school of optics/creol, and school of electrical engineering and...

Jannick P RollandSchool of OpticsCREOL andSchool of Electrical Engineering andComputer ScienceUniversity of Central Florida

Orlando FL 32816ndash2700

Henry FuchsDepartment of Computer Science

University of North CarolinaChapel Hill NC 27599-3175

Optical Versus Video See-ThroughHead-Mounted Displays in MedicalVisualization

Abstract

We compare two technological approaches to augmented reality for 3-D medical

visualization optical and video see-through devices We provide a context to discuss

the technology by reviewing several medical applications of augmented-reality re-

search efforts driven by real needs in the medical eld both in the United States and

in Europe We then discuss the issues for each approach optical versus video from

both a technology and human-factor point of view Finally we point to potentially

promising future developments of such devices including eye tracking and multifocus

planes capabilities as well as hybrid opticalvideo technology

1 Introduction

One of the most promising and challenging future uses of head-mounteddisplays (HMDs) is in applications in which virtual environments enhancerather than replace real environments This is referred to as augmented reality(Bajura Fuchs amp Ohbuchi 1992) To obtain an enhanced view of the real en-vironment users wear see-through HMDs to see 3-D computer-generated ob-jects superimposed on their real-world view This see-through capability can beaccomplished using either an optical HMD as shown in figure 1 or a videosee-through HMD as shown in figure 2 We shall discuss the tradeoffs betweenoptical and video see-through HMDs with respect to technological and hu-man-factor issues and discuss our experience designing building and testingthese HMDs in medical visualization

With optical see-through HMDs the real world is seen through half-trans-parent mirrors placed in front of the userrsquos eyes as shown in figure 1 Thesemirrors are also used to reflect the computer-generated images into the userrsquoseyes thereby optically combining the real- and virtual-world views With avideo see-through HMD the real-world view is captured with two miniaturevideo cameras mounted on the head gear as shown in figure 2 and the com-puter-generated images are electronically combined with the video representa-tion of the real world (Edwards Rolland amp Keller 1993 State et al 1994)

See-through HMDs were first developed in the 1960s Ivan Sutherlandrsquos1965 and 1968 optical see-through and stereo HMDs were the first computergraphics-based HMDs that used miniature CRTs for display devices a me-chanical tracker to provide head position and orientation in real time and ahand-tracking device (Sutherland 1965 1968) While most of the develop-ments in see-through HMDs aimed at military applications (Buchroeder

Presence Vol 9 No 3 June 2000 287ndash309

r 2000 by the Massachusetts Institute of Technology

Rolland and Fuchs 287

Seeley amp Vukobratovich 1981 Furness 1986Droessler amp Rotier 1990 Barrette 1992 Kandebo1988 Desplat 1997) developments in 3-D scientificand medical visualization were initiated in the 1980s atthe University of North Carolina at Chapel Hill (Brooks1992)

In this paper we shall first review several medical visu-alization applications developed using optical and videosee-through technologies We shall then discuss techno-logical and human-factors and perceptual issues relatedto see-through devices some of which are employed in

the various applications surveyed Finally we shall dis-cuss what the technology may evolve to become

2 Some Past and Current Applications ofOptical and Video See-Through HMDs

The need for accurate visualization and diagnosisin health care is crucial One of the main developmentsof medical care has been imaging Since the discovery ofX-rays in 1895 by Wilhelm Roentgen and the first X-rayclinical application a year later by two Birmingham (UK)doctors X-ray imaging and other medical imaging mo-dalities (such as CT ultrasound and NMR) haveemerged Medical imaging allows doctors to view as-pects of the interior architecture of living beings thatwere unseen before With the advent of imaging tech-nologies opportunities for minimally invasive surgicalprocedures have arisen Imaging and visualization can beused to guide needle biopsy laparoscopic endoscopicand catheter procedures Such procedures do requireadditional training because the physicians cannot see thenatural structures that are visible in open surgery Forexample the natural eye-hand coordination is not avail-able during laparoscopic surgery Visualization tech-niques associated with see-through HMDs promise tohelp restore some of the lost benefits of open surgery(for example by projecting a virtual image directly onthe patient eliminating the need for a remote monitor)

The following paragraphs briefly discuss examples ofrecent and current research conducted with optical see-through HMDs at the University of North Carolina atChapel Hill (UNC-CH) the University of CentralFlorida (UCF) and the United Medical and DentalSchools of Guyrsquos and Saint Thomasrsquos Hospitals in En-gland video-see-through at UNC-CH and hybrid opti-cal-video see-through at the University of Blaise Pascalin France

A rigorous error-analysis for an optical see-throughHMD targeted toward the application of optical see-through HMD to craniofacial reconstruction was con-ducted at UNC-CH (Holloway 1995) The superimpo-sition of CT skull data onto the head of the real patientwould give the surgeons lsquolsquoX-ray visionrsquorsquo The premise of

Figure 1 Optical see-through head-mounted

display (Photo courtesy of KaiserElectro-Optics)

Figure 2 A custom optics video

see-through head-mounted display

developed at UNC-CH Edwards et al

(1993) designed the miniature video

cameras The viewer was a large FOV

opaque HMD from Virtual Research

288 PRESENCE VOLUME 9 NUMBER 3

that system was that viewing the data in situ allows sur-geons to make better surgical plans because they will beable to see the complex relationships between the boneand soft tissue more clearly Holloway found that thelargest registration error between real and virtual objectsin optical see-through HMDs was caused by delays inpresenting updated information associated with track-ing Extensive research in tracking has been pursuedsince at UNC-CH (Welch amp Bishop 1997)

One of the authors and colleagues are currently devel-oping an augmented-reality tool for the visualization ofhuman anatomical joints in motion (Wright et al 1995Kancherla et al 1995 Rolland amp Arthur 1997 Parsonsamp Rolland 1998 Baillot amp Rolland 1998 Baillot et al1999) An illustration of the tool using an optical see-through HMD for visualization of anatomy is shown infigure 3 In the first prototype we have concentrated on

the positioning of the leg around the knee joint Thejoint is accurately tracked optically by using three infra-red video cameras to locate active infrared markersplaced around the joint Figure 4 shows the results ofthe optical superimposition of the graphical knee jointon a leg model seen through one of the lenses of ourstereoscopic bench prototype display

An optical see-through HMD coupled with opticaltracking devices positioned along the knee joint of amodel patient is used to visualize the 3-D computer-rendered anatomy directly superimposed on the real legin motion The user may further manipulate the jointand investigate the joint motions From a technologicalaspect the field of view (FOV) of the HMD should besufficient to capture the knee-joint region and the track-ing devices and image-generation system must be fastenough to track typical knee-joint motions during ma-

Figure 3 (a) The VRDA tool will allow superimposition of

virtual anatomy on a model patient (b) An illustration of the

view of the HMD user (Courtesy of Andrei State) (c) A

rendered frame of the knee-joint bone structures animated

based on a kinematic model of motion developed by Baillot

and Rolland that will be integrated in the tool (1998)

Figure 4 First demonstration of the superimposition of a graphical

knee-joint superimposed on a leg model for use in the VRDA tool (a) a

picture of the benchprototype setup a snapshot of the superimposition

through one lens of the setup in (b) a diagonal view and (c) a side view

(1999)


nipulation at interactive speed The challenge of captur-ing accurate knee-joint motion using optical markerslocated on the external surface of the joint was addressedby Rolland and Arthur (1997) The application aims atdeveloping a more advanced tool for teaching dynamicanatomy (advanced in the sense that the tool allowscombination of the senses of touch and vision) We aimthis tool to specifically impart better understanding ofbone motions during radiographic positioning for theradiological science (Wright et al 1995)

To support the need for accurate motions of the kneejoint in the Virtual Reality Dynamic Anatomy (VRDA)tool an accurate kinematic model of joint motion basedon the geometry of the bones and collision detectionalgorithms was developed (Baillot amp Rolland 1998Baillot et al 1999) This component of the research isdescribed in another paper of this special issue (Baillot etal 2000) The dynamic registration of the leg with thesimulated bones is reported elsewhere (Outters et al1999) High-accuracy optical tracking methods care-fully designed and calibrated HMD technology and ap-propriate computer graphics models for stereo pair gen-eration play an important role in achieving accurateregistration (Vaissie and Rolland 2000 Rolland et al2000)

At the United Medical and Dental Schools of Guyrsquosand Saint Thomasrsquos Hospitals in England researchersare projecting simple image features derived from preop-erative magnetic resonance and computer-tomographyimages into the light path of a stereo operating micro-scope with the goal of eventually allowing surgeons tovisualize underlying structures during surgery The firstprototype used low-contrast color displays (Edwards etal 1995) The current prototype uses high-contrastmonochrome displays The microscope is tracked intra-operatively and the optics are calibrated (includingzoom and focus) using a pinhole camera model Theintraoperative coordinate frame is registered using ana-tomical features and fiducial markers The image featuresused in the display are currently segmented by handThese include the outline of a lesion the track of keynerves and blood vessels and bone landmarks Thiscomputer-guided surgery system can be said to be

equivalent to an optical see-through system operating ona microscopic scale In this case the real scene is nowseen through magnifying optics but the eye of the ob-server is still the direct detecting device as in optical see-through

One of the authors and colleagues at the UNC-CHare currently developing techniques that merge videoand graphical images for augmented reality The goal isto develop a system displaying live real-time ultrasounddata properly registered in 3-D space on a scanned sub-ject This would be a powerful and intuitive visualizationtool as well The first application developed was the visu-alization of a human fetus during ultrasound echogra-phy Figure 5 shows the real-time ultrasound imageswhich appear to be pasted in front of the patientrsquos bodyrather than fixed within it (Bajura et al 1992) Real-time imaging and visualization remains a challenge Fig-ure 6 shows a more recent non-real-time implementa-tion of the visualization in which the fetus is renderedmore convincingly within the body (State et al 1994)

Recently knowledge from this video and ultrasoundtechnology has also been applied to developing a visual-ization method for ultrasound-guided biopsies of breastlesions that were detected during mammography screen-ing procedures (Figure 7) (State et al 1996) This ap-plication was motivated from the challenges we observedduring a biopsy procedure while collaborating on re-search with Etta Pisano head of the Mammography Re-search Group at UNC-CH The goal was to be able tolocate any tumor within the breast as quickly and accu-rately as possible The technology of video see-through

Figure 5 Real-time acquisition and superimposition of ultrasound

slice images on a pregnant woman (1992)


already developed was thus applied to this problem Theconventional approach to biopsy is to follow the inser-tion of a needle in the breast tissue with a remote moni-tor displaying real-time 2-D ultrasound depth imagesSuch a procedure typically requires five insertions of the

needle to maximize the chances of biopsy of the lesionIn the case in which the lesion is located fairly deep inthe breast tissue the procedure is difficult and can belengthy (one to two hours is not atypical for deep le-sions) Several challenges remain to be overcome beforethe technology developed can actually be tested in theclinic including accurate and precise tracking and atechnically reliable HMD The technology may haveapplications in guiding laparoscopy endoscopy or cath-eterization as well

At the University of Blaise Pascal in Clermont Fer-rand France researchers developed several augmented-reality visualization tools based on hybrid optical andvideo see-through to assist in surgery to correct scoliosis(abnormal curvature of the spine column) (PeuchotTanguy amp Eude 1994 1995) This application was de-veloped in collaboration with a surgeon of infantile sco-liosis The visualization system shown in figure 8 is froman optics point of view the simplest see-through systemone may conceive It is first of all fixed on a stand and itis designed as a viewbox positioned above the patient

Figure 6 Improved rendering of fetus inside the

abdomen (1994)

Figure 7 Ultrasound guided biopsy (a) Laboratory setup during

evaluation of the technology with Etta Pisano and Henry Fuchs (b) A

view through the HMD (1996)

Figure 8 Laboratory prototype of the

hybrid opticalvideo see-through AR tool

for guided scoliosis surgery developed by

Peuchot at the University of Blaise

Pascal France (1995)


The surgeon positions himself above the viewbox to seethe patient and the graphical information is superim-posed on the patient as illustrated in figure 9 The sys-tem includes a large monitor where a stereo pair of im-ages is displayed as well as half-silvered mirrors thatallow the superimposition of the real and virtual objectsThe monitor is optically imaged on a plane through thesemi-transparent mirrors and the spine under surgery islocated within a small volume around that plane An op-tical layout of the system is shown in figure 10

In the above hybrid optical-video system vertebraeare located in space by automatic analysis of the perspec-tive view from a single video camera of the vertebrae Astandard algorithm such as the inverse perspective algo-rithm is used to extract the 3-D information from theprojections observed in the detector plane (Dhome etal 1989) The method relies heavily on accurate videotracking of vertebral displacements High-accuracy algo-rithms were developed to support the application includ-ing development of subpixel detectors and calibrationtechniques The method has been validated on vertebral

specimens and accuracy of submillimeters in depth hasbeen demonstrated (Peuchot 1993 1994)

The success of the method can be attributed to thefine calibration of the system which contrary to mostsystems does not assume a pinhole camera model forthe video camera Moreover having a fixed viewer withno optical magnification (contrary to typical HMDs)and a constant average plane of surgical operation re-duces the complexity of problems such as registrationand visualization It can be shown for example that ren-dered depth errors are minimized when the virtual im-age planes through the optics (a simple semi-transparentmirror in Peuchotrsquos case) is located in the average planeof the 3-D virtual object visualized (Rolland et al1995) Furthermore the system avoids the challengingproblems of tracking optical distortion compensationand conflicts of accommodation and convergence re-lated to HMDs (Robinett amp Rolland 1992 Rolland ampHopkins 1993) Some tracking and distortion issueswill be further discussed in sections 31 and 32 respec-tively However good registration of real and virtualobjects in a static framework is a first step to good cali-bration in a dynamic framework and Peuchotrsquos resultsare state of the art in this regard

It is important to note that the method developed forthis application employs a hybrid optical-video technol-ogy In this case video is essentially used to localize realobjects in the surgical field and optical see-through isused as the visualization tool for the surgeon While thefirst system developed used one video camera the meth-ods have been extended to include multiple cameras

Figure 9 Graphics illustration of current and future

use of computer-guided surgery according to Bernard

Peuchot

Figure 10 Optical scheme of the hybrid

opticalvideo see-through AR tool shown in Fig 8


with demonstrated accuracy and precision of 001 mm(Peuchot 1998) Peuchot chose the hybrid system overa video see-through approach because lsquolsquoit allows the op-erator to work in his real environment with a perceptionspace that is realrsquorsquo Peuchot judged this point to be criti-cal in a medical application like surgery

3 A Comparison of Optical and VideoSee-Through Technology

As suggested in the description of the applicationsdescribed the main goal of augmented-reality systems isto merge virtual objects into the view of the real scene sothat the userrsquos visual system suspends disbelief into per-ceiving the virtual objects as part of the real environ-ment Current systems are far from perfect and systemdesigners typically end up making a number of applica-tion-dependent trade offs We shall list and discuss thesetradeoffs in order to guide the choice of technology de-pending upon the type of application considered

Both systems optical and video have two imagesources the real world and the computer-generatedworld These image sources are to be merged Opticalsee-through HMDs take what might be called a lsquolsquomini-mally obtrusiversquorsquo approach that is they leave the view ofthe real world nearly intact and attempt to augment it by

merging a reflected image of the computer-generatedscene into the view of the real world Video see-throughHMDs are typically more obtrusive in the sense thatthey block out the real-world view in exchange for theability to merge the two views more convincingly Inrecent developments narrow fields of view in video see-through HMDs have replaced large field-of-viewHMDs thus reducing the area where the real world(captured through video) and the computer-generatedimages are merged into a small part of the visual sceneIn any case a fundamental consideration is whether theadditional features afforded by video see-throughHMDs justify the loss of the unobstructed real-worldview

Our experience indicates that there are many tradeoffsbetween optical and video see-through HMDs with re-spect to technological and human-factors issues that af-fect designing building and assessing these HMDs Thespecific issues are laid out in figure 11 While most ofthese issues could be discussed from both a technologi-cal and human-factors-standpoint (because the two areclosely interrelated in HMD systems) we have chosen toclassify each issue where it is most adequately addressedat this time given the present state of the technologyFor example delays in HMD systems are addressed un-der technology because technological improvements areactively being pursued to minimize delays Delays also

Figure 11 Outline of sections 31 and 32 of this paper


certainly have impact on various human-factor issues(such as the perceived location of objects in depth anduser acceptance) Therefore the multiple arrows shownin figure 11 indicate that the technological and human-factor-categories are highly interrelated

31 Technological Issues

The technological issues for HMDs include latencyof the system resolution and distortion of the real scenefield of view (FOV) eyepoint matching of the see-through device and engineering and cost factors Whilewe shall discuss properties of both optical and video see-through HMDs it must be noted that contrary to opti-cal see-through HMDs there are no commercially avail-able products for video see-through HMDs Thereforediscussions of such systems should be considered care-fully as findings may be particular to only a few currentsystems Nevertheless we shall provide as much insightas possible into what we have learned with such systemsas well

311 System Latency An essential componentof see-through HMDs is the capacity to properly registera userrsquos surroundings and the synthetic space A geomet-ric calibration between the tracking devices and theHMD optics must be performed The major impedi-ment to achieving registration is the gap in time re-ferred to as lag between the moment when the HMDposition is measured and the moment when the syn-thetic image for that position is fully rendered and pre-sented to the user

Lag is the largest source of registration error in mostcurrent HMD systems (Holloway 1995) This lag intypical systems is between 60 ms and 180 ms The headof a user can move during such a period of time and thediscrepancy in perceived scene and superimposed scenecan destroy the illusion of the synthetic objects beingfixed in the environment The synthetic objects canlsquolsquoswimrsquorsquo around significantly in such a way that they maynot even seem to be part of the real object to which theybelong For example in the case of ultrasound-guidedbiopsy the computer-generated tumor may appear to belocated outside the breast while tracking the head of the

user This swimming effect has been demonstrated andminimized by predicting HMD position instead of sim-ply measuring positions (Azuma amp Bishop 1994)

Current HMD systems are lag limited as a conse-quence of tracker lag the complexity of rendering anddisplaying the images Tracker lag is often not the limit-ing factor in performance If displaying the image is thelimiting factor novel display architectures supportingframeless rendering can help solve the problem (Bishopet al 1994) Frameless rendering is a procedure for con-tinuously updating a displayed image as informationbecomes available instead of updating entire frames at atime The tradeoffs between lag and image quality arecurrently being investigated (Scher-Zagier 1997) If weassume that we are limited by the speed of rendering animage eye-tracking capability may be useful to quicklyupdate information only around the gaze point of theuser (Thomas et al 1989 Rolland Yoshida et al1998 Vaissie amp Rolland 1999)

One of the major advantages of video see-throughHMDs is the potential capability of reducing the relativelatencies between the 2-D real and synthetic images as aconsequence of both types of images being digital (Ja-cobs et al 1997) Manipulation of the images in spaceand in time is applied to register them Three-dimen-sional registration is computationally intensive if at allrobust and challenging for interactive speed The spatialapproach to forcing registration in video see-throughsystems is to correct registration errors by imaging land-mark points in the real world and registering virtual ob-jects with respect to them (State et al 1996) One ap-proach to eliminating temporal delays between the realand computer-generated images in such a case is to cap-ture a video image and draw the graphics on top of thevideo image Then the buffer is swapped and the com-bined image is presented to the HMD user In such aconfiguration no delay apparently exists between thereal and computer-generated images If the actual la-tency of the computer-generated image is large with re-spect to the video image however it may cause sensoryconflicts between vision and proprioception because thevideo images no longer correspond to the real-worldscene Any manual interactions with real objects couldsuffer as a result


Another approach to minimizing delays in video see-through HMDs is to delay the video image until thecomputer-generated image is rendered Bajura and Neu-mann (1995) applied chroma keying for example todynamically image a pair of red LEDs placed on two realobjects (one stream) and then registered two virtual ob-jects with respect to them (second stream) By trackingmore landmarks better registration of real and virtualobjects may be achieved (Tomasi and Kanade 1991)The limitation of the approach taken is the attempt toregister 3-D scenes using 2-D constraints If the userrotates his head rapidly or if a real-world object movesthere may be no lsquolsquocorrectrsquorsquo transformation for the virtualscene image To align all the landmarks one must eitherallow errors in registration of some of the landmarks orperform a nonlinear warping of the virtual scene thatmay create undesirable distortions of the virtual objectsThe nontrivial solution to this problem is to increase thespeed of the system until scene changes between framesare small and can be approximated with simple 2-Dtransformations

In a similar vein it is also important to note that thevideo view of the real scene will normally have some lagdue to the time it takes to acquire and display the videoimages Thus the image in a video see-through HMDwill normally be slightly delayed with respect to the realworld even without adding delay to match the syntheticimages This delay may increase if an image-processingstep is applied to either enforce registration or performocclusion The key issue is whether the delay in the sys-tem is too great for the user to adapt to it (Held ampDurlach 1987)

Systems using optical see-through HMDs have nomeans of introducing artificial delays into the real sceneTherefore the system may need to be optimized for lowlatency perhaps less than 60 ms where predictive track-ing can be effective (Azuma amp Bishop 1994) For anyremaining lag the user may have to limit his actions toslow head motions Applications in which speed ofmovement can be readily controlled such as in theVRDA tool described earlier can benefit from opticalsee-through technology (Rolland amp Arthur 1997) Theadvantage of having no artificial delays is that real ob-

jects will always be where they are perceived to be andthis may be crucial for a broad range of applications

312 Real-Scene Resolution and Distortion Ifreal-scene resolution refers to the resolution of the real-scene object the best real-scene resolution that a see-through device can provide is that perceived with thenaked eye under unit magnification of the real sceneCertainly under microscopic observation as described byHill (Edwards et al 1995) the best scene resolutiongoes beyond that obtained with a naked eye It is alsoassumed that the see-through device has no image-pro-cessing capability

A resolution extremely close to that obtained with thenaked eye is easily achieved with a nonmicroscopic opti-cal see-through HMD because the optical interface tothe real world is simply a thin parallel plate (such as aglass plate) positioned between the eyes and the realscene Such an interface typically introduces only verysmall amounts of optical aberrations to the real sceneFor example for a real-point object seen through a 2mm planar parallel plate placed in front of a 4 mm diaeye pupil the diffusion spot due to spherical aberrationwould subtend a 2 10 2 7 arc-minute visual angle for apoint object located 500 mm away Spherical aberrationis one of the most common and simple aberrations inoptical systems that lead to blurring of the images Sucha degradation of image quality is negligible compared tothe ability of the human eye to resolve a visual angle of 1minute of arc Similarly planar plates introduce low dis-tortion of the real scene typically below 1 There is nodistortion only for the chief rays that pass the plate paral-lel to its normal1

In the case of a video see-through HMD real-sceneimages are digitized by miniature cameras (Edwards etal 1993) and converted into an analog signal that is fedto the HMD The images are then viewed through theHMD viewing optics that typically use an eyepiece de-sign The perceived resolution of the real scene can thusbe limited by the resolution of the video cameras or theHMD viewing optics Currently available miniature

1 A chief ray is defined as a ray that emanates from a point in theFOV and passes through the center of the pupils of the system Theexit pupil in an HMD is the entrance pupil of the human eye


video cameras typically have a resolution of 640 3 480which is also near the resolution limit of the miniaturedisplays currently used in HMDs2 Depending upon themagnification and the field of view of the viewing opticsvarious effective visual resolutions may be reachedWhile the miniature displays and the video cameras seemto currently limit the resolution of most systems suchperformance may improve with higher-resolution detec-tors and displays

In assessing video see-through systems one must dis-tinguish between narrow and wide FOV devices Large-FOV ( $ 50 deg) eyepiece designs are known to be ex-tremely limited in optical quality as a consequence offactors such as optical aberrations that accompany largeFOVs pixelization that may become more apparent un-der large magnification and the exit pupil size that mustaccommodate the size of the pupils of a personrsquos eyesThus even with higher-resolution cameras and displaysvideo see-through HMDs may remain limited in theirability to provide a real-scene view of high resolution ifconventional eyepiece designs continue to be used Inthe case of small to moderate FOV (10 deg to 20 deg)video see-through HMDs the resolution is still typicallymuch less than the resolving power of the human eye

A new technology referred to as tiling may overcomesome of the current limitations of conventional eyepiecedesign for large FOVs (Kaiser 1994) The idea is to usemultiple narrow-FOV eyepieces coupled with miniaturedisplays to completely cover (or tile) the userrsquos FOVBecause the individual eyepieces have a fairly narrowFOV higher resolution (nevertheless currently less thanthe human visual system) can be achieved One of thefew demonstrations of high-resolution large-FOV dis-plays is the tiled displays A challenge is the minimizationof seams in assembling the tiles and the rendering ofmultiple images at interactive speed The tiled displayscertainly bring new practical and computational chal-lenges that need to be confronted If a see-through ca-pability is desired (for example to display virtual furni-ture in an empty room) it is currently unclear whether

the technical problems associated with providing overlaycan be solved

Theoretically distortion is not a problem in video see-through systems because the cameras can be designed tocompensate for the distortion of the optical viewer asdemonstrated by Edwards et al (1993) However if thegoal is to merge real and virtual information as in ultra-sound echography having a warped real scene signifi-cantly increases the complexity of the synthetic-imagegeneration (State et al 1994) Real-time video correc-tion can be used at the expense of an additional delay inthe image-generation sequence An alternative is to uselow-distortion video cameras at the expense of a nar-rower FOV merge unprocessed real scenes with virtualscenes and warp the merged images Warping can bedone using (for example) real-time texture mapping tocompensate for the distortion of the HMD viewing op-tics as a last step (Rolland amp Hopkins 1993 Watson ampHodges 1995)

The need for high real-scene resolution is highly taskdependent Demanding tasks such as surgery or engi-neering training for example may not be able to toler-ate much loss in real-scene resolution Because the large-FOV video see-through systems that we haveexperienced are seriously limited in terms of resolutionnarrow-FOV video see-through HMDs are currentlypreferred Independently of resolution an additionalcritical issue in aiming towards narrow-FOV video see-through HMDs is the need to match the viewpoint ofthe video cameras with the viewpoint of the user Match-ing is challenging with large-FOV systems Also meth-ods for matching video and real scenes for large-FOVtiled displays must be developed At this time consider-ing the growing availability of high-resolution flat-paneldisplays we foresee that the resolution of see-thoughHMDs could gradually increase for both small- andlarge-FOV systems The development and marketing ofminiature high-resolution technology must be under-taken to achieve resolutions that match that of humanvision

313 Field of View A generally challenging issueof HMDs is providing the user with an adequate FOVfor a given application For most applications increasing

2 The number of physical elements is typically 640 3 480 One canuse signal processing to interpolate between lines to get higher resolu-tions


the binocular FOV means that fewer head movementsare required to perceive an equivalently large scene Webelieve that a large FOV is especially important for tasksthat require grabbing and moving objects and that itprovides increased situation awareness when comparedto narrow-FOV devices (Slater amp Wilbur 1997) Thesituation with see-through devices is somewhat differentfrom that of fully opaque HMDs in that the aim of usingthe technology is different from that of immersing theuser in a virtual environment

3131 Overlay and Peripheral FOV The termoverlay FOV is defined as the region of the FOV wheregraphical information and real information are superim-posed The peripheral FOV is the real-world FOV be-yond the overlay FOV For immersive opaque HMDsno such distinction is made one refers simply to theFOV It is important to note that the overlay FOV mayneed to be narrow only for certain augmented-realityapplications For example in a visualization tool such asthe VRDA tool only the knee-joint region is needed inthe overlay FOV In the case of video HMD-guidedbreast biopsy the overlay FOV could be as narrow as thesynthesized tumor The real scene need not necessarilybe synthesized The available peripheral FOV howeveris critical for situation awareness and is most often re-quired for various applications whether it is provided aspart of the overlay or around the overlay If providedaround the overlay the transition from real to virtualimagery must be made as seamless as possible This is aninvestigation that has not yet been addressed in videosee-through HMDs

Optical see-through HMDs typically provide from 20deg to 60 deg overlay FOV via the half-transparentmirrors placed in front of the eyes a characteristic thatmay seem somewhat limited but promising for a varietyof medical applications whose working visualization dis-tance is within arm reach Larger FOVs have been ob-tained up to 825 3 67 deg at the expense of reducedbrightness increased complexity and massive expensivetechnology (Welch amp Shenker 1984) Such FOVs mayhave been required for performing navigation tasks inreal and virtual environments but are likely not requiredin most augmented-reality applications Optical see-

through HMDs however whether or not they have alarge overlay FOV have been typically designed openenough that users can use their peripheral vision aroundthe device thus increasing the total real-world FOV toclosely match onersquos natural FOV An annulus of obstruc-tion usually results from the mounts of the thin see-through mirror similar to the way that our vision may bepartially occluded by a frame when wearing eyeglasses

In the design of video see-through HMDs a difficultengineering task is matching the frustum of the eye withthat of the camera (as we shall discuss in section 314)While such matching is not so critical for far-field view-ing it is important for near-field visualization as in vari-ous medical visualizations This difficult matching prob-lem has lead to the consideration of narrower-FOVsystems A compact 40 3 30 deg FOV design de-signed for optical see-through HMD but adaptable tovideo see-through was proposed by Manhart Malcolmamp Frazee (1993) Video see-through HMDs on theother hand can provide (in terms of a see-throughFOV) the FOV displayed with the opaque type viewingoptics that typically range from 20 deg to 90 deg Insuch systems where the peripheral FOV of the user isoccluded the effective real-world FOV is often smallerthan in optical see-through systems When using a videosee-through HMD in a hand-eye coordination task wefound in a recent human-factor study that users neededto perform larger head movements to scan an active fieldof vision than when performing the task with the un-aided eye (Biocca amp Rolland 1998) We predict that theneed to make larger head movements would not arise asmuch with see-through HMDs with equivalent overlayFOVs but larger peripheral FOVs because users are pro-vided with increased peripheral vision and thus addi-tional information to more naturally perform the task

3132 Increasing Peripheral FOV in VideoSee-Through HMDs An increase in peripheral FOV invideo see-through systems can be accomplished in twoways in a folded optical design as used for optical see-through HMDs however with an opaque mirror insteadof a half-transparent mirror or in a nonfolded designbut with nonenclosed mounts The latter calls for inno-vative optomechanical design because heavier optics


have to be supported than in either optical or foldedvideo see-through Folded systems require only a thinmirror in front of the eyes and the heavier optical com-ponents are placed around the head However thetradeoff with folded systems is a significant reduction inthe overlay FOV

3133 Tradeoff Resolution and FOV While theresolution of a display in an HMD is defined in thegraphics community by the number of pixels the rel-evant measure of resolution is the number of pixels perangular FOV which is referred to as angular resolutionIndeed what is of importance for usability is the angularsubtends of a pixel at the eye of the HMD user Mostcurrent high-resolution HMDs achieve higher resolu-tion at the expense of a reduced FOV That is they usethe same miniature high-resolution CRTs but with op-tics of less magnification in order to achieve higher an-gular resolution This results in a FOV that is often nar-row The approach that employs large high-resolutiondisplays or light valves and transports the high-resolu-tion images to the eyes by imaging optics coupled to abundle of optical fibers achieves high resolution at fairlylarge FOVs (Thomas et al 1989) The current pro-posed solutions that improve resolution without tradingFOV are either tiling techniques high-resolution insetdisplays (Fernie 1995 Rolland Yoshida et al 1998)or projection HMDs (Hua et al 2000)

Projective HMDs differ from conventional HMDs inthat projection optics are used instead of eyepiece opticsto project real images of miniature displays in the envi-ronment A screen placed in the environment reflects theimages back to the eyes of the user Projective HMDshave been designed and demonstrated for example byKijima and Ojika (1997) and Parsons and Rolland(1998) Kijima used a conventional projection screen inhis prototype Parsons and Rolland developed a first-prototype projection HMD system to demonstrate thatan undistorted virtual 3-D image could be renderedwhen projecting a stereo pair of images on a bent sheetof microretroreflector cubes The first proof-of-conceptsystem is shown in figure 12 A comprehensive investiga-tion of the optical characteristics of projective HMDs isgiven by Hua et al (2000) We are also developing the

next-generation prototypes of the technology using cus-tom-made miniature lightweight optics The systempresents various advantages over conventional HMDsincluding distortion-free images occluded virtual ob-jects from real-objects interposition no image cross-talks for multiuser participants in the virtual world andthe potential for a wide FOV (up to 120 deg)

314 Viewpoint Matching In video see-through HMDs the camera viewpoint (that is the en-trance pupil) must be matched to the viewpoint of theobserver (the entrance pupil of the eye) The viewpointof a camera or eye is equivalent to the center of projec-tion used in the computer graphics model that computesthe stereo images and is taken here to be the center ofthe entrance pupil of the eye or camera (Vaissie ampRolland 2000) In earlier video see-through designsEdwards et al (1993) investigated ways to mount thecameras to minimize errors in viewpoint matching Theerror minimization versus exact matching was a conse-quence of working with wide-FOV systems If the view-points of the cameras do not match the viewpoints of theeyes the user experiences a spatial shift in the perceivedscene that may lead to perceptual anomalies (as further

Figure 12 Proof of concept prototype of a

projection head-mounted display with

microreector sheeting (1998)


discussed under human-factors issues (Biocca ampRolland 1998) Error analysis should then be con-ducted in such a case to match the need of the applica-tion

For cases in which the FOV is small (less than approxi-mately 20 deg) exact matching in viewpoints is pos-sible Because the cameras cannot be physically placed atthe actual eyepoints mirrors can be used to fold the op-tical path (much like a periscope) to make the camerasrsquoviewpoints correspond to the real eyepoints as shown infigure 13 (Edwards et al 1993) While such geometrysolves the problem of the shift in viewpoint it increasesthe length of the optical path which reduces the field ofview for the same reason that optical see-throughHMDs tend to have smaller fields of view Thus videosee-through HMDs must either trade their large FOVsfor correct real-world viewpoints or require the user toadapt to the shifted viewpoints as further discussed insection 323

Finally correctly mounting the video cameras in avideo see-through HMD requires that the HMD has aninterpupillary distance (IPD) adjustment Given the IPDof a user the lateral separation of the video cameras mustthen be adjusted to that value in order for the views ob-tained by the video cameras to match those that wouldhave been obtained with naked eyes If one were to ac-count for eye movements in video see-through HMDsthe level of complexity in slaving the camera viewpointto the user viewpoint would be highly increased To ourknowledge such complexity has not yet been consid-ered

315 Engineering and Cost Factors HMDdesigns often suffer from fairly low resolution limitedFOV poor ergonomic designs and excessive weight Agood ergonomic design requires an HMD whose weightis similar to a pair of eyeglasses or which folds aroundthe userrsquos head so the devicersquos center of gravity falls nearthe center of rotation of the head (Rolland 1994) Thegoal here is maximum comfort and usability Reasonablylightweight HMD designs currently suffer narrowFOVs on the order of 20 deg To our knowledge atpresent no large-FOV stereo see-through HMDs of anytype are comparable in weight to a pair of eyeglassesRolland predicts that it could be achieved with someemerging technology of projection HMDs (RollandParsons et al 1998) However it must be noted thatsuch technology may not be well suited to all visualiza-tion schemes as it requires a projection screen some-where in front of the user that is not necessarily attachedto the userrsquos head

With optical see-through HMDs the folding can beaccomplished with either an on-axis or an off-axis de-sign Off-axis designs are more elegant and also far moreattractive because they elimate the ghost images thatcurrently plague users of on-axis HMDs (Rolland2000) Off-axis designs are not commercially availablebecause very few prototypes have been built (and thosethat have been built are classified) (Shenker 1998)Moreover off-axis systems are difficult to design and arethus expensive to build as a result of off-axis components(Shenker 1994) A nonclassified off-axis design hasbeen designed by Rolland (1994 2000) Several factors

Figure 13 A 10 degree FOV video see-through

HMD Dglasses developed at UNC-CH Lipstick

cameras and a double fold mirror arrangement was

used to match the viewpoints of the camera and user

(1997)


(including cost) have also hindered the construction of afirst prototype as well New generations of computer-controlled fabrication and testing are expected to changethis trend

Since their beginning high-resolution HMDs havebeen CRT based Early systems were even monochromebut color CRTs using color wheels or frame-sequentialcolor have been fabricated and incorporated into HMDs(Allen 1993) Five years ago we may have thought thattoday high-resolution color flat-panel displays wouldbe the first choice for HMDs While this is slowly hap-pening miniature CRTs are not fully obsolete The cur-rent optimism however is prompted by new technolo-gies such as reflective LCDs microelectromechanicalsystems (MEMS)-based displays laser-based displaysand nanotechnology-based displays

32 Human-Factor and PerceptualIssues

Assuming that many of the technological chal-lenges described have been addressed and high-perfor-mance HMDs can be built a key human-factor issue forsee-through HMDs is that of user acceptance and safetywhich will be discussed first We shall then discuss thetechnicalities of perception in such displays The ulti-mate see-through display is one that provides quantita-tive and qualitative visual representations of scenes thatconform to a predictive model (for example conform tothat given by the real world if that is the intention) Is-sues include the accuracy and precision of the renderedand perceived locations of objects in depth the accuracyand precision of the rendered and perceived sizes of realand virtual objects in a scene and the need of an unob-structed peripheral FOV (which is important for manytasks that require situation awareness and the simple ma-nipulation of objects and accessories

321 User Acceptance and Safety A fair ques-tion for either type of technology is lsquolsquowill anyone actuallywear one of these devices for extended periodsrsquorsquo Theanswer will doubtless be specific to the application andthe technology included but it will probably centerupon whether the advanced capabilities afforded by the

technology offset the problems induced by the encum-brance and sensory conflicts that are associated with it

In particular one of us thinks that video see-throughHMDs may be met with resistance in the workplace be-cause they remove the direct real-world view in order toaugment it This issue of trust may be difficult to over-come for some users If wide-angle FOV video see-through HMDs are used this problem is exacerbated insafety-critical applications A key difference in such appli-cations may turn out to be the failure mode of eachtechnology A technology failure in the case of opticalsee-through HMDs may leave the subject without anycomputer-generated images but still with the real-worldview In the case of video see-through it may leave theuser with the complete suppression of the real-worldview as well as the computer-generated view

However it may be that the issue has been greatlylessened because the video view occupies such a smallfraction (approximately 10 deg visual angle) of thescene in recent developments of the technology It isespecially true of flip-up and flip-down devices such asthat developed at UNC-CH and shown in figure 13

Image quality and its tradeoffs are definitely criticalissues related to user acceptance for all types of technol-ogy In a personal communication Martin Shenker asenior optical engineer with more than twenty years ofexperience designing HMDs pointed out that there arecurrently no standards of image quality and technologyspecifications for the design calibration and mainte-nance of HMDs This is a current concern at a timewhen the technology may be adopted in various medicalvisualizations

322 Perceived Depth 3221 Occlusion Theability to perform occlusion in see-through HMDs is animportant issue of comparison between optical andvideo see-through HMDs One of the most importantdifferences between these two technologies is how theyhandle the depth cue known as occlusion (or interposi-tion) In real life an opaque object can block the view ofanother object so that part or all of it is not visibleWhile there is no problem in making computer-gener-ated objects occlude each other in either system it isconsiderably more difficult to make real objects occlude


virtual objects (and vice versa) unless the real world foran application is predefined and has been modeled in thecomputer Even then one would need to know the exactlocation of a user with respect to that real environmentThis is not the case in most augmented-reality applica-tions in which the real world is constantly changing andon-the-fly acquisition is all the information one will everhave of the real world Occlusion is a strong monocularcue to depth perception and may be required in certainapplications (Cutting amp Vishton 1995)

In both systems computing occlusion between thereal and virtual scenes requires a depth map of bothscenes A depth map of the virtual scene is usually avail-able (for z-buffered image generators) but a depth mapof the real scene is a much more difficult problem Whileone could create a depth map in advance from a staticreal environment many applications require on-the-flyimage acquisition of the real scene Assuming the systemhas a depth map of the real environment video see-through HMDs are perfectly positioned to take advan-tage of this information They can on a pixel-by-pixelbasis selectively block the view of either scene or evenblend them to minimize edge artifacts One of the chiefadvantages of video see-through HMDs is that theyhandle this problem so well

The situation for optical see-through HMDs can bemore complex Existing optical see-through HMDsblend the two images with beam splitters which blendthe real and virtual images uniformly throughout theFOV Normally the only control the designer has is theamount of reflectance versus transmittance of the beamsplitter which can be chosen to match the brightness ofthe displays with the expected light levels in the real-world environment If the system has a model of the realenvironment it is possible to have real objects occludevirtual ones by simply not drawing the occluded parts ofthe virtual objects The only light will then be from thereal objects giving the illusion that they are occludingthe virtual ones Such an effect requires a darkenedroom with light directed where it is needed This tech-nique has been used by CAE Electronics in their flightsimulator When the pilots look out the window theysee computer-generated objects If they look inside the

cockpit however the appropriate pixels of the com-puter-generated image are masked so that they can seethe real instruments The room is kept fairly dark so thatthis technique will work (Barrette 1992) David Mizell(from Boeing Seattle) and Tom Caudell (University ofNew Mexico) are also using this technique they refer toit as lsquolsquofused realityrsquorsquo (Mizell 1998)

While optical see-through HMDs can allow real ob-jects to occlude virtual objects the reverse is even morechallenging because normal beam splitters have no wayof selectively blocking out the real environment Thisproblem has at least two possible partial solutions Thefirst solution is to spatially control the light levels in thereal environment and to use displays that are brightenough so that the virtual objects mask the real ones byreason of contrast (This approach is used in the flightsimulator just mentioned for creating the virtual instru-ments) This may be a solution for a few applications Apossible second solution would be to locally attenuatethe real-world view by using an addressable filter deviceplaced on the see-through mirror It is possible to gener-ate partial occlusion in this manner because the effectivebeam of light entering the eye from some point in thescene covers only a small area of the beam splitter theeye pupil being typically 2mm to 4mm in photopic vi-sion A problem with this approach is that the user doesnot focus on the beam splitter but rather somewhere inthe scene A point in the scene maps to a disk on thebeam splitter and various points in the scene map tooverlapping disks on the beam splitter Thus any block-ing done at the beam splitter may occlude more of thescene than expected which might lead to odd visual ef-fects A final possibility is that some applications maywork acceptably without properly rendered occlusioncues That is in some cases the user may be able to useother depth cues such as head-motion parallax to re-solve the ambiguity caused by the lack of occlusion cues

3222 Rendered Locations of Objects in Depth Weshall distinguish between errors in the rendered and per-ceived locations of objects in depth The former yieldsthe latter One can conceive however that errors in theperceived location of objects in depth can also occur


even in the absence of errors in rendered depths as a re-sult of an incorrect computational model for stereo pairgeneration or a suboptimal presentation of the stereoimages This is true both for optical and video see-through HMDs Indeed if the technology is adequateto support a computational model and the model ac-counts for required technology and corresponding pa-rameters the rendered locations of objects in depthmdashaswell as the resulting perceived locations of objects indepthmdashwill follow expectations Vaissie recently showedsome limitations of the choice of a static eyepoint incomputational models for stereo pair generation for vir-tual environments that yield errors in rendered and thusperceived location of objects in depths (Vaissie andRolland 2000) The ultimate goal is to derive a compu-tational model and develop the required technology thatyield the desired perceived location of objects in depthErrors in rendered depth typically result from inaccuratedisplay calibration and parameter determination such asthe FOV the frame-buffer overscan the eyepointsrsquo loca-tions conflicting or noncompatible cues to depth andremaining optical aberrations including residual opticaldistortions

3223 FOV and Frame-Buffer Overscan Inaccu-racies of a few degrees in FOV are easily made if no cali-bration is conducted Such inaccuracies can lead to sig-nificant errors in rendered depths depending on theimaging geometry For some medical and computer-guided surgery applications for example errors of sev-eral millimeters are likely to be unacceptable The FOVand the overscan of the frame buffer that must be mea-sured and accounted for to yield accurate rendereddepths are critical parameters for stereo pair generationin HMDs (Rolland et al 1995) These parameters mustbe set correctly regardless of whether the technology isoptical or video see-through

3224 Specification of Eyepoint Location Thespecification of the locations of the userrsquos eyepoints(which are used to render the stereo images from thecorrect viewpoints) must be specified for accurate ren-dered depth This applies to both optical and video see-through HMDs In addition for video see-through

HMDs the real-scene video images must be acquiredfrom the correct viewpoint (Biocca amp Rolland 1998)

For the computer graphics-generation componentthree choices of eyepoint locations within the human eyehave been proposed the nodal point of the eye3 (Robi-nett amp Rolland 1992 Deering 1992) the entrancepupil of the eye (Rolland 1994 Rolland et al 1995)and the center of rotation of the eye (Holloway 1995)Rolland (1995) discusses that the choice of the nodalpoint would in fact yield errors in rendered depth in allcases whether the eyes are tracked or not For a devicewith eye-tracking capability the entrance pupil of the eyeshould be taken as the eyepoint If eye movements areignored meaning that the computer-graphics eyepointsare fixed then it was proposed that it is best to select thecenter of rotation of the eye as the eyepoint (Fry 1969Holloway 1995) An in-depth analysis of this issue re-veals that while the center of rotation yields higher accu-racy in position the center of the entrance pupil yields infact higher angular accuracy (Vaissie amp Rolland 2000)Therefore depending on the task involved and whetherangular accuracy or position accuracy is most importantthe centers of rotation or the centers of the entrance pu-pil may be selected as best eyepoints location in HMDs

3225 Residual Optical Distortions Optical dis-tortion is one of the few optical aberrations that do notaffect image sharpness rather it introduces warping ofthe image It occurs only for optics that include lenses Ifthe optics include only plane mirrors there are no dis-tortions (Peuchot 1994) Warping of the images leadsto errors in rendered depths Distortion results from thelocations of the userrsquos pupils away from the nodal pointsof the optics Moreover it varies as a function of wherethe user looks through the optics However if the opticsare well calibrated to account for the userrsquos IPD distor-tion will be fairly constant for typical eye movementsbehind the optics Prewarping of the computer-gener-ated image can thus be conducted to compensate for the

3 Nodal points are conjugate points in an optical system that satisfyan angular magnification of 1 Two points are considered to conjugateof each other if they are images of each other


optical residual distortions (Robinett amp Rolland 1992Rolland amp Hopkins 1993 Watson amp Hodges 1995)

3226 Perceived Location of Objects in DepthOnce depths are accurately rendered according to agiven computational model and the stereo images arepresented according to the computational model theperceived location and size of objects in depth becomean important issue in the assessment of the technologyand the model Accuracy and precision can be definedonly statistically Given an ensemble of measured per-ceived location of objects in depths the depth perceptwill be accurate if objects appear in average at the loca-tion predicted by the computational model The per-ceived location of objects in depth will be precise if ob-jects appear within a small spatial zone around thataverage location We shall distinguish between overlap-ping and nonoverlapping objects

In the case of nonoverlapping objects one may resortto depth cues other than occlusion These include famil-iar sizes stereopsis perspective texture and motionparallax A psychophysical investigation of the perceivedlocation of objects in depth in an optical see-throughHMD using stereopsis and perspective as the visual cuesto depth is given in Rolland et al (1995) and Rollandet al (1997) The HMD shown in figure 14 is mountedon a bench for calibration purpose and flexibility in vari-ous parameter settings

In a first investigation a systematic shift of 50 mm inthe perceived location of objects in depth versus pre-dicted values was found in this first set of study (Rollandet al 1995) Moreover the precision of the measuresvaried significantly across subjects As we learn moreabout the interface optics and computational modelused in the generation of the stereo image pairs and im-prove on the technology we have demonstrated errors

Figure 14 (a) Bench prototype head-mounted display with head-motion

parallax developed in the VGILab at UCF (1997) (b) Schematic of the optical

imaging from a top view of the setup


on the order of 2 mm The technology is now ready todeploy for extensive testing in specific applications andthe VRDA tool is one of the applications we are cur-rently pursuing

Studies of the perceived location of objects in depthfor overlapping objects in an optical see-through HMDhave been conducted by Ellis and Buchler (1994) Theyshowed that the perceived location of virtual objects canbe affected by the presence of a nearby opaque physicalobject When a physical object was positioned in front of(or at) the initial perceived location of a 3-D virtual ob-ject the virtual object appeared to move closer to theobserver In the case in which the opaque physical objectwas positioned substantially in front of the virtual objecthuman subjects often perceived the opaque object to betransparent In current investigations with the VRDAtool the opaque leg model appears transparent when avirtual knee model is projected on the leg as seen in fig-ure 4 The virtual anatomy subjectively appears to beinside the leg model (Baillot 1999 Outters et al 1999Baillot et al 2000)

323 Adaptation When a system does not offerwhat the user ultimately wants two paths may be takenimproving on the current technology or first studyingthe ability of the human system to adapt to an imperfecttechnological unit and then developing adaptation train-ing when appropriate This is possible because of theastonishing ability of the human visual and propriocep-tive systems to adapt to new environments as has beenshown in studies on adaptation (Rock 1966 for ex-ample)

Biocca and Rolland (1998) conducted a study of ad-aptation to visual displacement using a large-FOV videosee-through HMD Users see the real world throughtwo cameras that are located 62 mm higher than and165 mm forward from their natural eyepoints as shownin figure 2 Subjects showed evidence of perceptual ad-aptation to sensory disarrangement during the course ofthe study This revealed itself as improvement in perfor-mance over time while wearing the see-through HMDand as negative aftereffects once they removed it Moreprecisely the negative aftereffect manifested itself clearlyas a large overshoot in a depth-pointing task as well as

an upward translation in a lateral pointing task afterwearing the HMD Moreover some participants experi-enced some early signs of cybersickness

The presence of negative aftereffects has some poten-tially disturbing practical implications for the diffusion oflarge-FOV video see-through HMDs (Kennedy amp Stan-ney 1997) Some of the intended earlier users of theseHMDs are surgeons and other individuals in the medicalprofession Hand-eye sensory recalibration for highlyskilled users (such as surgeons) could have potentiallydisturbing consequences if the surgeon were to entersurgery within some period after using an HMD It is anempirical question how long the negative aftereffectsmight persist and whether a program of gradual adapta-tion (Welch 1994) or dual adaptation (Welch 1993)might minimize the effect altogether In any case anyshift in the camera eyepoints need to be minimized asmuch as possible to facilitate the adaptation process thatis taking place As we learn more about these issues wewill build devices with less error and more similarity be-tween using these systems and a pair of eyeglasses (sothat adaptation takes less time and aftereffects decreaseas well)

A remaining issue is the conflict between accommoda-tion and convergence in such displays The issue can besolved at some cost (Rolland et al 2000) For lower-end systems a question to investigate is how users adaptto various settings of the technology For high-end sys-tems much research is still needed to understand theimportance of perceptual conflicts and how to best mini-mize them

324 Peripheral FOV Given that peripheralvision can be provided in both optical and video see-through systems the next question is whether it is usedeffectively for both systems In optical see-throughthere is almost no transition or discrepancy between thereal scene captured by the see-through device and theperipheral vision seen on the side of the device

For video see-through the peripheral FOV has beenprovided by letting the user see around the device aswith optical see-through However it remains to be seenwhether the difference in presentation of the superim-posed real scene and the peripheral real scene will cause


discomfort or provide conflicting cues to the user Theissue is that the virtual displays call for a different accom-modation for the user than the real scene in variouscases

325 Depth of Field One important property ofoptical systems including the visual system is depth offield (Depth of field refers to the range of distancesfrom the detector (such as the eye) in which an objectappears to be in focus without the need for a change inthe optics focus (such as eye accommodation) For thehuman visual system example if an object is accuratelyfocused monocularly other objects somewhat nearer andfarther away are also seen clearly without any change inaccommodation Still nearer or farther objects areblurred Depth of field reduces the necessity for preciseaccommodation and is markedly influenced by the diam-eter of the pupil The larger the pupil the smaller thedepth of field For a 2 mm and 4 mm pupil the depthsof field are 6 006 and 6 003 diopters respectively Fora 4 mm pupil for example such a depth of field trans-lates as a clear focus from 094 m to 106 m for an object1 m away and from 11 m to 33 m for an object 17 maway (Campbell 1957 Moses 1970) An importantpoint is that accommodation plays an important roleonly at close working distances where depth of field isnarrow

With video see-through systems the miniature cam-eras that acquire the real-scene images must provide adepth of field equivalent to the required working dis-tance for a task For a large range of working distancesthe camera may need to be focused at the middle work-ing distance For closer distances the small depth of fieldmay require an autofocus instead of a fixed-focus cam-era

With optical see-through systems the available depthof field for the real scene is essentially that of the humanvisual system but for a larger pupil than would be acces-sible with unaided eyes This can be explained by thebrightness attenuation of the real scene by the half-trans-parent mirror As a result the pupils are dilated (we as-sume here that the real and virtual scenes are matched inbrightness) Therefore the effective depth of field is

slightly less than with unaided eyes This is a problemonly if the user is working with nearby objects and thevirtual images are focused outside of the depth of fieldthat is required for nearby objects For the virtual imagesand no autofocus capability for the 2-D virtual imagesthe depth of field is imposed by the human visual systemaround the location of the displayed virtual images

When the retinal images are not sharp following somediscrepancy in accommodation the visual system is con-stantly processing somewhat blurred images and tendsto tolerate blur up to the point at which essential detail isobscured This tolerance for blur considerably extendsthe apparent depth of field so that the eye may be asmuch as 6 025 diopters out of focus without stimulat-ing accommodative change (Moses 1970)

326 Qualitative Aspects The representationof virtual objects and in some cases of real objects isaltered by see-through devices Aspects of perceptualrepresentation include the shape of objects their colorbrightness contrast shading texture and level of detailIn the case of optical see-through HMDs folding theoptical path by using a half-transparent mirror is neces-sary because it is the only configuration that leaves thereal scene almost unaltered A thin folding mirror willintroduce a small apparent shift in depth of real objectsprecisely equal to e(n 2 1)n where e is the thickness ofthe plate and n is its index of refraction This is in addi-tion to a small amount of distortion ( 1) of the sceneat the edges of a 60 deg FOV Consequently real ob-jects are seen basically unaltered

Virtual objects on the other hand are formed fromthe fusion of stereo images formed through magnifyingoptics Each optical virtual image formed of the displayassociated with each eye is typically optically aberratedFor large-FOV optics such as HMDs astigmatism andchromatic aberrations are often the limiting factors Cus-tom-designed HMD optics can be analyzed from a visualperformance point of view (Shenker 1994 Rolland2000) Such analysis allows the prediction of the ex-pected visual performance of HMD users

It must be noted that real and virtual objects in suchsystems may be seen sharply by accommodating in dif-


ferent planes under most visualization settings Thisyields conflicts in accommodation for real and virtualimagery For applications in which the virtual objects arepresented in a small working volume around some meandisplay distance (such as arm-length visualization) the2-D optical images of the miniature displays can be lo-cated at that same distance to minimize conflicts in ac-commodation and convergence between real and virtualobjects Another approach to minimizing conflicts inaccommodation and convergence is multifocal planestechnology as described in Rolland et al 2000)

Beside brightness attenuation and distortion otheraspects of object representation are altered in video see-through HMDs The authorsrsquo experience with at leastone system is that the color and brightness of real ob-jects are altered along with the loss in texture and levelsof detail due to the limited resolution of the miniaturevideo cameras and the wide-angle optical viewer Thisalteration includes spatial luminance and color resolu-tion This is perhaps resolvable with improved technol-ogy but it currently limits the ability of the HMD userto perceive real objects as they would appear with un-aided eyes In wide-FOV video see-through HMDsboth real and virtual objects call for the same accommo-dation however conflicts of accommodation and con-vergence are also present As with optical see-throughHMDs these conflicts can be minimized if objects areperceived at a relatively constant depth near the plane ofthe optical images In narrow-FOV systems in which thereal scene is seen in large part outside the overlay imag-ery conflicts in accommodation can also result betweenthe real and computer-generated scene

For both technologies a solution to these variousconflicts in accommodation may be to allow autofocusof the 2-D virtual images as a function of the location ofthe user gaze point in the virtual environment or toimplement multifocal planes (Rolland et al 2000)Given eye-tracking capability autofocus could be pro-vided because small displacements of the miniature dis-play near the focal plane of the optics would yield largeaxial displacements of the 2-D virtual images in the pro-jected virtual space The 2-D virtual images would movein depth according to the user gaze point Multifocal

planes also allow autofocusing but with no need for eyetracking

4 Conclusion

We have discussed issues involving optical andvideo see-through HMDs The most important issuesare system latency occlusion the fidelity of the real-world view and user acceptance Optical see-throughsystems offer an essentially unhindered view of the realenvironment they also provide an instantaneous real-world view that assures that visual and proprioceptioninformation is synchronized Video systems forfeit theunhindered view in return for improved ability to seereal and synthetic imagery simultaneously

Some of us working with optical see-through devicesstrongly feel that providing the real scene through opti-cal means is important for applications such as medicalvisualization in which human lives are at stake Othersworking with video see-through devices feel that aflip-up view is adequate for the safety of the patientAlso how to render occlusion of the real scene at givenspatial locations may be important Video see-throughsystems can also guarantee registration of the real andvirtual scenes at the expense of a mismatch between vi-sion and proprioception This may or may not be per-ceived as a penalty if the human observer is able to adaptto such a mismatch Hybrid solutions such as that de-veloped by Peuchot (1994) including optical see-through technology for visualization and video technol-ogy for tracking objects in the real environment mayplay a key role in future developments of technology for3-D medical visualization

Clearly there is no lsquolsquorightrsquorsquo system for all applicationsEach of the tradeoffs discussed in this paper must be ex-amined with respect to specific applications and availabletechnology to determine which type of system is mostappropriate Furthermore additional HMD featuressuch as multiplane focusing and eye tracking are cur-rently investigated at various research and developmentsites and may provide solutions to current perceptual


conflicts in HMDs A shared concern among scientistsdeveloping further technology is the lack of standardsnot only in the design but also most importantly in thecalibration and maintenance of HMD systems

Acknowledgments

This review was expanded from an earlier paper in a SPIE pro-ceeding by Rolland Holloway and Fuchs (1994) and the au-thors would like to thank Rich Holloway for his earlier contri-bution to this work We thank Myron Krueger from ArtificialReality Corp for stimulating discussions on various aspects ofthe technology as well as Martin Shenker from MSOD andBrian Welch from CAE Electronics for discussions on currentoptical technologies Finally we thank Bernard Peuchot Derek

Hill and Andrei State for providing information about theirresearch that has significantly contributed to the improvementof this paper We deeply thank our various sponsors not onlyfor their financial support that has greatly facilitated our re-search in see-through devices but also for the stimulating dis-cussions they have provided over the years Contracts andgrants include ARPA DABT 63-93-C-0048 NSF CooperativeAgreement ASC-8920219 Science and Technology Center forComputer Graphics and Scientific Visualization ONR

N00014-86-K-0680 ONR N00014-94-1-0503 ONRN000149710654 NIH 5-R24-RR-02170 NIH1-R29LM06322-O1A1 and DAAH04-96-C-0086

References

Allen D (1993) A high resolution field sequential display forhead-mounted applications Proc IEEE virtual reality an-nual international symposium (VRAISrsquo93) (pp 364ndash370)

Azuma R amp Bishop G (1994) Improving static and dy-namic registration in an optical see-through HMD Com-puter Graphics Proceedings of SIGGRAPH rsquo94 (pp 197ndash

204)Bajura M Fuchs H amp Ohbuchi R (1992) Merging virtual

objects with the real world Computer Graphics 26 203ndash210

Bajura M amp Newmann H (1995) Dynamic registration cor-rection in video-based augmented reality systems IEEEComputer Graphics and Applications 15(5) 52ndash60

Baillot Y amp Rolland J P (1998) Modeling of a knee jointfor the VRDA tool Proc of Medicine Meets Virtual Reality(pp 366ndash367)

Baillot Y Rolland J P amp Wright D (1999) Kinematicmodeling of knee-joint motion for a virtual reality tool InProc of Medicine Meets Virtual Reality (pp 30ndash35)

Baillot Y Rolland J P Lin K amp Wright D L (2000) Au-tomatic modeling of knee-joint motion for the Virtual Real-ity Dynamic Anatomy (VRDA) Tool Presence Teleoperatorsand Virtual Environments 9(3) 223ndash235

Baillot Y (1999) First implementation of the Virtual RealityDynamic Anatomy (VRDA) tool Unpublished masterrsquos the-

sis University of Central FloridaBarrette R E (1992) Wide field of view full-color high-

resolution helmet-mounted display Proc of SID rsquo92 Sympo-sium (pp 69ndash72)

Biocca F A amp Rolland J P (1998) Virtual eyes can rear-range your body Adaptation to virtual-eye location in see-thru head-mounted displays Presence Teleoperators and Vir-tual Environments 7(3) 262ndash277

Bishop G Fuchs H McMillan L amp Scher-Zagier E J(1994) Frameless rendering Double buffering consideredharmful Proc of SIGGRAPHrsquo94 (pp 175ndash176)

Brooks F P (1992) Walkthrough project Final technical re-port to National Science Foundation Computer and Infor-mation Science and Engineering Technical Report TR92-026 University of North Carolina at Chapel Hill

Buchroeder R A Seeley G W amp Vukobratovich D(1981) Design of a catadioptric VCASS helmet-mounteddisplay Optical Sciences Center University of Arizona un-der contract to US Air Force Armstrong Aerospace Medi-cal Research Laboratory Wright-Patterson Air Force BaseDayton Ohio AFAMRL-TR-81-133

Campbell F W (1957) The depth of field of the human eyeOptica Acta 4 157ndash164

Cutting J E amp Vishton P M (1995) Perceiving the layoutand knowing distances The integration relative potencyand contextual use of different information about depth InW Epstein amp S Rogers (Eds) Perception of Space and Mo-tion (pp 69ndash117) Academic Press

Deering M (1992) High resolution virtual reality Proc ofSIGGRAPHrsquo92 Computer Graphics 26(2) 195ndash201

Desplat S (1997) Characterization des elements actuals etfuture de loculometre Metrovision-Sextant Technical Re-port Ecole Nationale Superieure de Physique de MarseillesNovember


Dhome M Richetin M Lapreste J P amp Rives G (1989)Determination of the attitude of 3-D objects from a singleperspective view IEEE Trans Pattern Analysis and MachineIntelligence 11(12) 1265ndash1278

Droessler J G amp Rotier D J (1990) Tilted cat helmet-mounted display Optical Engineering 29(8) 849ndash854

Edwards E K Rolland J P amp Keller K P (1993) Videosee-through design for merging of real and virtual environ-ments Proc of IEEE VRAISrsquo93 (pp 223ndash233)

Edwards P J Hawkes D J Hill D L G Jewell D SpinkR Strong A amp Gleeson M (1995) Augmentation ofreality using an operating microscope for otolaryngologyand neurosurgical guidance J Image Guided Surgery 1(3)172ndash178

Ellis S R amp Bucher U J (1994) Distance perception ofstereoscopically presented virtual objects optically superim-posed on physical objects in a head-mounted see-throughdisplay Proc of the Human Factors and Ergonomic SocietyNashville

Fernie A (1995) A 31 Helmet-mounted Display with DualResolution CAE Electronics Ltd Montreal Canada SID95 Applications Digest 37ndash40

Fernie A (1995) Improvements in area of interest helmet-mounted displays In Proc of SID95

Fry G A (1969) Geometrical Optics Chilton Book CompanyFurness T A (1986) The super cockpit and its human factors

challenges Proceedings of the Human Factors Society 3048ndash52

Held R amp Durlach N (1987) Telepresence time delay andadaptation NASA Conference publication 10032

Holloway R (1995) An analysis of registration errors in a see-through head-mounted display system for craniofacial sur-gery planning Unpublished doctoral dissertation Universityof North Carolina at Chapel Hill

Hua H Girardot A Gao C and Rolland JP (2000) En-gineering of head-mounted projective displays Applied Op-tics (in press)

Jacobs M C Livingston M A amp State A (1997) Manag-ing latency in complex augmented reality systems Proc of1997 Symposium on Interactive 3D Graphics ACMSIGGRAPH 235ndash240

Kaiser Electro-Optics (1994) Personal communication fromFrank Hepburn of KEO Carlsbad CA General descriptionof Kaiserrsquos VIM system (lsquolsquoFull immersion head-mounteddisplay systemrsquorsquo) is available via ARPArsquos ESTO home pagehttpestosysplancomESTO

Kancherla A Rolland J P Wright D amp Burdea G

(1995) A novel virtual reality tool for teaching dynamic 3Danatomy Proc of CVRMed rsquo95 163ndash169

Kandebo S W (1988) Navy to evaluate Agile Eye helmet-mounted display system Aviation Week amp Space TechnologyAugust 15 94ndash99

Kennedy R S amp Stanney K M (1997) Aftereffects in vir-tual environment exposure Psychometric issues In M JSmith G Salvendy amp R J Koubek (Eds) Design of Com-puting

Kijima R amp Ojika T (1997) Transition between virtual en-vironment and workstation environment with projectivehead-mounted display Proc of VRAIS rsquo97 130ndash137

Manhart P K Malcom R J amp Frazee J G (1993) AugeyeA compact solid Schmidt optical relay for helmet mounteddisplays Proc of IEEE VRAIS rsquo93 234ndash245

Mizell D (1998) Personnal communicationMoses R A (1970) Adlers Physiology of the Eye St Louis

MO MosbyOster P J amp Stern J A (1980) lsquolsquoMeasurement of Eye

Movementrsquorsquo in I Martin amp P H Venables (Eds) Tech-niques of Psychophysiology New York John Wiley amp Sons

Outters V Argotti Y amp Rolland J P (1999) Knee motioncapture and representation in augmented reality TechnicalReport TR99-006 University of Central Florida

Parsons J amp Rolland J P (1998) A non-intrusive displaytechnique for providing real-time data within a surgeonscritical area of interest Proc of Medicine Meets Virtual Real-ity 246ndash251

Peuchot B (1993) Camera virtual equivalent model 001pixel detectors Special issue on 3D Advanced Image Process-ing in Medicine in Computerized Medical Imaging andGraphics 17(45) 289ndash294

mdashmdashmdash (1994) Utilization de detecteurs subpixels dans lamodelisation drsquoune cameramdashverification de lrsquohypothesestenope 9e Congres AFCET reconnaissance des formes et in-telligence artificielle Paris

mdashmdashmdash (1998) Personal communicationPeuchot B Tanguy A amp Eude M (1994) Dispositif op-

tique pour la visualization drsquoune image virtuelle tridimen-sionelle en superimposition avec an object notamment pourdes applications chirurgicales Depot CNRS 94106623May 31

mdashmdashmdash (1995) Virtual reality as an operative tool during sco-liosis surgery Proc of CVRMed rsquo95 549ndash554

Robinett W amp Rolland J P (1992) A computational modelfor the stereoscopic optics of a headmounted display Pres-ence Teleoperators and Virtual Environments 1(1) 45ndash62


Rock I (1966) The nature of perceptual adaptation NewYork Basic Books

Rolland J P (1994) Head-mounted displays for virtual envi-ronments The optical interface International Optical De-sign Conference 94 Proc OSA 22 329ndash333

Rolland J P (1998) Mounted displays Optics and PhotonicsNews 9(11) 26ndash30

Rolland J P (2000) Wide angle off-axis see-through head-mounted display Optical EngineeringmdashSpecial Issue onPushing the Envelop in Optical Design Software 39(7) (inpress)

Rolland J P Ariely D amp Gibson W (1995) Towards quan-tifying depth and size perception in virtual environmentsPresence Teleoperators and Virtual Environments 4(1)24ndash49

Rolland J P amp Arthur K (1997) Study of depth judgmentsin a see-through mounted display Proceeding of SPIE 3058AEROSENSE 66ndash75

Rolland J P Holloway R L amp Fuchs H (1994) A com-parison of optical and video see-through head-mounted dis-plays Proceedings of SPIE 2351 293ndash307

Rolland J P amp Hopkins T (1993) A method for computa-tional correction of optical distortion in head-mounted dis-plays Technical Report TR93-045 University of NorthCarolina at Chapel Hill

Rolland J P Krueger M and Goon A (2000) Multifocusplanes in head-mounted displays Applied Optics 39(19) inpress

Rolland J P Parsons J Poizat D amp Hancock D (1998)Conformal optics for 3D visualization Proc of the Interna-tional Lens Design Conference Hawaii (June) 760ndash764

Rolland J P Yoshida A Davis L amp Reif J H (1998)High-resolution inset head-mounted display Applied Optics37(19) 4183ndash4193

Roscoe S N (1984) Judgments of size and distance with im-aging displays Human Factors 26(6) 617ndash629

Roscoe S N (1991) The eyes prefer real images in PictorialCommunication in Virtual and Real Environments EdStephen R Ellis Taylor and Francis

Scher-Zagier E J (1997) A humanrsquos eye view motion blurand frameless rendering ACM Crosswords 97

Shenker M (1994) Image quality considerations for head-mounted displays Proc of the OSA International Lens De-sign Conference 22 334ndash338

Shenker M (1998) Personal CommunicationSlater M and Wilbur S (1997) A framework for immersive

virtual environments (FIVE) Speculations on the role of

presence in virtual environments Presence Teleoperators andVirtual Environments 6(6) 603ndash616

State A Chen D Tector C Brandt C Chen H Ohbu-chi R Bajura M and Fuchs H (1994) Case study Ob-serving a volume rendered fetus within a pregnant patientProceedings of Visualization rsquo94 Washington DC 364ndash373

State A Hirota G Chen D T Garrett W E and Living-ston M (1996) Superior augmented-reality registration byintegrating landmark tracking and magnetic tracking Procof SIGGRAPH 1996 ACM SIGGRAPH 429ndash438

Sutherland I (1965) The ultimate display Information Pro-cessing 1965 Proc of IFIP Congress 65 506ndash508

Sutherland I E (1968) A head-mounted three-dimensionaldisplay Fall Joint Computer Conference AFIPS ConferenceProceedings 33 757ndash764

Thomas M L Siegmund W P Antos S E and RobinsonR M (1989) Fiber optic development for use on the fiberoptic helmet-mounted display in Helmet-Mounted DisplaysJ T Colloro Ed Proc of SPIE 1116 90ndash101

Tomasi C amp Kanade T (1991) Shape and motion from im-age streams a factorization method-Part 3 Detection andtracking of point features Carnegie Mellon technical reportCMU-CS-91-132

Vaissie L amp Rolland J P (1999) Eye-tracking integration inhead-mounted displays Patent Filed January

Vaissie L Rolland J P (2000) Alberliacutean errors in head-mounted displays choice of eyejoints locationTechnical Report TR 2000-001 University of CentralFlorida

Watson B amp Hodges L F (1995) Using texture maps tocorrect for optical distortion in head-mounted displaysProc of VRAISrsquo95 172ndash178

Welch B amp Shenker M (1984) The fiber-optic Helmet-Mounted Display Image III 345ndash361

Welch R B (1993) Alternating prism exposure causes dualadaptation and generalization to a novel displacement Per-ception and Psychophysics 54(2) 195ndash204

Welch R B (1994) Adapting to virtual environments andteleoperators Unpublished manuscript NASA-Ames Re-search Center Moffett Field CA

Welch G amp Bishop G (1997) SCAAT Incremental track-ing with incomplete information Proc of ACMSIGGRAPH (pp 333ndash344)

Wright D L Rolland J P amp Kancherla A R (1995) Us-ing virtual reality to teach radiographic positioning Radio-logic Technology 66(4) 167ndash172


Seeley amp Vukobratovich 1981 Furness 1986Droessler amp Rotier 1990 Barrette 1992 Kandebo1988 Desplat 1997) developments in 3-D scientificand medical visualization were initiated in the 1980s atthe University of North Carolina at Chapel Hill (Brooks1992)

In this paper we shall first review several medical visu-alization applications developed using optical and videosee-through technologies We shall then discuss techno-logical and human-factors and perceptual issues relatedto see-through devices some of which are employed in

the various applications surveyed Finally we shall dis-cuss what the technology may evolve to become

2 Some Past and Current Applications ofOptical and Video See-Through HMDs

The need for accurate visualization and diagnosisin health care is crucial One of the main developmentsof medical care has been imaging Since the discovery ofX-rays in 1895 by Wilhelm Roentgen and the first X-rayclinical application a year later by two Birmingham (UK)doctors X-ray imaging and other medical imaging mo-dalities (such as CT ultrasound and NMR) haveemerged Medical imaging allows doctors to view as-pects of the interior architecture of living beings thatwere unseen before With the advent of imaging tech-nologies opportunities for minimally invasive surgicalprocedures have arisen Imaging and visualization can beused to guide needle biopsy laparoscopic endoscopicand catheter procedures Such procedures do requireadditional training because the physicians cannot see thenatural structures that are visible in open surgery Forexample the natural eye-hand coordination is not avail-able during laparoscopic surgery Visualization tech-niques associated with see-through HMDs promise tohelp restore some of the lost benefits of open surgery(for example by projecting a virtual image directly onthe patient eliminating the need for a remote monitor)

The following paragraphs briefly discuss examples ofrecent and current research conducted with optical see-through HMDs at the University of North Carolina atChapel Hill (UNC-CH) the University of CentralFlorida (UCF) and the United Medical and DentalSchools of Guyrsquos and Saint Thomasrsquos Hospitals in En-gland video-see-through at UNC-CH and hybrid opti-cal-video see-through at the University of Blaise Pascalin France

A rigorous error-analysis for an optical see-throughHMD targeted toward the application of optical see-through HMD to craniofacial reconstruction was con-ducted at UNC-CH (Holloway 1995) The superimpo-sition of CT skull data onto the head of the real patientwould give the surgeons lsquolsquoX-ray visionrsquorsquo The premise of

Figure 1 Optical see-through head-mounted

display (Photo courtesy of KaiserElectro-Optics)

Figure 2 A custom optics video

see-through head-mounted display

developed at UNC-CH Edwards et al

(1993) designed the miniature video

cameras The viewer was a large FOV

opaque HMD from Virtual Research
















(1999)















abdomen (1994)

















Peuchot
































































(1997)




























































4 Conclusion






Acknowledgments




References




































































































(1999)















abdomen (1994)

















Peuchot
































































(1997)




























































4 Conclusion






Acknowledgments




References



































































































abdomen (1994)

















Peuchot
































































(1997)




























































4 Conclusion






Acknowledgments




References





























































































Peuchot
































































(1997)




























































4 Conclusion






Acknowledgments




References


















































































































































(1997)




























































4 Conclusion






Acknowledgments




References
















































































































































4 Conclusion






Acknowledgments




References























































































Acknowledgments




References






















































































school of optics/creol, and school of electrical engineering and...

Documents