Currently working on a bachelor project in computer engineering to collect traffic sign data.
Our approach will be to develop a mobile application for Android which uses computer vision to detect traffic signs while driving.
For our project its important to know of alternative methods of collecting traffic sign data so that we can weigh our solution against others. So, what are the currently used methods of collecting this data? Do you know of any other method that is similar to ours, or is most of this data collected from public/government-controlled databases?
Both Mapillary (VC-funded start-up) and OpenStreetCam (originally Telenav now Grab) use apps to collect sequences of imagery (primarily through mobile phone apps) & then process the imagery with ML techniques to identify traffic signs. No doubt there are many others.
The Mapillary team include very experienced computer vision professionals and have substantial VC funding behind them. I suspect to be successful you not only need vey sophisticated initial processing algorithms, but also a huge quantity of training data (tens of thousands of kilometres/miles), even when road signs have well-defined characterstics. For instance Mapillary does not recognise the UK “end of local speed limit” sign.
It’s certainly worth playing creating a short sequence on Mapillary and seeing how many traffic signs it finds.
For a lot of traffic sign data in OSM (and especially the related attributes, such as access restrictions and maxspeeds), the source is neither computer vision nor government databases, but humans being physically present and looking at the signs, then entering their observations into a computer by hand.
