Osm-yolo-crossings: AI import tool for zebra crossings

ivanbranco · August 4, 2023, 11:09am

I post it here in the forum since many people doesn’t read diaries entries. @NorthCrab shared in a diary entry his latest project. It’s a tool that autonomously detects and maps pedestrian crossings (zebras) in OpenStreetMap. You can find it here: GitHub

dieterdreist · August 4, 2023, 10:16pm

If I see it correctly this is based on orthographic imagery, which means it cannot detect a crossing type, just a crossing marking (which could be a traffic lights controlled crossing as well), right?

Minh_Nguyen · August 4, 2023, 10:51pm

Here are the tags it outputs:

github.com

Zaczero/osm-yolo-crossings/blob/def8ffbc492551c6c6fa963713e1a170fd14233b/crossing_type.py#L12-L28


      
          if crossing_type == CrossingType.UNCONTROLLED:
              return {
                  'highway': 'crossing',
                  'crossing': 'uncontrolled',
                  'crossing:markings': 'zebra',
              }
          elif crossing_type == CrossingType.TRAFFIC_SIGNALS:
              return {
                  'highway': 'crossing',
                  'crossing': 'traffic_signals',
                  'crossing:markings': 'zebra',
              }
          else:
              return {
                  'highway': 'crossing',
                  'crossing:markings': 'zebra',
              }

NorthCrab · August 5, 2023, 7:02am

For some context: I intend to support uncontrolled/traffic signals tagging at some point, but during development it turned out that it would be a much better idea to move it to a completely separate, and a new project, this way uncontrolled/traffic signals will be its own separate problem and I will have a lot of flexibility while working on it. I did have some initial success, so my hopes are up . When a new project is done, it would be able to “upgrade” ANY highway=crossing, not just the ones created by the YOLO project. At the moment, YOLO is only tagging highway crossing without crossing specification. This should be more than enough for now.

Peter_Elderson · January 14, 2025, 1:59pm

I have been using a data source of zebra crossings detected from the air, to check and improve zebra mapping. The number of false positives and false negatives is very high. Also, the matching of detected zebra crossings to existing crossings is good, but not excellent. The all-in error rate is much too high for an automated import/conflation; it would cause an enormous amount of checking and correcting afterwards.

I sincerely hope that this tool is better and does not introduce more errors than it solves!

SomeoneElse · January 14, 2025, 2:30pm

Given that the posts above are from 2023, presumably there’s data from between then and now in OSM already. How do we find it, in order to have a guess at things like false positive and false negative rates?

Peter_Elderson · January 14, 2025, 3:00pm

Exactly! I have no answer to that, but some people have amazing ideas and the skills to implement things. I just make do, and in this case I manually check all the dots. In the end I have a list of false positives, which I will feed back to the provider of the data, if they want it. False negatives are not yet on the agenda, but I now they are there because of frequent chance findings.

A proportion of the false positives were true positives when the aerial imagery was taken. The promise of updates on the provider’s website was a tad optimistic, I think. That is why I was looking for similar tools.