Machine vision for industrial applications is still a relatively young industry. The technology has, however, advanced at a great pace and here Sony Image Sensing Solutions' Stéphane Clauss and Tordivel's Thor Vollset examine advances in 3D vision applications and highlight the trade-offs in system specifications.

Back in the early days of the machine vision industry the systems were, to be blunt, simple. In the late nineties these basic measurement systems were based solely on pixel counts. You'd find the edge, count the number of pixels and that would give you the measurement. This meant the resolution was limited to the number of pixels, and the VGA sensors gave a total of 640x480 pixels. The next steps were to move towards calibrated systems that allowed sub-pixel measurements to be made, a 20x improvement in resolution. And by 2004 2D machine vision systems had become very accurate, that is, provided all items being captured were in the same plane. Once they aren't, flat errors are introduced and these are almost impossible to detect. 3D opens the way to avoid this problem, giving a method to detect movement, allowing you to verify fundamental dimensions and therefore prevent any wrong selections.

3D system requirements

Naturally the complexity of any given project will vary greatly and, as such, 3D machine vision systems do too. There are, therefore, trade-offs to be made between resolution and processing capabilities on the one hand and budget on the other. The three key requirements, however, are a good camera, a multi-core PC and efficient software.
For the camera, image quality is obviously the key requirement, and Tordivel specifies Gig E cameras from Sony; the XCG-97E (black/white, 1.3 megapixel super XGA resolution) or the XCG-U100CR (color, 2MP UXGA resolution) are used for the majority of its 3D systems. For low light applications, the XCG-H280 is used and for projects that require even greater levels of precision, the Sony XCG-5005E (black/white, 5MP resolution) is selected. But 3D doesn't always require megapixel images and for more simple applications Tordivel specifies the XCG-V60E (VGA 0.3MP resolution) camera modules as a low cost, entry-level alternative.
For the PC, the introduction of multi-core processing allows the output from each camera to be analyzed simultaneously on a separate core. For simple applications, this means data can be processed in less than 1s.
And, like cameras, the PC should be selected for the application. For simple applications, a PC with a dual-core low-power mobile processor and Windows 7 will suffice. But systems with 3 or more cameras will, naturally, require additional cores and power. As to the software, the more efficient (or less bloated) it is, the quicker it can run and the lower the overall system cost.

Application example

One of the more advanced systems to be based on Sony's machine vision cameras and Tordivel's software is a robot vision application that controls an ABB robot to select gears during the manufacturing process. It was created for a major European truck maker's assembly process in order to automate the gear assembly process and increase production capacity.
This system uses two Scorpion 3D Stinger cameras, each equipped with two XCG-97E Gig E cameras from Sony (4 in total), outputting to a Scorpion Industrial PC running the latest 6-core Intel processors. With no training, the system is able to identify all the different gears to an accuracy of 1mm in xyz from a picking area of 800x600x600 mm. The software also allows for real-time selection and location of moving objects.

What does the future hold?

The capabilities of both camera and PC technologies for any given cost point are improving constantly and the software is evolving to match these. 3D systems enable location rates greater than 99.9 per cent. Improving the accuracy 10x with 3D sub-pixel measurements can increase the value of a system or lower the hardware cost significantly and can also make innovative designs feasible.
As machine vision opens to a greater variety of customers, rather than solely industrial manufacturers, simplicity becomes essential. Last year Sony and Tordivel launched an App Repository (akin to the Apple app store, but for machine vision applications) to meet this need and, in the near future, this will contain apps that allow customers to more simply carry out a specific facet of 3D image analysis too.