Complexity and reliability: a contact measuring machine has many moving parts, subject to wear, and several different subsystem: mechanical, pneumatic, electric, electronic. Optical measuring machines have few or no moving part; their complexity lies mainly in software and their failure probability is low.

Safety: no contact, no moving parts means no possibility to damage the object being measured nor the operator.

Precision: contact measurement ranges from submicron to tens of microns, while optical measurement ranges from microns to millimeters. Anyway it must be said that only a minority of applications needs a micron or submicron precision; in most cases, the precision delivered by optical measurement is more than sufficient.

 

State of the art

Optical measurement is not one single technology but rather several different ones, covering different needs and partially competing where they overlap. Here we list the most widespread ones.
Optical contact systems represent the link between traditional and optical measurement. An optically recognizable contact probe, comprising a touch tip and a pattern (eg an array of LEDs), is carried around by hand or by a robot and put in touch with the object. The contact can be signalled via a wireless connection, while the position of the device is measured optically. In this case we have some advantage over traditional contact measurement in terms of flexibility and volume of measure, while throughput and precision stay low and the contact with the object is not avoided.
Laser trackers direct a laser beam onto the object and measure the time to get it back. Due to the high value of the speed of light, whence the need to measure very short times of flight, they are intrinsically not very precise; they also are sensitive to surface texture and usually do not behave well on edges. On the other hand, their range can be very large, up to kilometers, of course with decreasing precision, down to centimeters.
Laser scanners also direct a laser beam or blade onto the object, but they read it back by means of a camera which is spatially separate from the laser source. In this way a triangulation can be performed. Precision is usually high (tens of microns, some claims for microns) and range is in the order of meters.
Vision based measuring systems do not rely on a single moving spot like laser trackers, nor a pattern of spots like optical contact systems, and neither a line like laser scanners; they rather make use of couples or triples of whole images of the object, processed with machine vision algorithms for feature extraction. Triangulation is then used to reconstruct three-dimensional positions from matched features.
The advantage of this approach is evident: an image can yield thousands to millions features, and its acquisition time is very fast. This means a potentially very high throughput and less sensitivity to movement or vibration.

 

However the development of a vision based measuring system presents several challenges, which are not found (or, at least, not all together) in other optical systems:

• Objects have different surface textures. What can be recognized as a feature? And how to deal with smooth surfaces, which exhibit no feature at all?
• Matching corresponding features in different images is far from obvious; in fact, in the case of repetitive patterns, it can be difficult for the human eye too.
• Even sensors with the highest resolution available today would be insufficient to obtain an acceptable precision over an acceptable measuring volume if the pixel was the 'quantum of space'; it is necessary to exploit the grey level (or color) information with what is called subpixel interpolation.
• Lens distortions must be precisely compensated to avoid significant errors.
• Processing large images is time-consuming.

A first classification of vision based measuring systems takes into account feature extraction. We have passive feature extraction when the object exhibits well recognizable features, by its nature as in the case of cuts on a metal sheet, or thanks to adhesive or magnetic markers; we have active feature extraction when features are artificially created on a smooth surface by projecting some pattern of light.
A second classification deals with methods for matching corresponding features in different images. This can be done manually, by asking the operator to select corresponding features, or automatically. In this second case we must exploit some kind of similarity between features (eg their orientation) and some intrinsinc constraints: uniqueness (the correspondence of features is one-to-one) and continuity. Examples are sampling of disparity space, dynamic programming and Marr-Poggio matching.
If features are artificially obtained by projecting a suitable light pattern, there are in turn several possibilities. The light pattern can be modulated in space (sinusoidal intensity shape) so that correspondences can be obtained by determining their phase; or it can simply consist of 'light blades', with abrupt transitions light/dark; or it can be a 'coded' pattern, where correspondences are obtained by matching the codes.
A third classification concerns the calibration method. In order to perform measurements the optical system must be calibrated; that is, its intrinsic parameters (pixel size and slant, position of the optical centre, distortion coefficients) and extrinsic parameters (relative positions and orientations of the cameras) must be known. Such parameters can be obtained by reading a suitable, certified pattern before measurements, or they can be obtained during measurement ('self calibration') by matching a sufficient number of corresponding points, of which only a few have known distances. Self calibration is particularly suitable for wide fields of measure, where a very large certified calibration pattern would not be practical.

 

EyeGauge

How does EyeGauge (figure 1) fit inside the scene of three-dimensional vision based metrology? We have some technical innovation; as an example, calibration is performed with a proprietary nonlinear bundle adjustment algorithm taking into account all parameters (intrinsic, extrinsic and distortion) at once. The result is a precise calibration with an effective distortion correction (several radial and decentering coefficients) allowing to work with normal quality lenses, even short focus. For applications where the 3D reconstruction must really be fast, eg in-line metrology, EyeGauge can take advantage of the fact that the kind of computations performed are very suitable for parallelization. Parallel number crunchers like the new nVidia ® Tesla ® GPUs speed up the reconstruction of about two orders of magnitude.

figure 1: An axonometric view of the measuring head, trinocular version, via EyeGauge

The most notable strength of EyeGauge is anyway its flexibility. The same product can manage both objects exhibiting well recognizable edges, like cut metal sheets (figure 2), even bent or cast, and smooth surfaces (figure 3), just by varying the type of illumination, which is uniform in the first case and 'shaped' (light blades) in the second. In both cases, illuminators or pattern projectors made by Opto Engineering are the best choice.

Moreover, EyeGauge doesn't came with a fixed measuring volume. The cameras can easily be reoriented and recalibrated for volumes ranging from a few cubic centimeters to 600x400x200 mm; and at the time of writing (December 2008) we are completing the self-calibration feature, for measurement of large surfaces with projected patterns used both for calibration (together with a linear scale) and measurement.