Researchers have developed a system that improves machine vision through obscuring clouds and fog.
Referred to as 'confocal diffuse tomography',
[t]he newly developed system works via an algorithm that measures the movement of individual light particles or photons, as fired in fast pulses from a laser, and uses them to reconstruct objects that are obscured or hidden from the human eye.
What makes the technique extra special is the way that it can reconstruct light that's been scattered and bounced around by the barrier in the way.
In experiments, the laser sight was able to see objects hidden behind a 1-inch layer of foam.
Existing techniques for machine viewing in similar scenarios have major drawbacks. Some only work at microscopic scales, some require access to both sides of the diffusing medium, and others require prior knowledge of the object being viewed. The most comparable method relies on excluding scattered photons by time-gating of ballistic (non-scattered) photons and using them to construct images, however this approach degrades rapidly for greater propagation distances and more highly scattering media as ballistic photons drop towards zero.
Confocal diffuse tomography is different in that it reconstructs images from the scattered photons "modeling and inverting the scattering of photons that travel through a thick diffuser." It has a variety of potential applications including self driving vehicles whose LiDAR struggles in rain and fog, robotic vision, and viewing planetary surfaces through cloud cover.
The researchers caution that the approach is currently slow and requires significant optimization, but they are "excited to push [the approach] further."
Journal Reference:
David B. Lindell, Gordon Wetzstein. Three-dimensional imaging through scattering media based on confocal diffuse tomography [open], Nature Communications (DOI: 10.1038/s41467-020-18346-3)
(Score: 1, Interesting) by fustakrakich on Sunday September 13 2020, @09:27PM (9 children)
Or raise it... There are other parts of the spectrum that pass easily through rain and fog to warn you there's a truck you're about to hit. Using more than one sensor gives you stereo vision, you should be golden. LIDAR still looks like Rube Goldberg hype to me. Who wants a big disco ball on their roof?
La politica e i criminali sono la stessa cosa..
(Score: 0) by Anonymous Coward on Sunday September 13 2020, @09:35PM
Disco, Stu; doesn't advertise.
(Score: 0) by Anonymous Coward on Sunday September 13 2020, @10:06PM (7 children)
People who aren't shallow wouldn't mind. The reality is that if the technology ever gets popular, it will be seen as a status symbol to have that "disco ball." The whole idea that ugliness is an issue is really only brought up by people shallow enough to not realize just how ugly fashionable things typically are. Seriously, look at photos of fashionable people from just about any decade and they are disproportionately wearing the uglies clothing available.
(Score: 1) by fustakrakich on Sunday September 13 2020, @10:18PM (6 children)
Just watch for low clearance, Clarence
La politica e i criminali sono la stessa cosa..
(Score: 0) by Anonymous Coward on Monday September 14 2020, @03:38PM (5 children)
It's hardly an issue for most vehicles, larger trucks may have issues, but the vast majority of vehicles wouldn't have issues with that and infrastructure would be rebuilt with that in mind over time. Same as they did with the heavier vehicles on the roads now.
(Score: 1) by fustakrakich on Monday September 14 2020, @04:08PM (4 children)
I prefer not to have any large protrusions sticking out of my car. The side mirrors are bad enough. LIDAR is a stop gap measure at best, unless they can make flush with the surface with no moving parts. We should be getting full spectrum sensors soon.
La politica e i criminali sono la stessa cosa..
(Score: 0) by Anonymous Coward on Monday September 14 2020, @09:51PM (3 children)
If you're hitting the sides or roof of your car, then you should be in favor of this. The device is clearly smarter than you are as a driver.
(Score: 1) by fustakrakich on Monday September 14 2020, @10:26PM (2 children)
I don't want to be the driver. The idea is that the car drives itself, and I play with the radio and the air conditioner and the electric windows.
La politica e i criminali sono la stessa cosa..
(Score: 0) by Anonymous Coward on Tuesday September 15 2020, @12:00AM (1 child)
> I play with myself
FTFY
(Score: 0) by Anonymous Coward on Tuesday September 15 2020, @12:41AM
Don't get caught!
(Score: 2) by fakefuck39 on Sunday September 13 2020, @09:46PM (4 children)
so this sounds like this:
let's send a bunch of photons through a cloud. most of them are going to bounce around, hit an object, then bounce around on their way back to the sensor. that takes time. statistically, like 2 photons are going to go straight through the cloud, and come back straight, without hitting any fog. those will make it back a lot quicker.
so send the pulse, and collect data on the sensor only up to the point the straight-shooting photons make it. if you set on a fixed distance - say 20ft, this works really well. and since light moves fast, you just send a new pulse every 2ms, and you know what's 20ft ahead, with that data refreshed every 2ms. now you know what's ahead of you.
but here's the problem - almost no photons make it through, so you need many, many pulses from that fixed position to collect enough data to know what's 20ft ahead. and the thicker the fog, the more pulses you need for that single 20ft datapoint.
cars be movin' y'all.
i'm not educated in this field, but here's my groundbreaking idea. use sonar with its nonexistent resolution to determine the object closest to you. then you set your pulse collection time to that distance. compensate for the car's speed and decrease the collection distance with each light pulse. this way you get a bunch of datapoints from light on that object. you can have multiple frequency sonic pulses going at the same time, and use several laser colors, and have them all going at the same time - letting you track several objects. now you have good resolution from the light, and you can be driving while it scans. yes, it'll double the cost of the car, but let's face it, someone who is that lazy that they'll pay extra for autopilot in the few cases of fog and heavy rain, will pay that price. and then maybe the system could be packaged as a black box and mass-produced, and cost will come down a lot.
(Score: 2) by captain normal on Monday September 14 2020, @04:53AM
Yep...and I just wonder how it will see through heavy smoke. (disclaimer- I live in central California and last Tue and Wed it was dark at 4:30 PM nearly 2 weeks before Autumn Equinox).
Everyone is entitled to his own opinion, but not to his own facts"- --Daniel Patrick Moynihan--
(Score: 0) by Anonymous Coward on Thursday September 17 2020, @06:24PM (2 children)
The speed of sound is actually quite low. Why use sound when there's already lots of EM stuff that goes through "conventional" fog and smoke? So much so that the military have to use special "smoke":
https://www.ncbi.nlm.nih.gov/books/NBK224564/ [nih.gov]
https://www.researchgate.net/publication/340684584_Multi_Spectral_Smoke_Obscurants_for_the_M1A1_Abrams_Tank_and_M88A2_Hercules_Recovery_Vehicle [researchgate.net]
I wonder how much the article's method differs from this:
http://web.media.mit.edu/~guysatat/fog/ [mit.edu]
The tech nowadays should be much better than this tech which was available years ago:
https://www.youtube.com/watch?v=hOTo-6jXOYo [youtube.com]
https://www.youtube.com/watch?v=Mm83bT9Fyeo [youtube.com]
On a vaguely related note: https://www.youtube.com/watch?v=c_0s06ORTkY [youtube.com]
(Score: 0) by Anonymous Coward on Thursday September 17 2020, @06:32PM
There used to be some around, guess they're not there anymore or Google etc no longer shows them.
(Score: 2) by fakefuck39 on Thursday September 17 2020, @10:04PM
the slower speed of sound is completely irrelevant, since you're detecting objects far away. You won't get close to them by the time the sound comes back. you're asking why put a simple cheap sonic sensor to simply detect the distance of any object ahead, when instead you can put in a second lidar for another 20k? I guess the same reason you drive to work 30min instead of taking your helicopter to a helipad close to work and take a cab from there.
(Score: -1, Redundant) by Anonymous Coward on Sunday September 13 2020, @10:00PM
I'll just leave these here if you are interested in the origin of the confocal microscope,
https://web.media.mit.edu/~minsky/papers/ConfocalMemoir.html [mit.edu]
https://www.chemistryworld.com/opinion/minskys-microscope/9419.article [chemistryworld.com]
(Score: 4, Insightful) by ledow on Monday September 14 2020, @07:50AM (5 children)
Should we be making cars that use clever tricks to trick to "see through" fog?
Or we should be making cars that slow down when they cannot safely see a good distance?
Seems everyone wants the former, and I can't fathom why that's the sensible course.
So now you have a car that drives itself through thick fog at speed, and the only backup is a human who can't see through the fog and just has to trust the computer. The computer that's trusting the human to take over if something goes wrong.
It's like blindfolding both pilot and copilot and then expecting them to fly at full speed.
(Score: 0) by Anonymous Coward on Monday September 14 2020, @08:13AM (3 children)
Why would you even drive other than level 0 (no automation) or level 5 (full automation) car? There's no user backup in level 5.
And who says it doesn't slow down in fog? If it can't detect what's in front, then it shouldn't be going there. If it moves at all, it's going to need "tricks" to see through the fog anyway.
(Score: 1, Redundant) by ledow on Monday September 14 2020, @09:49AM (1 child)
I agree.
However an automated car racing through the fog in which other driver's cannot see and thus are moving slowly and carefully is more dangerous than a car that drives the same way a human would.
Personally, I do not understand the obsession with having them drive on "normal" roads alongside human drivers. It seems stupid and a backwards step and a way to introduce dangers that never existed before (e.g. the above).
Make a special lane, put a code/tape/signal on the lane somehow (does not need digging up the roads at all). Isolate that lane to automated cars only. Then they don't have to worry anywhere near as much about what a human driver does, and the technology is far simpler and more reliable.
Reliance on computer vision is dumb.
Driving alongside humans is dumb.
Driving a vehicle designed for a human is dumb (you're limited to what interface a human would use, which is unnecessarily restrictive).
Expecting the driver to take over is dumb.
Or you could just treat them like automated trains (which we've had for decades) which follow very simple routes, lanes, orders, and capabilities, and you could have had them 20 years ago with ancient, reliable tech.
And why we just immediately went from nothing to "you can drive on a fast freeway" without, say, using them for automated low speed theme park transport, airport transport, golf buggies, shopping assistance, etc. for a few years first to get the hang of things in a controlled environment at a safe speed that you could build trust on, I honestly cannot fathom.
(Score: 0) by Anonymous Coward on Monday September 14 2020, @03:43PM
None of that is realistic though. In the places that most people drive, there aren't extra lanes for this and there isn't space to build more.
Perhaps that would work out on city to city travel, but that's a relatively small amount of total traffic.
(Score: 0) by Anonymous Coward on Monday September 14 2020, @01:42PM
> Why would you even drive other than level 0 (no automation) or level 5 (full automation)
Seconded, but note that (as I understand it), Level 4 might be OK if it works at a trip level:
* Car, take me to 221B Baker Street
+ Sorry Sherlock I can't find that address, you will have to drive yourself
* Car, take me to 219–229 Baker Street and drop me off (then go find a place to park on your own)
+ OK
And similar go/no-go if road conditions or weather also prevent Level 4 from making the full trip.
Level 3 implementations that let the automation give up and hand-off to the human driver on short notice should never be allowed on the road, imo.
(Score: 0) by Anonymous Coward on Monday September 14 2020, @03:40PM
We can already slow down in fog, doing that is a half-step. Part of the point of using these sensors is that they aren't chained to our senses and as such don't inherently need to slow when it's dark or in poor visibility.