Drones (and by that we mean actual, self-flying quadcopters) have come a long way. Newer ones have cameras capable of detecting fast moving objects, but aren’t yet capable of getting out of the way of those objects. However, researchers at the University of Zurich have come up with a drone that can not only detect objects coming at them, but can quickly determine that they’re a danger and get out of the way.
The drone has cameras and accompanying algorithms to detect the movement in the span of a couple of milliseconds, rather than the 20-40 milliseconds that regular quad-copters would take to detect the movement. While regular cameras send the entire screens worth of image data to the copter’s processor, the cameras on the University’s drone are event cameras, which use pixels that detect change in light intensity and only they send their data to the processor, while those that don’t stay silent.
Since these event cameras are a new technology, the quadcopter processor required new algorithms to deal with the way the data is sent. After testing and tweaking, the algorithms are fast enough that the ‘copter can determine that an object is coming toward it and move out of the way.
It’s great to see the development of new techniques that will make drones better and more stable for the jobs they will do. It’s also nice that one day, we can fly a drone around without worrying about the neighborhood kids lobbing basketballs at them. While you’re waiting for your quadcopter delivered goods, check out this article on a quadcopter testbed for algorithm development.
Those Skynet drones just got a little harder to take out
“If you can dodge a ball, you can dodge anti-drone measures!” – Patches O’Hoolihan, probably.
Maybe…
Personally I’d be impressed if you sprayed the ball silver (Saying “You shall ride eternal shiny and chrome” of course) and lobbed it and it managed to dodge that.
You can be impressed even like this – what’s described in the paper is no small feat :) But of course it has it’s limitations as you correctly point out.
Curious what the specs and prices are like for “event camera” sensors (and why aren’t they called “event sensors”?). I’m guessing similar to FLIR sensors.
Background.. https://en.wikipedia.org/wiki/Event_camera
Yup, read that article, light on capabilities and no refs to manufacturer sites or retail sites.
Inivation makes three models.
Here are the specs: https://inivation.com/wp-content/uploads/2020/03/iniVation-Specifications.pdf (320×240 or 640×480, 12 to 165 million events per second, <1 ms latency, 200 microsecond accuracy)
Here are the prices: https://shop.inivation.com/collections/frontpage (2,000 to 5,000 euros)
“Don’t move! It’s vision is based on movement…”
There is a natural version of this, of sorts: some of spiders’ many eyes are largely light-dark detection that very rapidly assess nearby movement. That warns the spider, which can then orient its body to allow its narrow focal width but very high resolution eyes to acquire an image. Since they have rigid bodies, including eyes, they’re restricted to fairly minor eye movement by physically displacing the retina of the eye to a different point, so body orientation is a major way for them to get their best eyes in location to see things they need to see. Sensor integration from all eight eyes seems to allow them to form good spatial models with 360 degrees of vision above them even though they only have a 7 degree field of good vision. If you had to design a quadcopter with good visual acuity and very high visual speed, a spider eye system seems like a really good way of doing it.