One of our main goals here at Telenav is to constantly improve the maps we are using in our applications and services Having very detailed and accurate maps is of fundamental importance if we want to build high-quality and precise routing applications, ADAS systems, or self-driving guidance software. In this post, we’re going to talk about how we leveraged our massive datasets of anonymized GPS (probe) data in order to enhance the OSM maps, more specifically how we were able to detect missing roundabouts throughout the world.
The problem at hand is as follows: Given a dataset of GPS probe data and the current OSM geometries, could we identify missing roundabouts? More precisely, we are searching for geometries that lack a specific tag ( junction = roundabout) identifying them as such.
A relevant case would be the one below, where the geometry clearly defines a roundabout, but that specific tag is missing from the OSM map.
What we decided to do is to analyze car movement patterns and use this analysis to make inferences about the underlying map topology. In order to solve this not-so-trivial problem, we decided to harness the power of Machine Learning. We did this because we are aware of the huge recent developments in this field and of the powers of a well-designed Machine Learning algorithm when combined with huge datasets.
The intuition about why this approach is preferable is obvious when analyzing the available data and how different traffic patterns are when we are in the context of a roundabout compared to the context of a normal intersection. What we have achieved is to teach the algorithm to associate the circular traffic movements having a “hole” in the middle with a roundabout and to associate the evenly spread movements with a normal intersection.
After successfully developing this “smart” detection algorithm, we have selected from the world map approximately 117 000 potential points, where a roundabout would be likely to exist based on some predetermined criteria. Of course, these are far too many to manually check, so the automated solution is the only one suitable for this job.
After running those points through the Machine Learning algorithm, it has detected around 9000 missing roundabouts in Europe and North America, as those are the areas for which we have GPS probe data. The massive size of these results which translate to substantial improvements of the OSM map is obvious when visualizing them.
After a quick series of manual testing of a batch of results, we have discovered that the predictions are between 82% – 86% correct, depending on the level of confidence, which proves the efficiency of the Machine Learning oriented solution to this difficult task.
In the near future, we plan to release this data to be validated by the OSM community through the MapRoulette platform and we are eager to see the feedback we get. Having acquired even more knowledge in this field, we are now ready to tackle more difficult problems using more advanced Machine Learning and Deep Learning algorithms. This will surely enable us to improve the OSM maps even more.
We have uploaded all our predictions here in CSV format for those of you who are interested in playing with the data.
2 thoughts on “Enhancing OSM Maps using Machine Learning & Big Data”
I have looked at openstreetcam several times, but I have been unable to fathom how it is supposed to be used apart from some smart phone apps.
I make some use of a dash cam (with embedded gps) for mapping on OSM. It produces mov/qt files some of which I might perhaps upload to openstreetcam but I am quite unable to discover a simple upload option.
Likewise, I cannot even *find* any video footage to view or download!
Why is the simple and obvious just not present or not documented?
You are misinforming the reader by telling that you could produce a highly accurate map used in highly automated driving. You can with 82% certainty detect a roundabout where HAD requires cm. or mm. accuracy of the map. GPS is not accurate enough by far.
Comments are closed.