Researchers at the University of Melbourne are looking for spatial data that will help them understand the way hazards affect road networks. This is particularly useful as vehicles develop greater autonomy in navigating these hazards.
Eight tags have been added to our list to help facilitate this research. We’re inviting you to take part and help us identify and tag these incidents in images.
Tagging instructions
Have you observed one or more of these incidents while capturing imagery in Mapillary:
Collapse
Crash
Fire
Flooding
Landslide
Roadkill
Rockfall
Treefall
The researchers are looking for incidents that actually affect the road network. Here is an example of such an incident and how it can be tagged.
More detailed instructions on tagging and searching tags is available on our help portal.
The researchers have provided some background on their objectives: In pursuit of safer autonomous vehicles we are currently building a dataset that covers incidents that may occur on street networks. In this task we are looking to expand the following five categories of incidents: animals on or dangerously close to the road, snow, flooding, crashes, and treefall. Each of these should affect the driving surface in some way. For instance, snow is only relevant if the driving surface itself is covered in it. Likewise, a fallen tree that is not blocking the driving surface is not an incident either. We are interested in images from all across the globe, but images from the United Kingdom are particularly preferred.
I’m the main author of this research effort, and I’d be happy to answer any questions you might have!
For starters, I’ll add a little extra info on what’s been done so far.
Our research intent is to recognize incidents and hazards on the driving surface. To do so, we’ve been collecting and cleaning images from Google, Flickr, and Bing, up to a total of 13,000 relevant images so far. This process is still ongoing, and we reached out to Mapillary to see how the community could help, which is where the tagging task comes in.
Regarding the two questions already asked in this thread:
We encourage users to tag images even in the face of uncertainty. For instance, perhaps this rabbit could be in a dangerous position (e.g. on the road near a corner), in which case it would require detection. We reason that it’s better to have many examples that we can consider than to be overly rigorous up-front. We manually go through the tagged images once they’re collected to make sure that each image fits our criteria, so please feel free to tag generously!
Road debris and blockades are a significant hazard, and the only reason we didn’t consider it so far was because it turned out to be incredibly difficult to search for using search engines. Maybe we can open a tag and see how many people have relevant imagery, since it does occur and we’d do well to recognize when it does.
Thank you to everyone that’s contributing. Every image is a step towards a more complete and representative dataset, which is what we hope to achieve by the end of the research period.
Possibly beyond your time limit, but in the UK there’s a service called fixmystreet.com. You can report things like obstructions in road (and similar) and sometimes photos are added.
I do not know the licensing constraints of the data, but it might be an additional source for you.
That is awesome, we did not know about that initiative yet. Thanks for pointing it out! I’m on the road at the moment so I can’t take a look just yet, but it sounds promising!
Hi @ahlevering, any updates on your work? Do you have a publication that you can share with us? I am also interested in doing this, but specifically for landslides here in the mountain region of the Philippines.
Hi @tazmr , certainly I have some updates. We’ve completed the research about a year ago, and we’re in the process of publishing a paper on the topic. The data (12.500 images across 8 incident types) will be made available at that time as well. In a nutshell, the results we achieve when detecting incidents are very strong, but obviously this needs more testing in a real-world scenario.
For landslides we ended up collecting only a limited set of images, only a few hundred, from the top of my head. Since we last gathered images two years ago, maybe going through the same process will yield a couple hundred more. That said, I’ll update this thread once the publication is out. Let me know if you have any further questions!
