3d Machine Vision Systems
Machine vision refers to applications in which the PC automatically makes a decision based on visual input from a camera. Machine vision is a term typically used in industrial manufacturing, where applications range from culling blemished oranges from a conveyor belt to saving lives by inspecting to ensure that the correct drug capsule has been placed in the package before the product is shipped to the pharmacy. Three dimensional vision based measurement systems have made their presence into production metrology applications, notably in the electronics field. However, in the more traditional fields of durable goods now dominated by hard gauges and CMMs, 3D optical systems has been hindered by perceptions and real limitations. This paper will review where 3D vision is today, and what advances have been made to enable more quantitative, shop floor metrology applications. The field of 3D machine vision is a less established field, but one that is actively growing today. Three dimensional vision based measurements have come a long way in the past few years, moving from purely visualization tools that generate attractive color pictures, to serious measurement tools. These 3D systems include laser scanning, structured light, stereo viewing, and laser radar just to name a few
Machine vision in general has been used for everything from guiding the insertion of electronic chips on circuit boards to inspecting bottles at several per second in bottling lines. A natural extension of machine vision inspection is to provide programmable measurements for machined parts. In many applications, these measurements can be made in two dimensions for which there is an established based of machine vision tools working in the sub-thousandth of an inch range at multiple measurements per second. Each of these methods has their strong points and weak points for a given application.
The key performance parameters needed for durable good manufacturing include:
• Resolution in the mil, and sub-mil range
• Speeds sufficient to complete all measurements in a few seconds
• Ability to look at a wide range of surface types and finishes
This last point, the ability to look at a wide range of surface finishes has perhaps been the biggest limitation of 3D machine vision technology. In many cases, the surface needs to be diffuse to permit reflected light to be easily seen to achieve a good signal to noise ratio
The most popular commercial versions of range finding use the triangulation method where a beam of light is projected onto the object's surface at some angle and the image of this spot or line of light is viewed at some other angle (see Figure 1). As the object distance changes a spot of light on the surface will move along the surface by:
(change in spot position) = (change in distance) x (tan(incident angle) + tan(viewing angle))
Good resolution of a few tens of microns have been realized with these systems. Most triangulation gages today use laser light. When a laser beam is incident on an opaque, rough surface, the micro-structure of the surface can act as though it is made of a range of small mirrors, pointing in numerous directions. These micro-mirrors may reflect the light off in a particular direction, as generally machine marks do, or may direct the light along the surface of the part. Depending on how random or directional the pointing of these micro-mirrors may be, the apparent spot seen on the surface will not be a direct representation of the light beam as incident on the part surface. The effects that may be seen from a laser beam reflecting off a rough surface include:
• directional reflection due to surface ridges
• skewing of the apparent light distribution due to highlights
• expansion of the incident laser spot due to micro surface piping
The result of this type of laser reflection or "speckle" is a noisy signal from some surfaces. In like manner, there can be a problem with translucent surfaces, as the laser light will scatter through the medium and produce a false return signal. For a laser based sensor, a smooth, non-mirror like, opaque surface produces the best results. An active variation of restricting the view uses synchronized scanning. In the synchronized scanning approach (see figure 2), both the laser beam and viewing point is scanned across the field. In this manner, the detector only looks at where the laser is going
A more extreme variation of the synchronized scanning type of approach is to use an active point seeking triangulation system (see figure 3). With an active point seeking triangulation system, a single point detector is limited to a very narrow view, at a constant angle, which is then translated across the path of the laser beam.
As has already been stated, the key operational parameters needed for production machine vision include speed, resolution, and robustness especially to changing part surface conditions. Many systems that provide the best resolution are not the fastest, so a tradeoff must be made. Just as with touch probes, there are certain types of features or surfaces that optical 3D methods can be expected to work good on, and others where there may be problems. If has been pointed out that shiny, but not specular surfaces have offered one of the biggest challenges. In like manner, when a surface changes from a shiny area to a dull, many sensors may generate a bias error. In the simple case of triangulation, the measurement is based upon finding the centroid of a light spot of some finite size. If half that spot is on an area that reflects back to the sensor well, and the other half is not, the center of brightness of the spot will not be the geometric center, but rather weighted toward the brighter region. Testing the sensor on edge and surface transition features is a valuable first test to consider.
The next area of concern is the surface texture itself. A surface with machining marks has a texture which may scatter light into long lines that may confuse the sensor. In a similar manner, if the surface is translucent, the spot of light may spread out in an unpredictable manner, again offsetting the spot center, and hence the measurement made. Therefore, testing the sensor on a surface texture that matches the one to be measured is important. A final important feature consideration for many optical gages is the slope of the surface. A standard way to test such effects is to scan a diffuse sphere and fit the data to the known diameter (see Figure 7).At some angle around the sphere, the surface can be expected to lift off the true surface, before the signal is lost entirely. As with any gage, comparison of the optical gage against other measurement tools provides valuable information regarding confidence of the measurements. This is not always an easy comparison, as the repeatability of many traditional gages may not be as good as the optical gage.
Anytime one is trying to improve a process, one encounters the challenge of demonstrating a capability against a tool with less capability. The very high speeds and data densities available from optical gages offer some significant advances. Yet trying to compare those advantages against a gage which stakes everything on a half dozen points is a difficult road. However, the comparisons and qualification of the new process must be done. Comparisons against inadequate measurement tools can prove very frustrating, so at least one independent means of comparison is desirable. The use of reference artifacts that can be verified by independent means, preferable with primary standard reference (e.g., the laser distance interferometer is the international standard of length measurement), is a valuable aid in these comparison
1Machine vision systems can be used in various fields for numerous purposes as stated below:
Industrial inspection (Inspecting machine parts, Adaptive inspection systems)
Vision for Autonomous Vehicles (Detecting obstructions, Exploring new surroundings, AV surveillance)
Transport (Traffic monitoring, Aerial navigation, Transport Safety)
Surveillance(Intruder monitoring, Number plate identification, Tracking people)
Vision and Remote Sensing (Land Management, Crop classification, Surveying by satellite)
Medical (Planning radiation therapy, Chromosome analysis, Merging medical images)
As with any technology of this nature, the performance changes with the component technology. The primary advance that has made machine vision systems feasible for shop floor gauging applications has been the speed up in computing power. This has brought the processing times from 15 or 20 minutes on an expensive workstation to seconds on a standard PC. The other technologies that are influencing performance today include lower cost, digital cameras than provide better light range and pixel resolution with lower noise, and better light sources such as higher power laser diodes well as higher brightness and resolution LCD projectors. The consumer market largely influences all of these technologies, which is currently a much bigger driver than any manufacturing support system. However, as system prices decrease and performance improves, there is a wide range of new markets these systems will likely address ranging from dentistry to home 3D pictures