Slashdot Mirror
Deep Learning Identifies Wet Road Hazards From Sound Input (thestack.com)
An anonymous reader writes: Researches have used recurrent neural network architecture to develop an audio-interpretation system that can understand how wet a road is, using techniques more commonly employed in speech recognition and music analysis. Every year 384,032 persons are injured and 4,789 persons killed through wet roads, and it's a problem that also threatens to hamper the usefulness of self-driving cars, which are likely to either become dangerous or prohibitively cautious in the absence of good information about the safety of road surfaces.
Every year 384,032 persons are injured and 4,789 persons killed through wet roads
That's just the US. But hey, y'know, no-one else counts right?
some moaning & further overmedicating is anticipated?
" Every year 384,032 persons are injured and 4,789 persons killed through wet roads" should read:
Every year 384,032 persons are injured and 4,789 persons killed through wet roads and their inability to grasp the concepts of friction and velocity.
Or: ...are killed because of their piss poor driving skills.
09 F9 11 02 9D 74 E3 5B - D8 41 56 C5 63 56 88 C0 45 5F E1 04 22 CA 29 C4 93 3F 95 05 2B 79 2A B2
Is this just another excuse to install mics in every car? We already have backup and other cameras. Welcome to the surveillance society.
So how much worse would a manually seeded pattern correlation filter have worked for this? When you drive through water, the sound produced is quite distinctive and seems reasonably clearly correlated with the amount of water being displaced by the thread. I haven't looked at this signal in matlab, but is it really such a complex one that it is a prime candidate for deep learning algorithms?
I guess I just think it is easy to say 'deep learning' and then everyone thinks AI is just around the corner, when the reality is that many deep learning implementations are barely as good as tuned filters right now. Of course this is how you need to start the research, and if this has really been produced by a generic algorithm that was not tuned then that is cool, but these sorts of stories don't inspire me that transformational AI is close at all.
Many years ago, I remember hearing about a project to identify mechanical problems by audio recognition. The idea was that a computer could listen to an engine and tell whether a cylinder was misfiring, for example. I wonder whatever came of that?
-jcr
The only title of honor that a tyrant can grant is "Enemy of the State."
I fail to see how this is news. Wet roads produce a lot of high frequency 'noise' rather than dry roads that have little to no noise (unless maybe on certain types of gravel)
This sounds like the net simply managed to distinguish HF noise from LF noise.
Just ban driving on wet roads. Problem solved, because bans, prohibition, and regulations work.
"Every year 384,032 persons are injured and 4,789 persons killed through wet roads..."
No, these numbers are ten-year averages. It's highly unlikely that precisely the same number of people are injured and killed, every year.
If an accident happens, the accident is attributed to a SET of causes and then statistics claim each on of those clauses separately as causal.
So you have a crash, and it was unavoidable (e.g. random mechanical failure). Suppose you were speeding at the time, had a beer in you, maybe the radio was playing. Then that crash won't only be flagged as caused by the mechanical failure, it will be added to the {speeding} set, the {even low alcohol causes accidents} set, and the {radio is distracting too} set.
So "wet roads" is unlikely to be "wet roads" but rather "generally driving faster than the viable road speed", or "badly drained piece of road that causes aqua-planing".
There's no attempt to identify the primary cause, because double and triple counting is a useful lobbying tool.
This deep learning stuff is going to be huge one day
Please login to access my lawn
Unless there is extremely deep pools of water following the speed limit has enough built in safety factor to operate safely when it is wet.
love is just extroverted narcissism
Every year 384,032 persons are injured and 4,789 persons killed through wet roads,
Yet another reason to ban dihydrogen monoxide!
Here's some answers the artificial intelligence came up with:
How Wet Is It? ....it's so wet, the term "Ark-Industrial-Complex" has just been added to the Oxford Dictionary. ....it's so wet, the moisturizer jar gets more full with each use. ....it's so wet, my ride-on mower is now a float-on. ....it's so wet, it *must* be a liberal conspiracy to undermine the credibility of climate change deniers. ....it's so wet, I saw a fish in a lifeboat. ....it's so wet, my in-laws are vacationing in England for three weeks to dry out. ....it's so wet, I can let my goldfish play out in the yard. ....it's so wet, I saw a duck in a rain slicker. ....it's so wet, the water bugs in my basement are building an ark. ....it's so wet, I'm swimming laps in the hallway. ....it's so wet, I saw a squirrel wrapping his nuts in Saran Wrap. ....it's so wet, my umbrella needs a raincoat.
Blind drivers need to know how wet the road really is.
I wonder, what kind of roads and tyres are involved here?
I can't remember having a problem with wet roads while driving in a way that wouldn't be seriously uncomfortable - unless there's a few inches of water on the road or I'm driving crazy cars (like that 700hp Cadillac test car with slicks that didn't want to move with or without traction control). Am I just getting way too good tyres?
Unless the "wet" is frozen. But that's a completely different game.
But if anyone builds me a car that warns me of black ice in advance, I'd like it.
It's really simple. Police say it until they're blue in the face. Drive to the conditions and slow down. Just because you have all wheel drive, traction control, anti lock breaks, auto breaking, etc. doesn't give you license to drive 20% over the speed limit when everyone is going 50% under because the roads suck. Of course there will be the kidiot whining that he rolled his new suv and the highway department didn't do its job. Wah.
"Every security scheme that is based on secrets eventually fails." - Steve Jobs
Every year 384,032 persons are injured and 4,789 persons killed through [strike]wet roads[/strike] idiocy.
FTFY.
With all the accidents that occur, how much more likely is accidents during wet conditions? Seems to me at least a few of those accidents would still occur if the roads were dry. Though this research does have implications for self-driving like described int he article.