Computer Vision & LA Auto Show Debrief


Agenda, Attendee List, & Presentation files now available to Autotech Council members in the library.

Autotech Council’s meeting on Computer Vision will look at the latest advances of various technologies that comprise Computer Vision, and startups and vendors with specific solutions that are applicable to autonomous cars, and the automotive industry. After lunch, members stay to debrief the 2018 LA Auto Show. These debrief sessions mix news with opinions for those members who did not attend, while those who did actively contribute their insights to recent trends.


  • Date: 12/14/2018 08:30 AM
  • Location: SRI International, Menlo Park, CA, USA (Map)

Description

SUMMARY   |   AGENDA    |    ATTENDEES   |   LIBRARY

This month, Autotech Council members opens their discussion on Computer Vision to the public. Join Autotech Council members, OEMs, automotive suppliers, mobility startups and VCs to discover the current state of innovation in Computer Vision - a field that touches on optics, sensors, machine learning, and AI.

While humans have had millions of years to evolve our visual cortex to where we can instantly recognize shapes and objects, direction, and intent - even when occluded or in shade, the complexity of this task is huge. But Computer Vision is being tackled by suppliers and startups alike because it is essential to continuing to advance Self-Driving Vehicles as safe replacements for human drivers.

For the automotive space, Computer Vision often combines visual images from cameras with a 3D point cloud from a LIDAR, RADAR, or some other echolocation technology. By combining the 3D point cloud, the ability of machines to perform edge extraction, size estimation, and object recognition increases significantly, as well as the obvious benefits of ranging and velocity detection. But the addition of LIDAR, in particular, raises the cost curve substantially, so the ability to use Computer Vision, solely, to create an accurate 3D model of the world around the car could substantially lower the cost of equipping a vehicle for autonomous driving. (This is the Elon Musk approach.)

A proper self-driving car needs to copy a few specific things we humans do. It needs to:

  • Have some knowledge of where it is and where it wants to go, which for the car is represented by a 3D model of the world like a semantic map in HD.
  • It needs to know where it is in the world, and in relationship to that 3D HD point map.
  • It needs to manage the other road users, and avoid hazards. For this, it needs a 3D Perceptual System which detects, identifies, and vector-maps other objects around it.

Once it has achieved these things, the self-driving software can act on this understanding of the physical world to act and react and safely maneuver the vehicle to its destination.

The latter two bullets rely heavily on the field of Computer Vision, to make sense of the world around the vehicle. Elements of that job include:

  • Specialized cameras or other sensors
  • Scene reconstruction
  • Event detection
  • Edge extraction
  • Object recognition
  • Object Classification
  • 3D pose estimation
  • 3D point cloud, and SLAM
  • Intent or motion estimation
  • Machine learning
  • Training and fleet learning

Autotech Council’s meeting on Computer Vision will look at the latest advances of various technologies that comprise Computer Vision, and startups and vendors with specific solutions that are applicable to autonomous cars, and the automotive industry.