Head Motion and Latency Compensation on Localization of 3D Sound in Virtual Reality

Transcription

1 Head Motion and Latency Compensation on Locaization of 3D Sound in Virtua Reaity Jiann-Rong Wu, Cha-Dong Duh, Ming Ouhyoung Communication and Mutimedia Lab. Department of Computer Science and Information Engineering Nationa Taiwan University, Taiwan, R.O.C. Jei-Tun Wu Department of Psychoogy Nationa Taiwan University, Taiwan, R.O.C. ABSTRACT As part of designing a muti-sensory VR environment, in addition to evauating the visua subsystem, acoustic presence must be evauated. This paper proposes two experiments to examine the effects of human head movement and atency compensation in 3D sound ocaization. There are two hypotheses, the first hypothesis is that through the computer simuation of 3D sound, dynamic head movement can hep in the ocaization of sound in space, as compared to fixed head position. The second hypothesis is when there is atency introduced in the computer generation of 3D sound, a human subject can perform better in a sound ocating task with atency compensation than the one without compensation. The resuts of two proposed experiments corroborate with the above two hypotheses. Moreover, we are abe to identify that with dynamic head movement, the human capabiity can be enhanced by more than 90% in the ocaization of sound in space and at the same time reduce the front-back ambiguity. In the second experiment invoving a typica system atency at 300ms, it is shown that with compensation for atency the average time to perform sound ocating task can be reduced by more than 50% than that without compensation. A priori study of the atency effects indicates that if the atency is arger than I50ms, human performance of ocating a 3D sound source is noticeaby decreased. KEYWORDS: 1. INTRODUCTION Virtua reaity, 3D sound, ocaization of sound, atency compensation. Overa system atency, the eapsed time from input human motion to the immediate response of that input in the dispay, is one of the most frequenty cited shortcomings of current virtua environment (VE) technoogy when the atency is reativey arge. In a head-mounted dispay (HMD) based VE system with head tracker, if overa atency is onger than 200 ms, for instance, it wi cause motion sickness for ong time wearing. In a muti-sensory system, the situation is more compicated if Permission to make digita/hard copies ofa or part ofthis materia for persona or chssroom use is granted without fee provided that the copies are not made or distributed for profit or commercia advantage, the copyright notice, the tite ofthe pubication and its date appear, end notice is given that copyright is by pamission ofthe ACM, Inc. To copy otherwise, to repubish, to post on serva or to redistribute to ists. requires specific permission and/or fee ACM IQ&!? 97 Lausanne Switzerand Copyright 1997 ACM O ~x/97/9..%3.50 each sensory dispay has different atency. For exampe, in our interactive buiding wakthrough system, there are three kinds of sensory perception (vision, 3D sound, and mechanica sensory by treadmi) combined in the system with different ags. A preiminary research on the visua part of ag probem has been examined by a 3D tracking experiment on atency and its compensation method in virtua environments in The resuts showed that the Grey system predictor we proposed can reduce the atency significanty. Simiary, the auditory system may be sensitive to time ags in a VE. However, there is yet no research data avaiabe that describes the reationship between ocaization performance of sound source and head tracker atency [I]. Athough the power of a modem persona computer is enough to compute and payback 3D sound in rea time purey by software, the requirement of tempora consistency between audio and visua data makes the auditory subsystem suffer from the same atency probem of visua subsystem in an immersive VE. The new technoogy of 3D sound enabes us to conduct some experiments on the topics we aways want to do for studying the effect of combining 3D acoustic fideity in VE [1,2,3]. For exampe, in the experiments on the atency effect of a wakthrough project, a user can wak around an environment wearing an HMD. However, the graphics subsystem pus the space tracker hooked onto an HMD and the LCD pane dispay on the HMD atogether introduce a atency of 300 ms or more [4]. Therefore, when a user was using the system for more than 5 minutes, he or she sometimes fet dizzy and thus got motion sickness because of the reative high atency. This makes us think about another probem aso reated to virtua reaity: that in order to make the artificia environment ook rea and sound rea, we shoud aso introduce the 3D sound effect. What woud be the effect of the atency in a 3D sound environment? Is that simiar to the visua subsystem when the human perception can te the difference and behave differenty? Besides, there is another probem we are interested in, that is, in some reports we know that motion paraax can hep human beings in the perception of depth better than just wearing stereo gasses [S]. Simiary, there are aso reports on 3D sound experiments in free-fied conditions [6,7], which says that if a system et a human user move his or her head around in ocating a sound source in space, the precision is better than that of keeping the human head fixed in space. In Poack and Rose s study [S], ow bandwidth and high bandwidth therma noise sound with short duration were used in these experiment, and an improvement of 1 O-I 5% was observed. Therefore, from the resuts, 15

2 we know that dynamic head movement is better for ocating sound, but how much better in a computer generated 3D sound system? In this paper, our first experiment examines the benefits of head movements in sound ocaization, and if so, by what range in degrees. The hypothesis of the first experiment is that dynamic head movement can improve the ocaization of sound in space as compared to that of static head position. Our second experiment focuses on the benefits of reducing the atency in the audio signa that is typicay introduced in virtua environments. The hypothesis of the second experiment is that when a human subject is performing a rea task and there is atency invoved in the system, one s capabiity woud be owered. Since we can use a computer to compensate for atency using prediction agorithms, in a system where there is significant atency, the hypothesis becomes that with atency compensation, human subjects can perform better than that without. In the foowing, we wi introduce the impementation of our software based 3D sound system used in the above two experiments. Experienced readers in 3D sound can skip most of Section D SOUND SYSTEM IMPLEMENTATION 3D sound generay means that a istener hears sounds in a directions, where the sound is simuated by a computer. For a headphone based 3D sound system, the system shoud be abe to pace sounds outside one s head, as we as to the istener s front and rear. In the foowing, we wii introduce the impementation of our headphone based 3D sound system used in our two experiments. This is a rea-time software based 3D sound generation system. Consider the spatia environment as a digita system, one can generate perceptibe audio by means of digita signa processing [9,10,1 I]. If one treat the free-fied spatia environment as a inear, time-invariant digita system, one can understand the behaviour of sound in space by its impuse response. Basicay, the spatia hearing environment has severa factors that contribute to the resut of its impuse response: the azimuth, the eevation, and the position of the sound source, incuding the distance between the sound source and the istener, a can affect the vaue of the impuse response. Furthermore, some factors are human reated, such as the shape of istener s pinnae and canas, the size of the istener s head, the height of the istener s nose, etc. From measurements of the impuse response, one can get the headreated transform functions (HRTF). If the HRTFs are avaiabe, the 3D sound effects can be produced from a inear convoution between sound signas and its corresponding spatia reated impuse response [ 12,131. However, the inear convoution is a time consuming computation. In some virtua environment systems, ike NASA s VIEW system, rea-time convoution is handed by a DSP chip. Because of that PCs computation capabiity becomes more and more powerfu, it is possibe to generate CD quaity 3D audio purey in software on a PC. Our system is consisted of a PC with a Inte Pentium-IUMHz processor, a Creative Lab Sound Baster audio card, and a Pro.2 headphone. By optimization of direct computation of convoution, we can achieve rea time performance. In our design of simuating 3D sound in VEs, a space tracker is hooked to a headphone for reporting the head s position and orientation. Depending on the movement of the space tracker hooked on one s head, the 3D audio payer can seect the nearest impuse response and generate 3D sound accordingy. The HRTFs used in our system are impuse responses obtnincd from the MIT Media Lab. The compacted data are equaized with speaker s impuse response and packed in stereo. The MIT Medin Lab proposed a set of HRTFs measured by Bi Gardner and Keitft Martin in 1994 [14]. They used a dummy head mode, KEMAR mannequin head, as the istener s head. A probe microphone is attached to the position of eardrum, and can record the sound from outside. The impuse response is computed with maximum ength sequence (MLS) technique. A speaker is mounted 1.4 meters From the KEMAR mode. They measured in tota 7 IO different positions at a samping rate of 44.1 KHz with two types of pinnae, ranging from the eevation from -40 degrees to 90 degrees and the azimuth from 0 degree to 360 degrees. Each impuse response is 12%point fength. Then we have to equaize the compact data by an inverse titcr of the headphone. Because the origina compact data used ear cann resonance, we have to remove the ear cana resonance to avoid doube resonance [IS]. An inverse fiter of the headphone can be used to eiminate the effect of the superfuous cana resonance. Gardner and Martin provided a set of measured impuse response for severa combinations of the headphone and pinnae. WC choose the AKGK240 headphone with norma pinnae as our target pairs and compute its inverse fiter to equaize the impuse response, The payback system reproduces the 3D audio output from a mono-audio source by using the 128 point inear convoution with the specified impuse responses (one for eft ear and the other for right ear) continuousy. The mono-audio source is based on puse code moduation (PCM). During payback of 3D sound, changing the ocation of sound source is necessary for our experiments. Moving head in 3D sound environment shoud change the pair of impuse responses to simuate the correspondence of the newy sound source position reated to head orientation. When changing impuse response, there wi be some power gap between two impuse response which cause some cicks in the payback. Appying interpoation can eiminate most of the effect of cick and make payback more smoothy. 3. EXPERIMENT 1: HEAD FIXED AND HEAD MOVEMENT OF 3D SOUND The goa of the experiment is to verify whether dynamic movement of human head can reay improve the ocaization of sound in space. The difference between this experiment and the previous experiments in free-fied conditions is that our sound source is simuated by a computer. The hypothesis of this experiment is that dynamic head movement can improve the ocaization of sound in space as compared to that of static head position, here caed Head Movement Hypothesis. 3.1 EXPERIMENTAL DESIGN Assume that the sound source is fixed and continuous in space, the subjects must point out the direction of sound by a pointer either in the case of fixing one s head or aowed to rotate one s head. For tracking head movements, a magnetic space tracker ( Fock of Bird from Ascension technoogy Co.) was introduced to report the head s position and orientation. Figure 1 shows a subject points out the ocation of sound in space. The degree of azimuth pointed is measured by referring to a arge compass with the projected ine of the pointer on the ground. 16

3 back ambiguity of the fixed head -part of ocaization errors has been eimhated in advance.. The resuts corroborate with our Head Movement H~ pothesis, t(9)=42s, pc.01. Tabe 1. Locaization errors (degrees) of ten subjects of the sound ocating experiment. Figure 1. A photo of our experimenta set-up for the Experiment 1. In the set of impuse responses, the higher eevation has ower precision. From [16] point of view, we know that minimum audibe movement ange (MAMA) at eevation 0 degree is smaer than that of another eevation. Therefore, to be precise, we choose eevation at zero to conduct the foowing two experiments to preserve higher precision [ 171. s For the motion based experiment, the static precision of sound source pay an important roe in conducting the experiments. That is, the position offset between the fixed sound source and the subject s head must be preserved as the same for getting the stabe pair of (evauation, azimuth). However; there is no guarantee that head position is fixed during rotation. The soution is to modify the ocation of sound source with respect to an offset of the difference of current head tracker s position and the initia tracker s position. With different pinnae among subjects, in the idea case, it is better to have different HRTFs for each subject. However, the impuse responses we got were reported from the standard KAMAR mannequin head, that s the reason why we chose this for ater experiments. 3.2 PROCEDURE In the experiment, we invited IO vounteer subjects to participate in our experiments. A subjects were abe to hear 3D sound. Each subject was given two sessions of experiments, the first one required that the subject s head be fixed in space and facing the front. The second aowed the human subject move around the space as he/she wished, of course when one was moving one s head around, a headphone set with space tracker is fixed on one s head. The sound source was aways fixed in space even though the subject can turn his head around. Within each session, 12 directions were seecred in random order. A directions were generated in advance. For one-haf of subjects, the experimenta order was keeping the head fixed first and then aowed dynamic head movement, and for the others, the order was reversed. Each subject was trained to be famiiar with the 3D sound perception for about 5 minutes before the experiment. 3.3 RESULTS Tabe 1 shows the ocaization errors, which are the average error in degrees within tweve trias, of ten subjects. Note that the front- By examining the data for dynamic head movement, the ocaization of sound in space is more precise, that is, the average error in dynamic head movement is degrees whie if the head is fixed in space, the error is about IS degrees: As a resut, we are abe to identify that with dynamic head movement, the human capabiity is enhanced by more than 90% in ocating sound in space. Finay, there are more so caed front-back ambiguity when the human head is fixed in space, since that s the case when one can not identify whether the sound ocates in one s front or in one s back. When the head is fixed in space, on average, there are three front-back errors out of tweve tests caused by confusion, however, in the case of dynamic head movement, there is no such error. As a brief comment, the resut shows that dynamic head movement can not ony hep in the precision of ocating sound in space, but aso hep to reduce the front-back ambiguity. 4. EXPERIMENT 2: LATENCY AND ITS COMPENSATION OF 3D SOUND Considering a typica architecture wakthrough system that can generate 3D sound as we as graphica objects, since the simuated sound has to be- synchronized with the graphics subsystem, whenever there is atency in the graphics subsystem, there is equa amount of atency introduced to the 3D sound system. Therefore, since there is usuay a atency of 300ms in a typica wakthrough system, we wi use the same atency (300 ms) in the ater 3D sound experiment. The hypothesis of the experiment is that when there introduces reativey arge atency in a VE, one s capabiity of ocating sound source woud be significanty reduced. Since we can use a computer to compensate for atency based on prediction agorithms, 0 the hypothesis becomes that with atency compensation. human beings can perform better than those without, and we caed it Latency Compensation Hypothesis. Before conducting the experiment, there is a question to be answered which is What s the argest overa system atency that

4 Tabe 2. A tabe of mean difference between any pair of two cases. The symbo * indicates that the difference of two means is greater than the critica vaue set by Tukev S HSD. wi not affect the perception of 3D sound focaization? The answer is very important since it can hep a VE s designer to decide whether his/her system needs to dea with the atency probem when using 3D sound. 4.1 A PRIORI STUDY ON LATENCY AND RECOGNITION OF 3D SOUND In the priori experiment, we woud ike to find the mythica threshod in the atency just mentioned above. There were five conditions of VEs indicating different atency engths as foows. Condition 1: 0 ms, no atency incuded for reference. Condition 2: 50 ms atency incuded. Condition 3: 100 ms atency incuded. Condition 4: I50 ms atency incuded. Condition 5: 200 ms atency incuded. A sound source trac ng task is designed for the priori experiment, and its experimenta design is described in Section 4.3 in detai. Five subjects took part in the priori experiment. Each subject accepted a of 5 conditions with sequentia order (counterbaanced between subjects) For each condition, randomy seected sound source was put in space seected in random order. When the teiting music (used as sound source) started, the subject was required to find the ocation of sound source by turning his/her head as fast as possibe so that the sound source appears exacty in front of him/her. There were in tota 15 trias examined and the corresponding reaction times of ocation used were reported. Every subject had a training for about 5 minutes. The reaction time of 3D ocaization were tested by a one-way within-subject anaysis of variance (ANOVA). The resut indicated that there exists difference of reaction time of 3D ocaization among those five conditions, F[4,16)=10.8, MSe=0.3 15, pc.01. That is, among ten pairs of these conditions, there is at east one pair that is significanty different. A mutipe comparison test, Tuke.vS HSD, was appied to find whether there is significant difference among five conditions. Tabe 2 shows the resuts of mean difference between any pair of conditions. When atency is ess than 150 ms, there is no significant difference. However, when the atency reaches 200 ms. it showed that there exists difference between condition of 200 ms and any one of conditions with atency ess than 150 ms and thus indicates that the performance in a VE with 200 ms atency is significanty worse than the one without atency in ocating a sound in space. From our observation, if the atency is arger than 150 ms, the performance of ocaization of 3D sound is noticeaby decreased. Since a typica appication such as our architecture wakthrough system suffers from 300 ms atency, finding a way to compensate the atency is obviousy necessary for improving the precision of ocating a sound in space. 4.2 PREDICTION ALGORITHMS FOR LATENCY COMPENSATION 3D sound atency is the time deay between head movement and its corresponding motion of virtua sound source payed on the headphone. To compensate for atency, many proposed methods used prediction in tracking. Severa HMD systems have been impemented with head tracker prediction [ S, 19,20,21], where a ook Bhead agorithm is impemented which uses the 3D position and orientation as the input data. Figure 2 shows the system diagram of a genera prediction system. At each time t, a system behaviour formua can be generated from appying the historica data sequence X (the nearest data observed in time domain) to the prediction agorithm. In Figure 2, the number of historica data is set to i, and therefore the system wi use i observed data as the inputs. Appying a specified prediction ength ro the system behaviour formua, a new predicted data at time f, P, *wi appear. Historica data sequence Prediction Length t Prediction * Agorithm Figure 2. A system diagram of a genera prediction system. Predicted data There are two usefu prediction agorithms, Kaman fitering and Grey system theory, both of which have been evauated to be amost the same in rea task experiments on visua perception [22]. The Grey system based prediction agorithm is used in our experiment because of its ower computation compexity than that of the Kaman fitering based prediction. For detai information of the above two prediction agorithms, pease refer to the origina papers [4,18,19,20, EXPERIMENTAL DESIGN A tracing task was designed to evauate the effects of atency on ocating sound in space, where the sound source was randomy generated in space during the task. When a subject hears 3D sound in space, he/she is required to trace the 3D sound source by facing to the sound source, and when the subject was certain that he/she is exacty facing the sound source, he/she can press the eft button of a mouse to signa his/her decision. If the azimuth degrees betweenhead orientation and the sound source was cose enough (under a given threshod of &5 degrees), the payback of 3D sound stopped. This means that the subject has finished the

5 i tracing task. However, if the azimuth difference in degree exceeds a given threshod, the 3D sound payed continuousy. The ocaization process continued unti the 3D sound stopped. For exampe, in Figure 3, assuming that the 3D sound source is fixed in azimuth ange of 300 degrees, the task wi be finished if the subject faces between 295 to 305 degree and presses button to confirm his/her decision, otherwise, the 3D sound pays continuousy. X (P EmrTokmncz~-5 Figure 3. The tracing task for ocaization of sound in azimuth ange. Task competion time spent in the tracing task was considered as the key parameter to evauate the effects of different vaues of atency. From our previous experience on the visua subsystem, the arger atency introduced in a VE, the more time. it took to finish a task [22]. On the other hand, average error distance may be considered to be a good indicator to evauate the atency effect. However, there is a probem caused from the case that one can move his head sower and spend more time to ocate a sound source more precisey even if arge atency is incuded. Therefore, we sti anayze task competion time in our experiments. The reason that a subject has to expicity press a button to signa his/her possibe competion can be expained in the foowing figure. can identify the sound source in front of him more precisey. That s the reason why we used the technique of facing the sound source in the ocaization task of our experiments. 4.4 PROCEDURE Ten subjects were invited to, participate in the experiment. During the experiment, subjects were instructed to face the sound source by turning their head as fast as possibe. Each subject was given two separate eves in random order, one is with atency compensation and the other is without atency compensation, where the atency is 300 ms. We designed six sets of cases with different initia sound source ocation, and the average task competion time is used for each subject. Since each test aso invoves two conditions, i.e., with and without atency compensation, each subject actuay took I2 trias in tota. The training phase was the same as in Experiment I. 4.5 RESULTS Tabe 3 shows the task competion time used, which is the average time spent in second within tweve trias, by ten subjects. The resut corroborates with our Latency Compensation Hypothesis, t(9)=3.439, pc.01. Tabe 3. Task competion time spent (seconds) of ten subjects of the experiment. 1 Subject 1 without atency with atency compensation compensation S s s I 5.19 s4 I I s I 2.60 S6 t B: endingheadposition Since head swing from one direction (point A) to the opposite direction (point B) is continuous, there exists a point C in the path that wi have a zero vaue in the azimuth difference. However, this does not mean that a subject has exacty ocated the sound source, since he/she is just swinging from one direction to the other trying to ocate the sound source, and so this intermediate zero vaue ij meaningess. According to our observation from Experiment 1, a usefu heuristic can be used: when one tries to ocaize a sound source, one tends to move one s head to minimize the intramura difference. That is, when one hears sound in perfect baance, one Mean I * t(9)=3.439, PC.01 On the average, the time to ocaize a target sound ocation in space with atency compensation is seconds, and the time without atency compensation is about seconds. Therefore, the average task competion time in sound ocaization is about 50% shorter when atency compensation technique is used. Simiary, with atency compensation, on average, the number of times of pressing a button to signa the competion of the task is two, whie without atency compensation, it is 4.5. That is a significant improvement in human computer interaction. For the button pressing part, pease refer to Section 4.3 for detaied expanation. This indicates that with the same prediction agorithm as used in a visua subsystem, compensation for atency in the 3D sound system can aso significanty improve human ocaization of 3D sound. That is, the prediction agorithms such as those based on Grey system or Kaman fitering not ony can reduce the atency in HMD but aso can improve the ocaization of 3D sound. 19

6 5. CONCLUSION AND FUTURE WORK 9. We have conducted two experiments to examine two hyporheses. The first hypothesis is that ocaization of sound in space with head movement is easier than that of keeping head fixed. The second hypothesis is that in a virtua reaity appication with 3D sound capabiity, with atency compensation, a subject in the task of ocating 3D sound performs better than that without atency compensation. Two conducted experiments corroborate with our hypotheses, Note that the resuts of the second experiment are origina. Moreover, a priori study of the atency effects indicates that if the atency is arger than 50ms, 3D sound ocaization performance is noticeaby decreased. We have therefore estabished the simiarity between computer graphics and 3D sound, namey, the effects of motion paraax and atency compensation. ACKNOWLEDGEMENT We woud ike to thank the MIT Media Lab s HRTF measurements, which they have put in WWW for pubic use. We thank detaied comments from anonymous reviewers when we first submitted this paper to SIGCH 97. This project is partiay supported by the grant NSC E REFERENCE s. Kenda, G. S., A 3-D Sound Primer: Directiona Hearing and Stereo Reproducrion, Computer Music Journa, Vo. 19, No. 4, pp.23-46, Winter Hanmann, W. M., Locaization of Sound in Rooms, Journa of the Acoustica.S&iety of America. Vo. 74, No. 5, pp , November Middebrooks, J. C., Green, D. M., Sound Locaization by Human Listeners, Annuaf Review of Psychofof, Vo. 42,pp , Wu, J.-R., Lei, Y.-W., Chen, B.-Y., Ouhyoung, M., User Interface Issues for a Buiding Wakthrough System with Motion Prediction, Proc. of IEEE 1996 Internationa Conference on Consumer Eectronics, pp , Chicago, Kaawsky, R S., Tie Science of Yirtua Reaity and Virtua Environments, Addison-Wesey, Waach H., The Roe of Head Movements and Vestibuar and Visua Cues in Sound Locaization, Journa ofexperimenta Psychbog, V01.27, pp , Thurow, W. R., Manges, J. W., and Runge, P. S., Head Movements During Sound Locaization, The Journa of the Acoustica sociey of America, Vo.42, pp , Poack, I., Rose, M., Effect of Head Movement on the Locaization of Sounds in the Equatoria Pane, Precept. Psvchopiys, Vo. 2, pp.59 I-596, Rodgers, C. A. P., Pinna Transformation and Sound Reproduction, Journa of the Audio Engineering Socie$ Vo. 29, pp , 19s I. 10. Morimoto, M., Ando, Y., On the Simuation of Sound Locaization, in R. W. Gatehouse (Ed.), Locakation of Sound: Theorv and,4&ications. Groton, CT: Amphora Press, * I. Wightman, F. L., Kister, D. J., Headphone Simuation of Fre&ied Listening 1: Stimuus Synthesis, Journa of the Acoustica Socien, of America, Vo. SS, pp , February Fisher, H., Freedman, S. J., The Roe of the Pinnae in Auditory Locaization, Journa of Audito,? Research, Vo. 8, ~~15-26, Wenze, E. M., Wightman, F. L., Foster, S. H., A Virtua Dispay System for Conveying Three-dimensona Acoustic Information, In 32 Annua Meeting of the Human Factors and Ergonomics Society, Santa Monicn: Human Factors and Ergonomics Society, Gardner, B., Martin, K., HRTF Measurements of a KEMAR Dummy-Head Microphone, M/T Media Lab Perceptua Computing Technica Report #780, May Begaut, D. R., 3-D Sound for Virtua Reaity and Mitimedia, Academic Press, Strybe, T. Z., Manigas, C. L., Perrott, D. R., Minimum Audibe Movement Ange as a Function of the Azimuth and Eevation of the Source, Huntan Factors, Vo. 34, pp , Grantham, D. W., Detection and Discrimination of Simuated Motion of Auditory Targets in the Horizonta Pane, Journa of the Acoustica Socie@ of America, Vo, 79, No. 6, pp I 949, June IS. Liang, J., Shaw, C., Green, M., On Tempora-spatia Reaism in the Virma Reaity Environment, Proc. 4th Annua Symposium on User Interface Sofnvare and Technoogy, Hiton Head SC, pp , I99 I. 19. Azuma, R. and Bishop, G., Improving Static and Dynamic Registration in an Optica See-through HMD, gxaph 94 Conference Proceedings, pp , Maz&k, T. And Gervautz M., Two-Step Prediction and Image Defection for Exact Head Tracking in Virtua Environments, Computer Graphics Forum (Eurographics 95), Vo. 14, NO. 3, pp. c30-~41, Wu, J.-R., Ouhyoung, M., Reducing The Latency In Head-Mounted Dispays By a Nove Prediction Method Using Grey System Theory, Computer Graphics Forum (EuroGraphics 94), Vo. 13, NO. 3, pp. c503-~ Wu, J.-R., Ouhyoung, M., A 3D Tracking Experiment on Latency and Its Compensation Methods in Virtua Environments, Proc. of UN 95 (User Interface and Sofware Technooa 1995)). pp , ACM Press, Pittsburgh, 1995., 20