TMC is an independent, primarily volunteer organization that relies on ad revenue to cover its operating costs. Please consider whitelisting TMC on your ad blocker or making a Paypal contribution here: paypal.me/SupportTMC

next step in replacement of lidar requirement for autonomous vehicles

Discussion in 'Autonomous Vehicles' started by kurdakov, Oct 8, 2016.

  1. kurdakov

    kurdakov Member

    Joined:
    Jul 25, 2016
    Messages:
    30
    Location:
    Moscow Russia
    cannot figure if following belongs more to 'technical' or to autonomous topic, but if anything wrong, mods will make right action.

    So the story is that, you know, Elon has idea, that radar can replace lidar for AP for all situations, while others think differently.

    From what we can see so far in behavior of autopilot - it can be claimed, that lidar really might be beneficial. it is not difficult to imagine that simple use of more spatial information can improve current autopilot behavior here and there.

    Yes, future vision algorithms like Nvidia demonstrates might provide elegant solution for current problems, still if there are more approaches, the better, besides there are other companies who seem to rely on radar potential Toyota Bets Big On Autonomous Tech, Swallows Millimeter Wave Radar Maker this high resolution radar won't appear till 20s so what about current Tesla radar?

    Seems there is one approach to get more spatial information from current hardware ( maybe additional polarization filter will be required for the visual camera ):
    is to use normals from polarized video feed to greatly enhance spatial resolution of radar depth information ( which is currently obtained and obviously used to successfully drive car with AP 8.0 )

    the paper is here
    http://web.media.mit.edu/~shiboxin/files/Kadambi_ICCV15.pdf ( it talks about kinect - so I suppose, that maybe Tesla engineers did not got a look, but as currently radar gets depth maps analogous to kinect and lidar - then why not to use developed algorithms?)

    the approach seems is more robust than previous attempts
    https://people.mpi-inf.mpg.de/~mzollhoef/Papers/SGASIA2014_SH/paper.pdf
    https://cseweb.ucsd.edu/~ravir/papers/hybrid/hybrid-final.pdf

    but those older approaches possibly might be used too.

    so it might be computationally intensive, but still I would try - maybe that would allow autopilot with similar to current hardware requirements to become much better.
     
  2. Tam

    Tam Active Member

    Joined:
    Nov 25, 2012
    Messages:
    1,047
    Location:
    Visalia, CA
    #2 Tam, Oct 12, 2016
    Last edited: Oct 12, 2016
    It may be an over-generalization to think that Tesla is against LIDAR.

    It is true that every single one of them who is now testing driverless system is using LIDAR except for Tesla Autopilot which is not a driverless system.

    But note what Elon said about LIDAR:

    "We do use LIDAR for our dragon spacecraft when docking with the space station. And I think it makes sense [in that case] and we’ve put a lot of effort into developing that. So it’s not that I don’t like LIDAR in general, I just don’t think it makes sense in a car context. I think it’s unnecessary."

    At this infancy stage that requires an alert and competent driver at the wheel for Autopilot, of course you do not need LIDAR just yet since a human would fulfill that function!

    So it looks like if you cannot disprove that LIDAR is "unnecessary" in a car, Tesla will continue to not consider LIDAR for its car.
     
  3. Tam

    Tam Active Member

    Joined:
    Nov 25, 2012
    Messages:
    1,047
    Location:
    Visalia, CA
    I think even Microsoft Kinect camera system is more sophisticated than current Tesla camera.

    Kinect is stereo which is a requirement for depth perception while Tesla camera is monocular which it cannot interpret depths.

    Kinect is full VGA color while Tesla is colorblind with monochrome capability.

    Kinect projects infrared markers to help to interpret depths as well even in a case of white object in a room against white wall ...
     
  4. TooManyAirMiles

    Joined:
    Aug 2, 2015
    Messages:
    30
    Location:
    Canada
    I think it is a cost and style issue as well. There is some work going on with much smaller solid state LIDAR systems that could really change that market. I don't think anyone (other then maybe Uber) feels that a top hat with rotating mirrors on a car is supportable as a general roll out. I cant imagine the maintenance issues - just not practical and really really ugly. Take a look at these sensors to see where this is going…. and this is generation 1. Can you imagine a roof design with 2 of these in the corners - but say 1/5 the size? I can.
    Disclosure - I have nothing to do with this company at all - just find it interesting.
    Products | Quanergy
     
    • Informative x 1
  5. Tam

    Tam Active Member

    Joined:
    Nov 25, 2012
    Messages:
    1,047
    Location:
    Visalia, CA
  6. Tam

    Tam Active Member

    Joined:
    Nov 25, 2012
    Messages:
    1,047
    Location:
    Visalia, CA
    This one also shows on figure 4 a stereo camera setup which Tesla is not.

    The question is if you want to replace current Tesla camera with these systems described on the papers, would that be an improvement?

    The problem is none of the papers has shown any real encounter of traffic nightmare scenarios.

    Thus, these papers might need more years of research in order to apply to the car industry.
     
  7. kurdakov

    kurdakov Member

    Joined:
    Jul 25, 2016
    Messages:
    30
    Location:
    Moscow Russia
    #7 kurdakov, Oct 12, 2016
    Last edited: Oct 12, 2016
    nonetheless authors of http://web.media.mit.edu/~shiboxin/files/Kadambi_ICCV15.pdf
    think that autonomous cars could benefit from this approach for not only to restore 3d scene but also see in rain and smog ( though that is still is in research phase )
    Polarization Data Enhances 3D Scanning Resolution

    they also progressed toward real time ( 30fps ) processing of scenes ( have some ideas on this Polarized 3D - Achuta Kadambi - MIT Media Lab )

    so, at least the approach has some potential for being explored more.
     
  8. Tam

    Tam Active Member

    Joined:
    Nov 25, 2012
    Messages:
    1,047
    Location:
    Visalia, CA
    But if we need to rush to a foreseeable future of 2020, we might need to count on other systems that are currently being trialed on the roads already.

    However, I do agree that these research papers do have potentials.
     

Share This Page