Monday, September 2, 2013

3D or not 3D?


Auto-maker Ford has been making some noise about their “dirt detection system”. This uses 16 “digital cameras” and “reflected light” (now there’s an innovation!) to find tiny blemishes in painted car and truck bodies. The video below shows the system but is lacking in the kind of technical content that satisfies us machine vision specialists.

 


The vehicle bodies pass through an array of fluorescent tubes, which we see clearly reflected as white lines in the paint. Does this mean a form of triangulation is used to create a point cloud? Would that give sufficient resolution to find specks of dirt in the paint?

Would darkfield lighting be a more effective approach?

The associated press release, “The Dirt on Ford’s New 3D Dirt Detection Technology: There Isn’t Any”, is pretty clear about the 3D part, in which case the cameras, (which we are told run at 15 fps,) must be pretty high resolution.

Would anyone like to crunch the numbers to determine if this system can really see “microscopic” paint flaws?
 

3 comments:

Unknown said...

I didn't do the crunching to see what kind of 3D resolution they could achieve we this system but from what we can see at 0:42 they are using a JAI camera with a GigE interface, probably monochrome, running at 15 fps. Considering the camera housing form factor they must be using a JAI BM-500GE.

The optic in front of it is a 12.5mm Schneider cinegon lens. The "12-0906" marking on it is a total give away.

Supposing the distance to the camera is ~3m this gives a pixel resolution of about 1mm.

Unknown said...

I didn't do the crunching to see what kind of 3D resolution they could achieve we this system but from what we can see at 0:42 they are using a JAI camera with a GigE interface, probably monochrome, running at 15 fps. Considering the camera housing form factor they must be using a JAI BM-500GE.

The optic in front of it is a 12.5mm Schneider cinegon lens. The "12-0906" marking on it is a total give away.

Supposing the distance to the camera is ~3m this gives a pixel resolution of about 1mm.

R Hamilton said...

Creating a vision system is well and good but I would be interested in how they define a "defect" and how it is discriminated against the background.

Given the differing colors of the paint, the uneven nature of the body panels and the randomness of any defects (no "a priori" knowledge), it may be the easiest route to simply build a model of good paint jobs and look for anything that exceeds some predefined threshold. Odd reflections (flares) or high frequency distortions of the light edges could be detected.

Knowing what image processing algorithms are in use would help in the understanding.