Self-Driving / Autonomous Cars: General discussion

Discussion in 'General' started by gooki, Oct 9, 2018.

To remove this ad click here.

  1. Pushmi-Pullyu

    Pushmi-Pullyu Well-Known Member

    It astounds me that some other company is following Tesla down the rabbit hole of trying to use cameras as the primary sensors for self-driving cars.

    Can cameras see in the dark? No, no better than the human eye can. (In fact, I've seen arguments that ordinary video cameras can't match the human eye's ability to discern fined shades of gray, which means the night "vision" is even worse.)

    Can cameras discern the distance to objects, or directly discern where one object ends (that is, where the edge of an object is) and other begins? Heck no! A self-driving system has to use software to interpret those video images, hopefully figuring out distances and figuring out which objects are where. But one doesn't have to do much research on the subject to learn that this is problematic and rather error-prone, despite decades of effort on the part of roboticists.

    To clarify: The problem isn't due to any lack of resolution on the part of the cameras. When comparing to the human vision system, the limitation isn't the camera, it's the image processing. The human brain has a highly developed visual cortex, a sophisticated dedicated image processing center which is the result of billions of years of evolution. The computer has maybe 2 or 4 general-purpose microprocessors running some software, with far less processing power. That's hardly a fair comparison!

    It truly boggles me that some companies are ignoring the very clear advantages of active scanning using lidar and/or phased-array, high-res radar. With active scanning, you get an instant error-free read on the distance to objects, as well as detecting their shape; no need to rely on unreliable optical object recognition software which may or may not be able to figure out where the objects are, and how far away.

    And of course, active scanners don't "care" if it's day or night; if it's bright sunlight or pitch black. Unlike cameras and the human eye, they aren't limited to "seeing" only in the direction headlights point... and the inadequacy of headlights used in nearly all cars is something the IIHS has been talking rather loudly about recently.

    Now, there is one legitimate criticism of lidar: It degrades in rain worse than camera images. Okay, so use phased-array high-res radar instead. Problem solved.

    Even if optical object recognition could be made reliable, how could they possibly overcome the limitation that cameras can't see in the dark any better (or perhaps even worse) than the human eye? Will they mount dozens of headlights all around the car, so the self-driving system can (as will be required) see in all directions, rather than merely a narrow cone directly ahead of the car? And will they increase the brightness of those headlights, so they will adequately illuminate the landscape for a sufficient distance from the car that the self-driving system will have enough time to react... unlike the Uber car which ran down and killed a woman pushing her bicycle across the road in the dark? And if they do start using much brighter headlights, how much worse will the problem be of human drivers being blinded by the headlights of oncoming cars? And if those self-driving cars use illumination in a 360° circle, the problem will also occur with cars that you're following.

    The more one looks at the problem, the worse it looks (in my opinion) for trying to depend on cameras as the primary sensors, day and night.

    As I've said many times, the goal of those developing self-driving cars should not be to slavishly imitate a human driver. If human drivers were capable of driving safely, then we wouldn't need self-driving cars. Self-driving cars should be better and safer than humans at driving. One way they can be safer is by using active scanners, and not cameras, as the primary sensors.
     
    Last edited: Dec 19, 2018
  2. To remove this ad click here.

  3. bwilson4web

    bwilson4web Well-Known Member Subscriber

    I'm open minded about this emerging technology but lean towards optical systems. The Tesla optically based systems are more widely deployed than the lidar based systems. Furthermore, Tesla has active data capture back to the developers to rapidly improve the systems. Furthermore, Tesla has announced custom processors to improve processing. This rapid, feedback between design and deployment gives the optically based, Tesla system an advantage.

    Based on my Prius Prime experience, I do like radar based, distance measurements to augment the optical sensors. They can reach out through fog but lack the precision needed to make an accurate map.

    Bob Wilson
     
  4. Pushmi-Pullyu

    Pushmi-Pullyu Well-Known Member

    Radars currently used in production cars use only low-res Doppler radar, not a high-res (or more precisely, higher-res) phased radar array. See my previous posts in this thread.

     
  5. gooki

    gooki Well-Known Member

    The solution to seeing in the dark is infra red lights to compliment the stand lights on the vehicle.

     
    Last edited: Dec 19, 2018
  6. gooki

    gooki Well-Known Member

    I disagree with this. Humans at their peak are perfectly capable of driving vehicles safely. The issue is our performance is not consistent.

    It's the inconsistency of our performance that autonomous driving can easily fix. An AV solution should never get drunk, tired, impatient, or distracted.
     
    Roy_H likes this.
  7. To remove this ad click here.

  8. gooki

    gooki Well-Known Member

    I'd be very interested in a comparison of Lidar vs camera plus radar in fog. I suspect at this stage of development, both of them will be ****.
     
  9. Pushmi-Pullyu

    Pushmi-Pullyu Well-Known Member

    I'd like to point out that humans continue to drive in conditions under which it's impossible to drive safely. Driving at night when it's raining, when literally all you can see is headlights and tail-lights, is very far from safe driving conditions. That used to terrify me when I was still driving. (I suffer from a lack of depth perception... just like current self-driving cars, which generally don't use binocular "vision", but rather rely on analyzing individual images for distance cues.)

    Likewise, humans continue to drive in heavy downpours where the only rational act would be to pull over, park, and wait for the rain to lessen. Even worse is the habit that people have of driving into a fog bank, sometimes without slowing down... and other times slowing substantially. On a highway in heavy traffic, that's a recipe pretty much guaranteed to result in multiple accidents, and I think we've all seen news reports of multi-car pileups on the freeway under those conditions.

    No, absolutely not. Under any of those conditions, relying on visible light cameras is hopelessly inadequate, where human lives are dependent on the safe operation of the self-driving car. Unless you're going to claim that "performance" means refusing to continue driving under those circumstances, Gooki, then your argument doesn't hold up to scrutiny.

    If self-driving cars are going to be safer than human drivers, then they must use active sensors.

     
  10. Pushmi-Pullyu

    Pushmi-Pullyu Well-Known Member

    I just ran across an article which claims there is a software solution to the problem with rain and snow scattering lidar scanning. If true, that would certainly remove one of the major objections to using lidar.

    "Driverless cars have a new way to navigate in rain or snow"

    I'm somewhat skeptical of the claims here since I haven't read about this anywhere else. But it does suggest that perhaps the problems with lidar may be reduced or possibly even eliminated with the right system.

     
  11. bwilson4web

    bwilson4web Well-Known Member Subscriber

    Interesting: https://insideevs.com/waymo-chrysler-pacifica-hybrids-under-attack/

    People in Arizona are fighting back against the rise of the machines, specifically Waymo’s autonomous Chrysler Pacifica minivans testing there. There have been at least 21 incidents of people attacking the vehicles in the last two years, including slashing tires and throwing rocks at them, according to The New York Times.

    The LIDAR on top draws attention to the cars. Road rage against self-driving cars ... great!

    Bob Wilson
     
  12. To remove this ad click here.

Share This Page