Chapter 3
Active 3D Imaging Systems
Marc-Antoine Drouin and Jean-Angelo Beraldin
Abstract Active 3D imaging systems use artificial illumination in order to capture and record digital representations of objects. The use of artificial illumination allows the acquisition of dense and accurate range images of textureless objects that are difficult to acquire using passive vision systems. An active 3D imaging system can be based on different measurement principles that include time-of-flight, triangulation and interferometry. While time-of-flight and interferometry systems are briefly discussed, an in-depth description of triangulation-based systems is provided. The characterization of triangulation-based systems is discussed using both an error propagation framework and experimental protocols.
3.1 Introduction
Three-dimensional (3D) imaging systems (also known as 3D vision systems) capture and record a digital representation of the geometry and appearance (e.g. colortexture) information of visible 3D surfaces of people, animals, plants, objects and sites. This digital surrogate of the physical world is then processed in order to extract useful information from the raw data and finally, communicate the results. Active 3D imaging systems use an artificial illumination, usually either a spatially coherent light source (e.g. laser) or an incoherent one (e.g. halogen lamp), to acquire dense range maps with a minimum of ambiguity. The term is also used for systems that project non-visible electromagnetic radiation, such as near infra-red onto the scene. The use of an artificial light source makes it possible for active 3D imaging systems to generate a model of a surface geometry even when the surface appears featureless to the naked eye or to a photographic/video camera and, hence, require minimal operator assistance. Furthermore, the 3D information can be made relatively insensitive to ambient illumination and surface color. They are, by their
M.-A. Drouin ( ) · J.-A. Beraldin
National Research Council of Canada, Ottawa, Ontario, Canada
e-mail: Marc-Antoine.Drouin@nrc-cnrc.gc.ca
J.-A. Beraldin
e-mail: Jean-Angelo.Beraldin@nrc-cnrc.gc.ca
N. Pears et al. (eds.), 3D Imaging, Analysis and Applications, |
95 |
DOI 10.1007/978-1-4471-4063-4_3, © Her Majesty the Queen in Right of Canada 2012 |
|
96 |
M.-A. Drouin and J.-A. Beraldin |
nature, non-contact measurement instruments and produce a quantifiable 3D digital representation (e.g. point cloud or range image) of a surface in a specified finite volume of interest and with a particular measurement uncertainty.
3.1.1 Historical Context
The desire to capture and record shape using optical instruments can be traced back to the invention of rudimentary surveying instruments and the camera obscura [60]. The invention of photography, in which images are recorded on semipermanent recording media, is certainly the catalyst of modern methods. In the 1860s, François Willème invented a process known as photo-sculpture that used many cameras [49, 65]. Profiles of the subject to be reproduced were taken on photographic plates, projected onto a screen, and transferred to a piece of clay using a pantograph. The process supplied many profiles, which were used to rough down the piece of clay, leaving a large amount of manual work. Commercial applications developed rapidly and studios stayed in operation from 1863 to 1867, when it was realized that the photo-sculpture process was not more economical than the traditional sculpture technique. A professional sculptor was needed and the photo-sculpture process required a significant investment in cameras, projection and reproduction systems, and skilled labour to operate them.
It is only with the advances made during the last 50 years in the field of solidstate electronics, photonics, computer vision and computer graphics that the process of capturing and recording detailed shapes by optical means regained substantial interest. Indeed, obvious changes have been instrumental in the growth of active 3D imaging systems technology i.e. the availability of affordable and fast digital computers and reliable light sources (lasers, halogen lamps, LED). It is now possible to build reliable, accurate, high-resolution 3D active vision systems that can capture large amounts of 3D data. In addition, the ability to process these dense point clouds in an efficient and cost-effective way has opened up a myriad of applications in areas as diverse as military, medical, entertainment, industrial and commercial activities.
3.1.2 Basic Measurement Principles
Active three-dimensional (3D) imaging systems can be based on different measurement principles. The three most used principles in commercially available systems are time-of-flight, interferometry and triangulation. Seitz describes time-of-flight as based on an accurate clock, interferometry as one that uses accurate wavelengths and triangulation as a method based on geometry[59]. Figure 3.1 summarizes the typical accuracy of each type of active 3D imaging system technology found on the market as a function of the operating distance. It can be observed from that figure that each optical technique covers a particular range of operation. Many in-depth classifications of optical distance measurement principles have been published in important references in the field of 3D vision, e.g. [14, 16, 43, 50].
3 Active 3D Imaging Systems |
97 |
Fig. 3.1 Diagram showing typical accuracy at different operating distances of the most common active 3D imaging technologies for surface digitization. Figure courtesy of [56]
The fundamental work on time-of-flight systems can be traced back to the era of RADAR, which is based on radio waves. With the advent of lasers in the late 1950s, it became possible to image a surface with angular and range resolutions much higher than possible with radio waves. Different strategies have been devised to exploit this basic measurement principle of time-of-flight [7, 10, 43]. Figure 3.1 shows two of them.
Interferometry is based on the superposition of two beams of light [43]. Typically, a laser beam is split into two paths. One path is of known length, while the other is of unknown length. The difference in path lengths creates a phase difference between the light beams. The two beams are then combined together before reaching a photo-detector. The interference pattern seen by the detector resulting from the superposition of those two light beams depends on the path difference (a distance). Note that commercially available systems based on other principles such as conoscopic holography are available for small operating distances (see Fig. 3.1).
The remainder of this chapter will focus on triangulation-based methods which use the same principle as the passive triangulation systems presented in the previous chapter. The reader can find more information on interferometry-based methods in [43] and time-of-flight methods are described in [7, 10, 14, 43].
3.1.3 Active Triangulation-Based Methods
In the previous chapter, passive triangulation systems were presented, namely standard stereo configurations and structure-from-motion. In this chapter, we assume
98 |
M.-A. Drouin and J.-A. Beraldin |
that the reader is familiar with camera calibration and the epipolar geometry of two camera views, as presented in Chap. 2.
Both active and passive triangulation systems are based on the same geometric principle: intersecting light rays in 3D space. Typically, an active system replaces one camera of a passive stereo system by a projection device. This projection device can be a digital video projector, an analogue slide projector or a laser. (Note, however, that many active systems use two cameras with a projection device. Although this is at first sight redundant, there are sound reasons behind this design choice, such as a reduction in ‘missing parts’ due to self-occlusion.)
There are many ways to classify active triangulation sensors, according to their opto-mechanical components, construction and performance. One of the key dimensions within this taxonomy is the way in which the active 3D imaging system illuminates the scene. Here we will consider three distinct categories: spot scanners, stripe scanners and systems that use structured light patterns.
The simplest of these is the spot scanner (also known as a point-based scanner) where, typically, a collimated or focused laser beam illuminates a very small circular or elliptical part of the scene for each image capture. One advantage of this approach is that the spatial correspondence problem is non-existent, because the illumination of the object’s surface is spread temporally (i.e. in the time dimension). Moreover, a spot scanner allows us to control the spatial sampling on the scene surfaces. The laser power can also be controlled on a per 3D sample basis. However, this is at the expense of additional opto-mechanical complexity because the spot must be scanned either by mounting the sensor on a 2D translation stage, or by orienting the laser around two axes of rotation using two galvanometer-mounted mirrors.
Typically, in the second type of scanner, a collimated laser beam is passed through a cylindrical lens in order to generate a ‘sheet of light’ which illuminates the scene with a thin stripe. Other implementations are possible, for example the cylindrical lens could be replaced by a diffractive optical element (DOE) or a diffraction grating. This stripe now only needs to be scanned in one direction relative to the scene in order to assemble a range image and, again, this may be done by translation of the sensor (or, alternatively, the object) or the rotation of a single mirror. These 3D imaging devices are called stripe scanners or profile scanners. Note that although the complexity of scanning has reduced from two-dimensional to one-dimensional, some ambiguity in the direction of illumination is introduced, which needs to be resolved using the epipolar constraint.
The final type of scanner that we discuss is the type that projects a structured light pattern onto the scene [43]. These 3D imaging devices are also known as area scanners [2]. Typically these systems do not scan the projected light over the scene object at all, since the object is usually completely illuminated by the pattern, although the term ‘scanner’ is often still applied in an informal sense. These systems provide the advantage of the shortest capture times, thus minimizing distortion due to motion in dynamic scenes. The correspondence problem, however, is more challenging (although not as difficult as for passive stereo) and the projected light is structured either spatially, temporally, or both in order to determine