EyeDriver Lets Drivers Steer Car With Their Eyes

Hugh Pickens writes "NPR reports that German researchers have tested a new technology called eyeDriver that tracks a driver's eye movement and, in turn, steers the car in whatever direction they're looking at speeds up to 31 mph. 'The next step will be to get it to drive 60 miles per hour,' says Raul Rojas, an artificial intelligence researcher at Berlin's Free University. A Dodge Caravan fitted with eyeDriver has been tested on the tarmac at an abandoned airport at Tempelhof Airport. However, it remains unclear when — or if — the technology will be commercialized, as questions about safety and practicability abound: What about looking at a cute girl next to the road for a few seconds? Not to mention taking phone calls or typing a text while driving. But the researchers have an answer to distracted drivers: 'The Spirit of Berlin' is also an autonomous car equipped with GPS navigation, scores of cameras, lasers, and scanners that enable it to drive by itself. And should the technology-packed vehicle have a major bug, there's still an old fashioned way of stopping it. Two big external emergency buttons at the rear of the car allow people outside to shut down all systems."

Comments

Re:Boobies

[tibit]

It's not even that "abnormal distractions would be bad" -- it would be completely, absolutely crazy to drive like that.

Landing a plane, on the other hand -- that I could potentially agree on. Some studies show pilots staring at the far end of the runway say from 200ft down to ground contact, so that could potenitally work. It's sort of a reflex thing they do in visual conditions.

Driving on a long stretch of straight road sometimes looks like that too, when you analyze the eye movements.

But in "normal" driving, not only the instantaneous eye position is uncorrelated with desired steering wheel input, but also long-term averages are generally uncorrelated.

Basically, to drive with eye position as the control input is to become blind, to a large extent. I have played a little bit with using eye movement in various control input scenarios, and the only conclusion I came to (in an informal study) was that you can obviously learn rather well to use eye position as the control input. Heck, with audio feedback you can learn even faster, but you become progressively more oblivious to what's going on around you. This may, perhaps, be thought to be "OK" at first sight, as you'd think it rather keeps you focused on a particular area of the surroundings -- just that on curves it usually ends up being anywhere but on the road. So you literally feel like driving in a tunnel. Forget looking at street signs, or navigating in unfamiliar environment. And you better never had to change lanes.

Remember that the raison d'être of our visual system is exploration of our environment. This also happens when we drive. Eye position will depend on what interesting stuff is out there, not on which way you are driving.

There's a lot of "eye movments to control this or that" type of studies. Unfortunately, the idea came from science fiction, and belongs on the same shelf with "cleaning up" SD interlaced surveillance video full of compression artifacts to "clearly" see a face that's six pixels across. Such studies certainly have a lot of appeal to the general public, and to anyone who doesn't quite think it through or understand the basic conflicts of purpose involved in using eye movments for something they just can't do while simultaneously maintaining visual awareness. This is the same fantasy as using eye movments for interacting with machine user interfaces: all fine and dandy, as long as you don't need to see/explore the damn interface. If you have the UI all memorized, and ideally are provided with audible cues to help you navigate, you can use eye movments. But the moment you want to look around, it becomes all screwed up.

Now, in situations where you don't give a damn about maintaining visual input -- you can use eye movements for whatever control inputs you please, and they are quite good for that. Heck, the input is at least 3-dimensional: you can choose not only the direction vector, but to some extent the amplitude of the initial saccade.

So -- eye movements are great for controlling a car, as long as you're in the passenger seat, and the driver makes sure you won't run over the old lady, and won't drive off the end of a closed bridge -- IOW, as long as you don't need to actually see most of what's outside the window.

There are of course ways to devise some special patterns of eye movements that switch the modality of the controlled device/interface, so that you can work around the conflict between controlling and visual exploration. But those hardly feel natural. Those are very fine things to do if the alternative is even worse -- say, if you're paralyzed and all that's left is eye motion. But without training and adaptation the eye movement control has anything but "natural feel" to it.