Slashdot Mirror

← Back to Stories (view on slashdot.org)

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

Posted by EditorDavid on from the driving-in-the-dark dept.

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.

32 comments

  1. Can't Stump The Trump by Anonymous Coward · · Score: 0, Funny

    This was made possible by Trumps great America. Mr. president, I salute you!

    TRUMP 2020!

  2. Here's an idea by Anonymous Coward · · Score: 0

    Hey, is it too far fetched to try to create a virtual 3D works inside the computer based on the observed images? You might use motion compensation algorithms to determine which pixels move together given your vector and build your own 3D reality. Then apply the vector to your world and confirm that they're right when you make your next observation or adapt your world if there's a mismatch. If done correctly one should be able to build a real world copy map for a third-person shooter without involving any 3D modellers.

    I mean, that's how I imagine the human brain does it.

    1. Re:Here's an idea by Anonymous Coward · · Score: 0

      This kind of speculation about "how the human brain does it" is almost certainly incorrect, but is quite a popular idea. But let's talk about the problem of forming a 3D model of the world and trying to make decisions based on that model. Your model still has to understand the behavior of things within the model, obviously things that appear to be "cars" have predictable behavior. The model will also need to identify pedestrians in their multitude of forms, bicycles, scooters, large dogs, ladders, spare tires, bathroom cabinets that have fallen off the back of a truck. It is an exhausting list, and it still doesn't tell the system what it needs to do in response to those things, singly or collectively. What is a threat, and what can be disregarded, and more importantly what incorrect labels has the system applied? The rules that should be applied depend on the particular situation, and the identification of the situation depend on the particular rules that are in effect at that moment. You need actual, situated intelligence to make this work with any reliability, and there is no evidence that a camera with some feature detection software (a "convolutional deep learning network") is capable of rising to that level.

    2. Re: Here's an idea by Anonymous Coward · · Score: 0

      And that is exactly what apple maps and google maps did in 3D view

  3. fascinating! by Anonymous Coward · · Score: 0

    So when my self-driving car is suddenly unable to drive at night, do I take it to a mechanic or to a psychologist?

    "Hey doc, do you think this might have something to do with those convolutional things they are using now? Kin you fix it?"

    1. Re:fascinating! by ShanghaiBill · · Score: 1

      So when my self-driving car is suddenly unable to drive at night, do I take it to a mechanic or to a psychologist?

      Neither. You wipe the mud off the camera lens.

  4. 5 years? by fluffernutter · · Score: 5, Insightful

    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.

    --
    Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
    1. Re:5 years? by 110010001000 · · Score: 1

      Exactly. The hype around AI and autonomous cars is ridiculous.

    2. Re:5 years? by sims+2 · · Score: 2

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

      --
      Minimum threshold fixed. Thanks!
    3. Re:5 years? by Anonymous Coward · · Score: 0

      Yep, when you drop the Slashdot/Reddit circle jerk to push UBI and look at where we actually are, its obvious we have many hurdles which will take much longer than 5 years to solve.

      There's hype from romanticists and optimists like Musk fueling this perception that 5 years from now we'll all be in self-driving cars, and then there's reality. Of course the /. neckbeards will down vote this because they like to think software engineers are gods that have every problem solved.

    4. Re:5 years? by Kjella · · Score: 2

      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.

      --
      Live today, because you never know what tomorrow brings
    5. Re: 5 years? by Anonymous Coward · · Score: 0

      Driven by investment marketing companies targeting greedy investors and sensationalist media chasing advertising revenue. Wait a few more years and you'll just get shoulder shrugs while they harp about the next great thing.

    6. Re: 5 years? by Anonymous Coward · · Score: 1

      Eh, no. Usually the Slashdot crowd is far more skeptical, but we've been getting an increasing number of Reddit refugees the past couple years. Oh, and I guess we do get engineers for said companies and they do love drinking the koolaid!

    7. Re:5 years? by Anonymous Coward · · Score: 0

      5 years? You haven't been keeping up with the marketing puff.

      Elon Musk has promised a fully autonomous Tesla vehicle by the end of 2017...

      And of course we all know he never lies and always meets his targets.

    8. Re:5 years? by Anonymous Coward · · Score: 0

      My 2 year old Tesla has no problems reading speed limit signs.

    9. Re:5 years? by Anonymous Coward · · Score: 0

      Failed to read a sign? Just go by what the stored map says.
      Reading signs is only necessary to quickly update to temporary situations and human drivers fail to read signs all the time without causing major incidents.

    10. Re:5 years? by neoRUR · · Score: 1

      I worked on autonomous Humvee's about 15 years ago and that same technology is whats driving the cars now.
      The things that have changed are the sensors and computers are faster and better, the algorithms and NN's are better and deeper and there is so much data now.
      We already have autonomous cars and people are diving them and the stuff that they can detect even at fast moving speeds is quite a bit.
      They can definitely read signs and getting t\o be under all light, weather conditions.
      The higher level decision making processes are simple, but they will get better.
      I would say 5-7 years is about the right time frame.

    11. Re:5 years? by Marxist+Hacker+42 · · Score: 1

      I thought Beret Guy's company already did this

      --
      SJW: a person who perceives an injustice, and while correcting it, commits a greater injustice.
  5. Uh what? by 110010001000 · · Score: 2

    "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.

    1. Re:Uh what? by sysrammer · · Score: 1

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

      I'm assuming they've finally added parallel processing to the mix. Eventually the output of each stream will be sequentially handled by the mainline driver.

      --
      His ignorance covered the whole earth like a blanket, and there was hardly a hole in it anywhere. - Mark Twain
    2. Re:Uh what? by dinfinity · · Score: 2

      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.

    3. Re:Uh what? by sysrammer · · Score: 1

      Absolutely. I imagine this would be applicable to other areas of robotic vision. I notice that they also are able to handle varying illumination and sign occlusion better. I wonder how it would scale: Is it possible to more or less double the system to split the view into eighths instead of quarters? What's the point of diminishing returns?

      Looking at nature, a fly has thousands of prisms in its eyes. They're good for motion detection, not so much for identification, though. Spiders have 8 eyes, again not so good at images, but they do offload part of their processing to 3 gpu's (primarily light detection/direction).

      --

      And, back to reality. I have people to do and things to meet, and your link let me down the rat-hole (again).

      I thank you for the link. And now for something completely different.

      --
      His ignorance covered the whole earth like a blanket, and there was hardly a hole in it anywhere. - Mark Twain
  6. Oblig. by sysrammer · · Score: 2

    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.)

    --
    His ignorance covered the whole earth like a blanket, and there was hardly a hole in it anywhere. - Mark Twain
  7. Up the headlines from kindergarten level, please by RightwingNutjob · · Score: 1

    People who read slashdot ought to know that 'teach' is a meaningless filler word when it comes to AI and ought to be able to understand a slightly more technical headline.

  8. Re:Up the headlines from kindergarten level, pleas by religionofpeas · · Score: 1

    The headline comes straight from the source article. I'd rather see that than some lame attempt to translate it, and end up with something worse.

  9. convolutional neural network? by Anonymous Coward · · Score: 0

    NN's are great for machine learning, but there also famous for being black boxes that give every appearance of performing their intended function, but you can never quite be sure what you've actually taught them.

    When it comes to something that can kill a whole bunch of people Id rather they sorted out the autonomous bit the good old fashioned millions-of-lines-of-code way. Doubly so when it comes to the inevitable attribution of blame, "it wasn't me officer it was the car!" "it wasn't the car detective it was the software!" "it wasn't OUR software your honor, it was the neural network, seems it thought that group of nuns was a freeway"

  10. Re:Up the headlines from kindergarten level, pleas by RightwingNutjob · · Score: 1

    I'd rather see a well-executed attempt to translate it into adult language than a cut-and-paste any monkey can do. Which raises the question of why the new ownership of slashdot feels it can copy and paste baby talk from non-technical news sources at all. It didn't use to be that way.

  11. Re: Up the headlines from kindergarten level, plea by Anonymous Coward · · Score: 0

    "Selected designs, strategies, parameters and datasets" would be more accurate but not so snappy, but one of those sets is called the training set.

    What irks me more is articles saying "90%" accurate without reference to false positive and negative rates.

  12. Re:Up the headlines from kindergarten level, pleas by religionofpeas · · Score: 1

    If it's still going to have a link to the baby talk site, it's pointless to translate the title.

  13. Signs? Who cares about signs? by oddaddresstrap · · Score: 1

    At night out here in the boonies, I want my car to see Bambi, the deer that's trying to kill me.

  14. Once again by Anonymous Coward · · Score: 0

    I instantly ignore any headline that includes the word, 'may'. Conjecture in these matters really isn't worth reporting. Nothing to see, nothing to see.