Marker Tracking and HMD Calibration for a Video-based Augmented Reality Conferencing System

Transcription

1 Marker Traking and HMD Calibration for a Video-based Augmented Reality Conferening System Hirokazu Kato 1 and Mark Billinghurst 2 1 Faulty of Information Sienes, Hiroshima City University 2 Human Interfae Tehnology Laboratory, University of Washington kato@sys.im.hiroshima-u.a.jp, grof@hitl.washington.edu Abstrat We desribe an augmented reality onferening system whih uses the overlay of virtual images on the real world. Remote ollaborators are represented on Virtual Monitors whih an be freely positioned about a user in spae. Users an ollaboratively view and interat with virtual objets using a shared virtual whiteboard. This is possible through preise virtual image registration using fast and aurate omputer vision tehniques and HMD alibration. We propose a method for traking fiduial markers and a alibration method for optial see-through HMD based on the marker traking. 1. Introdution Computers are inreasingly used to enhane ollaboration between people. As ollaborative tools beome more ommon the Human-Computer Interfae is giving way to a Human-Human Interfae mediated by omputers. This emphasis adds new tehnial hallenges to the design of Human Computer Interfaes. These hallenges are ompounded for attempts to support three-dimensional Computer Supported Collaborative Work (CSCW). Although the use of spatial ues and three-dimensional objet manipulation are ommon in fae-to-fae ollaboration, tools for three-dimensional CSCW are still rare. However new 3D interfae metaphors suh as virtual reality may overome this limitation. Virtual Reality (VR) appears a natural medium for 3D CSCW; in this setting omputers an provide the same type of ollaborative information that people have in fae-to-fae interations, suh as ommuniation by objet manipulation, voie and gesture [1]. Work on the DIVE projet [2], GreenSpae [3] and other fully immersive multi-partiipant virtual environments has shown that ollaborative work is indeed intuitive in suh surroundings. However most urrent multi-user VR systems are fully immersive, separating the user from the real world and their traditional tools. As Grudin [4] points out, CSCW tools are generally rejeted when they fore users to hange the way they work. This is beause of the introdution of seams or disontinuities between the way people usually work and the way they are fored to work beause of the omputer interfae. Ishii desribes in detail the advantages of seamless CSCW interfaes [5]. Obviously immersive VR interfaes introdue a huge disontinuity between the real and virtual worlds. An alternative approah is through Augmented Reality (AR), the overlaying of virtual objets onto the real world. In the past researhers have explored the use of AR approahes to support fae-to-fae ollaboration. Projets suh as Studierstube [6], Transvision [7], and AR2 Hokey [8] allow users an see eah other as well as 3D virtual objets in the spae between them. Users an interat with the real world at the same time as the virtual images, bringing the benefits of VR interfaes into the real world and failitating very natural ollaboration. In a previous paper we found that this meant that users ollaborate better on a task in a fae-to-fae AR setting than for the same task in a

2 fully immersive Virtual Environment [9]. We have been developing a AR onferening system that allows virtual images (Virtual Monitors) of remote ollaborators to be overlaid on the users real environment. Our Augmented Reality onferening system tries to overomes some of the limitations of urrent desktop video onferening, inluding the lak of spatial ues [1], the diffiulty of interating with shared 3D data, and the need to be physially present at a desktop mahine to onferene. While using this system, users an easily hange the arrangement of Virtual Monitors, plaing the virtual images of remote partiipants about them in the real world and they an ollaboratively interat with 2D and 3D information using a Virtual Shared Whiteboard. The virtual images are shown in a lightweight head mounted display, so with a wearable omputer our system ould be made portable enabling ollaboration anywhere in the workplae. In developing a multi-user augmented reality video onferening system, preise registration of the virtual images with the real world is one of the greatest hallenges. In our work we use omputer vision tehniques and have developed some optimized algorithms for fast, aurate real time registration and onvenient optial see-through HMD alibration. In this paper, after introduing our onferening appliation, we desribe the video-based registration and alibration methods used. 2. System overview Our prototype system supports ollaboration between a user wearing see-through head mounted displays(hmd) and those on more traditional desktop interfaes as shown in figure 1. This simulates the situation that ould our in AR user with an optial see-through head mounted display and a amera Desktop omputer users with a amera Figure 1. System onfiguration. ollaboration between a desk bound expert and a remote field worker. The user with the AR head mounted interfae an see video images from desktop users and be supported by them. Remote desktop users an see the video images that the small amera of the AR user grabs and give support to the AR user. This system dose not support video ommuniation among desktop users. If neessary, however, they ould simultaneously exeute a traditional video ommuniation appliation. In this setion we first desribe the AR head mounted interfae and then the desktop interfae Augmented Reality Interfae The user with the AR interfae wears a pair of the Virtual i-o iglasses HMD that have been modified by adding a small olor amera. The iglasses are full olor, an be used in either a see-through or oluded mode and have a resolution of 263x234 pixels. The amera output is onneted to an SGI O2 (R5SC 18MHz CPU) omputer and the video out of the SGI is onneted bak into the head mounted display. The O2 is used for both image proessing of video from the head mounted amera and virtual image generation for the HMD. Performane speed is 7-1 frames per se for full version, 1-15 fps running without the Virtual Shared Whiteboard. The AR user also has a set of small marked ards and a larger piee of paper with six letters on it around the outside. There is one small marked ard for eah remote ollaborator with their name written on it. These are plaeholders (user ID ards) for the Virtual Monitors showing the remote ollaborators, while the larger piee of paper is a plaeholder for the shared white board. To write and interat with virtual objets on the shared whiteboard the user has a simple light pen onsisting of an LED, swith and battery mounted on a pen. When the LED touhes a surfae the swith is tripped and it is turned on. Figure 2 shows an observer's view of the AR user using the interfae. The software omponents of the interfae onsist of two parts, the Virtual Monitors shown on the user ID ards, and the Virtual Shared Whiteboard. When the system is running, omputer vision tehniques are used to identify speifi user

3 Figure 2. Using the Augmented Reality Interfae. Figure 3. Remote user representation in the AR interfae. ID ards (using the user name on the ard) and display live video of the remote user that orresponds to the ID ard. Vision tehniques are also used to alulate head position and orientation relative to the ards so the virtual images are preisely registered with the ID ards. Figure 3 shows an example of a Virtual Monitor, in this ase the user is holding an ID ard whih has live video from a remote ollaborator attahed to it. Shared whiteboards are ommonly using in ollaborative appliations to enables people to share notes and diagrams. In our appliation we use a Virtual Shared Whiteboard as seen in figure 4. This is shown on a larger paper board with six similar registration markings as the user ID ards. Virtual annotations written by remote partiipants are displayed on it, exatly aligned with the plane of the physial ard. The loal partiipant an use the light-pen to draw on the ard and add their own annotations, whih are in turn displayed and transferred to the remote desktops. The user an erase Figure 4. Virtual Shared White Board. their own annotations by touhing one orner of the ard. Currently our appliation only supports virtual annotations aligned with the surfae of the ard, but we are working on adding support for shared 3D objets. The position and pose of this paper board an be estimated by using the same vision methods used for the virtual monitors. However, sine the user's hands often olude the registration markers, the estimation has to be done by using only visible markers. We an reliably estimate the ard position using only one of the six markers. The LED lightpen is on while it touhes the paper board. When this happens the system estimates the position of the pen tip relative to the paper board from the 2D position of the LED in the amera image and the knowledge that the tip of the pen is ontat with the board. Users an pik up the ard for a loser to look at the images on the virtual whiteboard, and an position it freely within their real workspae Desktop Interfae The AR user ollaborates with remote desktop users that have a more traditional interfae. The desktop users are on networked SGI omputers. Users with video ameras on their omputer see a video window of the video image that their amera is sending, the remote video from the AR head mounted amera and a share white board appliation. The video from the AR user's head mounted amera enables the desktop user to ollaborate more effetively with them on real world tasks. They an freely draw on the shared white board using the mouse, and whiteboard annotations and video frames from their amera are send to the AR user.

4 3. Video-based registration tehniques Our AR onferening interfae relies heavily on omputer vision tehniques for ID reognition and user head position and pose determination. In the remainder of the paper we outline the underlying omputer vision methods we have developed to aomplish this. These methods are general enough to be appliable for a wide range of augmented reality appliations. Augmented Reality Systems using HMDs an be lassified into two groups aording to the display method used: Type A: Video See-through Augmented Reality Type B: Optial See-through Augmented Reality In type A, virtual objets are superimposed on a live video image of the real world aptured by the amera attahed to the HMD. The resulting omposite video image is displayed bak to both eyes of the user. In this ase, interation with the real world is a little unnatural beause the amera viewpoint shown in the HMD is offset from that of the user's own eyes, and the image is not stereographi. Performane an also be signifiantly affeted as the video frame rate drops. However, this type of system an be realized easily, beause good image registration only requires that the relationship between 2D sreen oordinates on the image and 3D oordinates in the real world is known. In type B, virtual objets are shown diretly on the real world by using a see-through display. In this ase, the user an see the real world diretly and stereosopi virtual images an be generated so the interation is very natural. However, the image registration requirements are a lot more hallenging beause it requires the relationships between the amera, the HMD sreens and the eyes to be known in addition to the relationships used by type A systems. The alibration of the system is therefore very important for preise registration. Azume reported a good review of the issues faed in augmented reality registration and alibration[11]. Also many registration tehniques have been proposed. State proposed a registration method using stereo images and a magneti traker[12]. Neumann used a single amera and multiple fiduial markers for robust traking[13]. Rekimoto used vision tehniques to identify 2D matrix markers[14]. Klinker used square markers for fast traking[15]. Our approah is similar to this method. We have developed a preise registration method for the optial see-through augmented reality system. Our method overomes two primary problems; alibration of the HMD and amera, and estimating an aurate position and pose of fiduial markers. We first desribe a position and pose estimation method, and then HMD and amera alibration method, beause our HMD alibration method is based on the fiduial marker traking. 4. Position and pose estimation of markers 4.1. Estimation of the Transformation Matrix Size-known square markers are used as a base of the oordinates frame in whih Virtual Monitors are represented (Figure 5). The transformation matries from these marker oordinates to the amera oordinates (T m ) represented in eq.1 are estimated by image analysis. X V11 V12 V13 Wx Xm Y V V V W y Y m = Z V31 V32 V33 Wz Zm 1 1 Xm Ym = V 3 3 W3 1 1 = Zm Marker Camera Sreen Coordinates Marker Coordinates (Xm, Ym, Zm) T m Xm Y m Zm (x, y) Camera Coordinates (X, Y, Z) (eq. 1) Figure 5. The relationship between marker oordinates and the amera oordinates is estimated by image analysis.

5 After thresholding of the input image, regions whose outline ontour an be fitted by four line segments are extrated. Parameters of these four line segments and oordinates of the four verties of the regions found from the intersetions of the line segments are stored for later proesses. The regions are normalized and the sub-image within the region is ompared by template mathing with patterns that were given the system before to identify speifi user ID markers. User names or photos an be used as identifiable patterns. For this normalization proess, eq.2 that represents a perspetive transformation is used. All variables in the transformation matrix are determined by substituting sreen oordinates and marker oordinates of deteted marker's four verties for (x, y ) and (X m, Y m ) respetively. After that, the normalization proess an be done by using this transformation matrix. hx N11 N12 N13 Xm hy = N N N Y m h N N (eq. 2) When two parallel sides of a square marker are projeted on the image, the equations of those line segments in the amera sreen oordinates are the following: ax 1 + by =, ax 2 + by = (eq. 3) For eah of markers, the value of these parameters has been already obtained in the line-fitting proess. Given the perspetive projetion matrix P that is obtained by the amera alibration in eq.4, equations of the planes that inlude these two sides respetively an be represented as eq.5 in the amera oordinates frame by substituting x and y in eq.4 for x and y in eq.3. P11 P12 P13 P 22 P23 P = 1, 1 hx X hy Y = P h Z ap 1 11X+ ( ap bp 1 22) Y + ( ap bp ) Z = apx + ( ap + bp ) Y+ ( ap + bp + ) Z = (eq. 4) (eq. 5) Given that normal vetors of these planes are n 1 and n 2 respetively, the diretion vetor of parallel two sides of the v 1 u 1 v 2 Figure 6. Two perpendiular unit diretion vetors: v 1, v 2 are alulated from u 1 and u 2. square is given by the outer produt n n. Given that 1 2 two unit diretion vetors that are obtained from two sets of two parallel sides of the square is u 1 and u 2, these vetors should be perpendiular. However, image proessing errors mean that the vetors won't be exatly perpendiular. To ompensate for this two perpendiular unit diretion vetors are defined by v 1 and v 2 in the plane that inludes u 1 and u 2 as shown in figure 6. Given that the unit diretion vetor whih is perpendiular to both v 1 and v 2 is v 3, the rotation omponent V 3x3 in the transformation matrix T m from marker oordinates to amera oordinates speified in eq.1 t t is [V 1 V 2 V 3t ]. Sine the rotation omponent V 3x3 in the transformation matrix was given, by using eq.1, eq.4, the four verties oordinates of the marker in the marker oordinate frame and those oordinates in the amera sreen oordinate frame, eight equations inluding translation omponent W x W y W z are generated and the value of these translation omponent W x W y W z an be obtained from these equations. The transformation matrix found from the method mentioned above may inlude error. However this an be redued through the following proess. The vertex oordinates of the markers in the marker oordinate frame an be transformed to oordinates in the amera sreen oordinate frame by using the transformation matrix obtained. Then the transformation matrix is optimized as sum of the differene between these transformed oordinates and the oordinates measured from the image goes to a minimum. Though there are six independent variables in the transformation matrix, only the rotation omponents are optimized and then the translation omponents are reestimated by using the method mentioned above. By iteration of this proess a number of times the transformation u 2

6 matrix is more aurately found. It would be possible to deal with all of six independent variables in the optimization proess. However, omputational ost has to be onsidered An Extension for the Virtual Shared White Board White Board Coordinates Frame (Xw, Yw, Zw) Figure 7. Layout of markers on the Shared White Board. The method desribed for traking user ID ards is extended for traking the shared whiteboard ard. There are six markers in the Virtual Shared White Board, aligned around the outside of the board as shown in figure 7. The orientation of the White Board is found by fitting lines around the fiduial markers and using an extension of the tehnique desribed for traking user ID ards. Using all six markers to find the board orientation and align virtual images in the interior produes very good registration results. However, when a user draws a virtual annotation, some markers may be oluded by user's hands, or they may move their head so only a subset of the markers are in view. The transformation matrix for Virtual Shared White Board has to be estimated from visible markers so errors are introdued when fewer markers are available. To redue errors the line fitting equations are found by both onsidering individual markers and sets of aligned markers. Eah marker has a unique letter in its interior that enables the system to identify markers whih should be horizontally or vertially aligned and so estimate the board rotation. Though line equations in the amera sreen oordinates frame are independently generated for eah of markers, the alignment of the six markers in Virtual Shared White Board means that some line equations are idential. Therefore by extrating all aligned sides from visible markers for the linefitting, eah line equation is alulated by using all the ontour information that is on the extrated sides. Furthermore by using all the equations of the deteted parallel lines, the diretion vetors are estimated and the board orientation is found Pen Detetion The light-pen is on while touhing the shared whiteboard board. Estimation of the pen tip loation is found in the following way. First, the brightest region in the image is extrated and the enter of the gravity is deteted. If brightness and area of the regions are not satisfied with heuristi rules, the light-pen is regarded as turned off status. Sine pen position (X w, Y w, Z w ) is expressed relative to the Virtual Shared Whiteboard it is deteted in the whiteboard oordinate frame. The relationship between the amera sreen oordinates and the whiteboard oordinates is given by eq.6. (x, y ) is a position of the enter of gravity that is deteted by image proessing. Also Z w is equal to zero sine pen is on the board. By using these values in eq.6, two equations inluding X w and Y w as variables are generated and their values are alulated easily by solving these equations. hx X V11 V12 V13 Wx Xw hy Y V V V W y Y w = P = P h Z V31 V32 V33 Wz Zw HMD and Camera Calibration (eq. 6) In an optial see-through HMD, a ray from a physial objet reahes the foal point of the eye through the HMD sreen. Then, a 3D position represented in the eye oordinates whose origin is the foal point of the eye an be projeted on the HMD sreen oordinates by the perspetive projetion model. This assumes that the Z axis perpendiularly rosses the HMD sreen, and the X and Y axes are parallel to X and Y axes of the HMD sreen oordinates frame respetively. Figure 8 shows oordinates frames in our alibration

7 proedure. As mentioned in setion 4, position and pose estimation of a marker is done by alulating the transformation matrix from marker oordinates to amera oordinates: T m (eq.1). The perspetive projetion matrix P (eq.4) is required in this proedure. Camera alibration is to find the perspetive projetion matrix P that represents the relationship between the amera oordinates and the amera sreen oordinates. In order to display virtual objets on HMD sreen as if those are on the marker, the relationship between the marker oordinates and the HMD sreen oordinates is required. Relationship between HMD sreen oordinates and eye oordinates is represented by the perspetive projetion. Also, relationship between amera oordinates and eye oordinates is represented by rotation and translation transformations. eq.7 shows those relationship. ixs Xe X Xm iy s Y e Y Y m = Q = = se QseTe QseTeTm i Ze Z Zm Qse : Perspetive transformation matrix T : Rotation and translation matrix e HMD Sreen Coordinates (xs, ys) Marker Coordinates (Xm, Ym, Zm) Eye Coordinates (Xe, Ye, Ze) Camera Sreen Coordinates (x, y) Camera Coordinates (X, Y, Z) Figure 8. Coordinates frames in our alibration proedure. (eq. 7) Matrix T m representing the transformation from marker oordinates to amera oordinates is obtained in use of the system as mentioned in Setion 4. HMD alibration is therefore to find the matrix Q se T e for both of eyes Camera Calibration - Finding the matrix P lines for the amera alibration. Coordinates of all ross points of a grid are known in the ardboard loal 3D oordinates. Also those in the amera sreen oordinates an be deteted by image proessing after the ardboard image is grabbed. Many pairs of the ardboard loal 3D oordinates (X t, Y t, Z t ) and the amera sreen oordinates (x, y ) are used for finding the perspetive transformation matrix P. The relationships among the amera sreen oordinates (x, y ), the amera oordinates (X, Y, Z ) and the ardboard oordinates (X t, Y t, Z t ) an be represented as: hx X Xt Xt C11 C12 C13 C14 Xt hy Y Y t Y t C21 C22 C23 C 24 Y t = P = P T = C = t h Z Zt Zt C31 C32 C33 1 Zt sxf x R11 R12 R13 Tx sf y y P = 1, T = R 21 R22 R23 Ty t R31 R32 R33 Tz 1 1 (eq. 8) where P is the perspetive transformation matrix whih should be found here, f is the foal length, s x is the sale fator [pixel/mm] in diretion of x axis, s y is the sale fator in diretion of y axis, (x, y ) is the position that Z axis of the eye oordinates frame passes, T t represents the translation and rotation transformation from the ardboard oordinates to the amera oordinates and C is the transformation matrix obtained by ombining P and T t. Sine many pairs of (x, y ) and (X t, Y t, Z t ) have been obtained by the proedure mentioned above, matrix C an be estimated. However, the matrix C annot be deomposed into P and T t in general beause matrix C has 11 independent variables but matries P and T t have 4 and 6 respetively, so the sum of the independent variables of P and T t is not equal to the one of C. A salar variable k is added into P to make these numbers equal as the following: hx sxf k x R11 R12 R13 Tx Xt hy sf y y R R R T y Y t = h 1 R31 R32 R33 Tz Zt (eq. 9) We use a simple ardboard frame with a ruled grid of

8 As a result, the matrix C an be deomposed into P and T t. The variable k means the slant between x-axis and y- axis and should be zero ideally but it may be a small noise value. 5.2 HMD Calibration - Finding the matrix Q se T e Formulation of the matrix Q se T e is same as one of the matrix C in eq.8. Therefore many pairs of the oordinates (x s, y s ) and (X, Y, Z ) an be used for finding the transformation matrix ombining Q se and T e. In order to obtain suh kinds of data, we use marker traking tehnique introdued in setion 4. HMD alibration proedure is done for eah eye. A rosshair ursor is displayed on the orresponding HMD sreen. The user handles a fiduial marker and fits its enter on the ross-hair ursor as shown in figure 9. The fiduial marker is simultaneously observed by the amera attahed on the HMD and the entral oordinates are deteted in the amera oordinates. While the user manipulates the marker from near side to far side, some marker positions are stored by liking a mouse button. In this proedure, positions of the ross-hair ursor mean HMD sreen oordinates (x s, y s ) and marker positions mean amera oordinates (X, Y, Z ). After iterating this operation in some positions of ross-hair ursor, many pairs of (x s, y s ) and (X, Y, Z ) are obtained. At last the transformation matrix ombining Q se and T e is found. Some alibration methods for optial see-through HMD have been proposed. However, most of those require that users hold their head position during the alibration[15]. This onstraint is a ause of diffiulties of HMD alibration. Obviously our alibration method dose not need this kind of onstrains. So this alibration method an be used onveniently. 6. Evaluation of registration and alibration 6.1. Auray of the marker detetion In order to evaluate auray of the marker detetion, deteted position and pose were reorded while the square marker with 8[mm] of side length was moved in depth diretion with some slants. Figure 1 shows errors of position. Figure 11 shows deteted slant. This result shows that auray dereases the further the ards are from the amera. error [mm] 3 2 slant [deg] distane [mm] 9 Figure 1. Errors of position. slant [deg] slant [deg] distane [mm] Figure 9. HMD alibration. Figure 11. Deteted slant.

9 6.2. Evaluation of HMD alibration Our HMD alibration method was evaluated by using a program that displays a square of same size as the marker on it. A user with HMD looks at a displayed square on the marker and reports the deviation of a displayed square from the marker. This evaluation was done for 3 tasks: Task 1: holding the marker. (Eye-marker distane: 3mm) Task 2: putting the marker on a desk. (Eye-marker distane: 4mm) Task 3: putting the marker far away on a desk. (Eye-marker distane: 8mm) Also we had 3 onditions: (a) Evaluation with standard parameters. (b) Evaluation with alibrated parameters. () Evaluation with alibrated parameters, but user took off the HMD one after alibration. Standard parameters mean ones whih had been alibrated by another user. 1 times ross-hair ursor fittings were done for eah eye. Table 1 shows results of this user testing. This result seems to be good. However, it inludes a problem: Foal point of the HMD is on 2-3[m] distane, but a user have to see a virtual objet on 3-8[mm] distane. Hereby the user see the virtual objet out of fous. This means that reporting a preise deviation is very diffiult beause of this defoused situation. As a result, test user might report good-will answer. However, we an see the user Table 1. Results of user testing. time (min) A 3 B 2 C 2 D 2 E 2 ondition task 1 (mm) task 2 (mm) task 3 (mm) ( a) ( b) 5 ( ) ( a) ( b) 5 5 ( ) 5 5 ( a) ( b) ( ) 2 2 ( a) ( b) ( ) 1 2 ( a) ( b) ( ) improvement of the registration by using alibrated parameters. 7. Conlusions In this paper we have desribed a new Augmented Reality onferening appliation and the omputer vision tehniques used in the appliation. Our omputer vision methods give good results when the markers are lose to the user, but auray dereases the further the ards are from the amera. Also our HMD alibration method whih does not require a non-moving user give good results without user's patiene. In future, we will improve this AR onferening prototype and exeute user testing for its evaluation as a ommuniation system. Referenes [1] A. Wexelblat, "The Reality of Cooperation: Virtual Reality and CSCW", Virtual Reality: Appliations and Explorations. Edited by A. Wexelblat. Boston, Aademi Publishers, [2] C. Carlson, and O. Hagsand, "DIVE - A Platform for Multi-User Virtual Environments", Computers and Graphis, Nov/De 1993, Vol. 17(6), pp [3] J. Mandeville, J. Davidson, D. Campbell, A. Dahl, P. Shwartz, and T. Furness, "A Shared Virtual Environment for Arhitetural Design Review", CVE '96 Workshop Proeedings, 19-2th September 1996, Nottingham, Great Britain. [4] J. Grudin, "Why CSCW appliations fail: Problems in the design and evaluation of organizational interfaes", Proeedings of CSCW '88, Portland, Oregon, 1988, New York: ACM Press, pp [5] H. Ishii, M. Kobayashi, K. Arita, "Iterative Design of Seamless Collaboration Media", Communiations of the ACM, Vol 37, No. 8, August 1994, pp [6] D. Shmalsteig, A. Fuhrmann, Z. Szalavari, M. Gervautz, "Studierstube - An Environment for Collaboration in Augmented Reality", CVE '96 Workshop Proeedings, 19-2th September 1996, Nottingham, Great Britain. [7] J. Rekimoto, "Transvision: A Hand-held Augmented

10 Reality System for Collaborative Design", Proeeding of Virtual Systems and Multimedia '96 (VSMM '96), Gifu, Japan, 18-2 Sept., [8] T. Ohshima, K. Sato, H. Yamamoto, H. Tamura, "AR2Hokey:A ase study of ollaborative augmented reality", Proeedings of VRAIS'98, pp [9] M. Billinghurst, S. Weghorst, T. Furness, "Shared Spae: An Augmented Reality Approah for Computer Supported Cooperative Work", Virtual Reality Vol. 3(1), 1998, pp [1] A. Sellen, "Speeh Patterns in Video-Mediated Conversations", Proeedings CHI '92, May 3-7, 1992, ACM: New York, pp [11] R. Azuma, "SIGGRAPH95 Course Notes: A Survey of Augmented Reality", Los Angeles, Assoiation for Computing Mahinery, [12] A. State, G. Hirota, D. T. Chen, W. F. Garrett, M. A. Livingston, "Superior Augmented Reality Registration by Integrating Landmark Traking and magneti Traking", Proeedings of SIGGRAPH96, pp , [13] U. Neumann, S. You, Y. Cho, J. Lee, J. Park, "Augmented Reality Traking in Natural Environments", Mixed Reality - Merging Real and Virtual Worlds (Ed. by Y. Ohta and H. Tamura), Ohmsha and Springer-Verlag, pp.11-13, [14] J. Rekimoto, "Matrix: A Realtime Objet Identifiation and Registration Method for Augmented Reality", Proeedings of Asia Paifi Computer Human Interation 1998 (APCHI'98), Japan, Jul , [15] G. Klinker, D. Striker, D. Reiners, "Augmented Reality: A Balaning At Between High Quality and Real-Time Constraints", Proeedings of ISMR '99, 1999, pp