning system between the map run and the test run of 2.7m. Within the run this systematic error varies with a standard deviation of 294mm.
This test provides evidence that RailLoc improves repeatability, when compared to the existing positioning system, by an incredible factor of 100. The RailLoc error is dominated by the variance of 27mm whilst the existing positioning system is dominated by the systematic error of 2.7m.
Longer section (225km) used to quantify the repeatability of run-on-run position measurements
Figure 3: In this test, RailLoc has a systematic error of 9mm and a statistical error of 7cm. The existing positioning system (NPS) has a systematic error of 2.1m and a statistical error of 1.1m.
Figure 3 presents the histogram of differences between RailLoc and the existing system’s position measurements as the measurement train travels along the much longer section (225km), quantifying the repeatability of these systems. The key observations are as follows:
RailLoc position measurements from one run to the next agree extremely well even over this much longer section. There is only a very small systematic difference of 9mm in this test and the standard deviation of errors is 70mm (2.6 times larger than in the Map 1 test). This still means that run-on-run 85% of measurements will be within +/- 100mm.
The existing positioning system exhibits a much larger variance in measurement repeatability than the expected 1m. In this test the difference in positions reported by the existing positioning system between the map run and the test run vary from around -2m to 5m, a range of 7m.
This test also provides evidence that RailLoc improves repeatability, when compared to the existing positioning system, by an incredible factor of 100. The RailLoc error is still dominated by the variance, now of 70mm. In this test the existing positioning system is dominated by the data spread of up to 7m.
And why is all of this important? High accuracy position measurements are crucial in enabling automatic run-on-run alignment of infrastructure condition measurements recorded by on-train sensors. Which is another blog in itself.
A full report detailing the project in depth is due to be published shortly. We will update this page with a link as it is available. In the mean time, if you would like more information, please contact us at info@machineswithvision.com .
