Researchers Teach Self-Driving Cars To 'See' Better At Night (sciencemag.org)

Researchers may have developed a way for self-driving cars to continue navigating at night (or on rainy days) by performing an AI analysis to identify traffic signs by their relative reflectiveness. Slashdot reader sciencehabit shares an article from Science: Their approach requires autonomous cars to continuously capture images of their surroundings. Each image is evaluated by a machine learning algorithm...looking for a section of the image that is likely to contain a sign. It's able to simultaneously evaluate multiple sections of the image -- a departure from previous systems that considered parts of an image one by one. At this stage, it's possible it will also detect irrelevant signs placed along roads. The section of the image flagged as a possible sign then passes through what's known as a convolutional neural network [which] picks up on specific features like shapes, symbols, and numbers in the image to decide which type of sign it most likely depicts... In the real world, this should mean that an autonomous car can drive down the street and accurately pinpoint and decipher every single sign it passes.

Comments

5 years?

[fluffernutter]

We're not going to have self driving in 5 years if they can't even freaking read signs at night yet! Further proof that automated driving is much further back then we are led to believe.

Re:5 years?

[sims 2]

5 years huh?
Relevant xkcd.
https://xkcd.com/678/

Re:5 years?

[Kjella]

Somehow I'm not so concerned about this one, if you can follow the rules of the road in daytime and the sun is shining it's the same rules when it's night and raining. The rest is "just" a sensor sensitivity/noise cancellation problem that can be worked on in parallel to everything else. You can probably do a lot with combination LIDAR/optical systems to make LIDAR identify candidate surfaces then do optical do actually identify the sign. And you're looking for a predetermined number of surfaces of particular sizes, if you have identified the shape it should be possible to correlate candidates rather than try to analyze. You probably also have a lot of temporal data that could be used to enhance the search, after all traffic signs generally don't move or change. I'd be much more concerned with everything else that's out there, is it pedestrians or wild animals or a tree falling over the road that doesn't have any particularly known shape or size or color.

Uh what?

[110010001000]

"It's able to simultaneously evaluate multiple sections of the image -- a departure from previous systems that considered parts of an image one by one".

Wait, uh, this is cutting edge AI? What autonomous system can't evaluate multiple sections of an image

"convolutional neural network [which] picks up on specific features like shapes, symbols, and numbers in the image to decide which type of sign it most likely depicts." Uh, what? You mean the have an algorithm that can decide on types of street signs based on the image? Wow. Truly cutting edge. Autonomous cars are truly right around the corner.

Re:Uh what?

[dinfinity]

You are correct; it's worded terribly.

The paper ( http://journals.plos.org/ploso... ) specifically mentions that their method allows for GPGPU processing due to parallel block evaluation instead of 'sliding window' evaluation of the image:
"Unlike the sliding-window method, which scans an image in a sequential manner, parallel window-searching divides the input image into several blocks and simultaneously performs classification on one block using each GPU core."

This is a step forward in commodification of self-driving car technology.

Oblig.

[sysrammer]

Ewe must be gnu hear. Links to XKCD may be relevant, they may be insightful, even pithy; but in all cases they must be obligatory. (The rules do allow abbreviation, but the inclusion of the word is...well, you know.)