I wrote recently about developments in augmented windscreens that researchers believe will eventually lead to predictive capabilities being provided to drivers.
These capabilities have been explored by a German team called UR:BAN, who are working to better understand the complex behaviors of all road users so that they can better predict how they will behave.
Recent studies suggest that such systems will soon be able to predict what the driver will do up to three seconds before they actually do it.
We already have a glimpse of this future with the array of driver assistance technologies that are fitted in the latest models. Alongside the German team, a group of researchers from Cornell University are also working on developing predictive capabilities for our vehicles.
They are building a comprehensive database of our behaviors when we drive to try and understand the various trigger movements that signify that a particular behavior is about to be undertaken.
They are combining this behavioral science with things like GPS and street map information, vehicle speed and so on, to present an accurate prediction about what the driver might do.
The challenge is to bring all of these disparate data sources together so that they allow for quick and effective predictions to be made.
To do this, the team use AI to crunch the data and look out for the key, telltale signs that something important is about to happen.
Training the Machine
They trained the algorithm by first of all videoing drivers, with one camera monitoring their actions, and another trained on the road ahead.
With data from 10 different drivers and over 1,000 miles of driving over a two month period, the researchers set about annotating the data by hand to label what maneuvers were taking place in the vehicle.
This process identified 700 unique events, such as lane changes, turns and even carrying on in a straight line.
This was then used to train the algorithms to be able to accurately detect the circumstances whereby the driver will make turns or change lanes.
After a period of training, the algorithms were able to do this with an accuracy of around 90%. It was able to make these accurate predictions roughly 3.5 seconds before the action actually occurred.
Suffice to say, the algorithm was only on the hunt for a relatively limited range of behaviors over a relatively limited timescale. More work is needed, for instance, on coping with hazardous driving conditions or during challenging situations where visibility may be low.
These are the kind of situations where accidents are most likely, and therefore, where predictive capabilities will be most valuable in averting them.
So, whilst the development is a good indicator of the potential for this kind of predictive technology, there is still a lot of work to be done before it begins to be offered in the marketplace.
As vehicles become increasingly automated however, this is certainly an interesting area of exploration and one to keep a firm eye on.