![[Previous Chapter]](/icons/button-back.gif){kind=icon}
![[Table of Contents]](/icons/button-contents.gif){kind=icon}
![[Next Chapter]](/icons/button-next.gif){kind=icon}
| Publications Page | HITL Home |
5.1 Objectives and Hypotheses
Inequality between DFOV angle and GFOV angle often exists in virtual interfaces, due either to software limitations or the inability of the user to properly set this variable. This disparity can be quantified by examining the ratio DFOV/GFOV. Deviations of this ratio from 1.0 result in image scale changes (i.e., magnification or minification) of the visual scene (see Section 2.3.3.3). These scale changes mimic the visual stimuli experienced in traditional VOR gain adaptation studies that require subjects to wear magnifying/minifying spectacles. Therefore, it is reasonable to infer that certain DFOV/GFOV ratios may induce VOR gain adaptations as well as sickness symptoms in users of immersive head-coupled virtual interfaces. The following experiment examined this possibility by systematically manipulating GFOV while holding DFOV constant.
This experiment, however, did not employ a traditional VOR adaptation protocol. Classic VOR adaptation research often entails exposure to a visual-vestibular rearrangement through passive rotation of the subject at specific frequencies with the subject’s head immobilized with respect to the body. The subject would stare at a meaningless visual scene (often random dots or vertical stripes) with no task to perform. This traditional protocol attempts to amplify the effects of the stimulation over a very limited and highly specified set of conditions.
The interesting question of this experiment, however, was not simply whether DFOV/GFOV deviations could drive VOR gain changes but whether they could do so while subjects actively moved their heads in a natural, unrestricted manner, across several frequencies, while they performed a meaningful task in a VE. In other words, would DFOV/GFOV deviations from 1.0 cause VOR gain adaptations in realistic VE user scenarios?
Given the five main objectives of this dissertation identified in Chapter 1, this experiment addressed objectives 1, 2, 3, and 5. Specifically, this experiment was designed to ascertain: 1) if VOR gain adaptation could occur in virtual interfaces that involved natural head-coupled interaction with meaningful visual scenes over a short exposure period, 2) if this adaptation could be directionally modulated through systematic variation of the ratio DFOV/GFOV, 3) if any occurring VOR adaptation was frequency specific or if it generalized across the tested frequencies, 4) if simulator sickness reports were also effected by DFOV/GFOV deviations, and 5) if VOR adaptation covaried with sickness reports. In addition, preliminary data were to be collected on post-exposure re-adaptation time-course information for any occurring VOR adaptation and sickness incidence.
Henceforth, the ratio DFOV/GFOV is termed ‘image scale’ to simplify matters. It was hypothesized that an image scale of 0.5 (corresponding to a 0.5X visual scene magnification, which is thus termed a minification) would reduce the gain of the VOR, an image scale of 2.0 (corresponding to a 2.0X scene magnification) would result in a VOR gain increase and an image scale of 1.0 (i.e., no scene magnification) would result in no VOR gain changes. Given that the VOR can be a noisy signal and that the gain adaptation achieved rarely matches stimulus demand, a threshold for determining adaptation was required. A threshold of 5% change in VOR gain was chosen, partially from a review of short-term VOR adaptation research results, partially from its potential implications on retinal image stability (if not compensated for by visual tracking mechanisms), and partially on ‘scientific intuition’ as to what amount of gain change would be considered meaningful after a 30 minute exposure period. Therefore, the hypotheses presented above specified adaptation as a statistically significant VOR gain change of at least 5% from baseline values.
In addition, it was hypothesized that VOR adaptation magnitude would be the largest at the lowest test frequencies since several VOR adaptation experiments have shown frequency specificity and subjects tend to reduce amplitude and frequency of head movements while immersed in virtual environments (Pausch, et al., 1996). Sickness incidence was hypothesized to be the highest in the scale magnification condition (due to the sensory rearrangement and increased optic flow velocity), less but still significant in the minification condition (due to the sensory rearrangement), and least in the neutral (correctly scaled) condition. Lastly, it was hypothesized that there would be a moderate correlation (approximately 0.40 to 0.60) between sickness reports and VOR adaptation.
A total of 11 adult subjects (6 males/5 females, mean age 28.5, age range 19 to 39) volunteered to participate in this experiment. All subjects reported to be in good health with no history of epilepsy or vestibular medical problems. Subjects were tested for normal or corrected visual acuity of 30/20 or better. Five subjects wore contacts for corrected visual acuity and three of these reported some level of diagnosed astigmatism. Only contacts were acceptable for vision correction during the experiment due to incompatibilities between the eye tracking system and spectacles. Subjects voluntarily abstained from drugs and alcohol for 12 hours prior to participating in each session. Nine subjects reported that they had experienced motion sickness in the past and two subjects reported experiencing simulator sickness in the past. Two subjects claimed to be nonsusceptable to motion sickness, six subjects reported slight susceptibility and three reported moderate susceptibility. At the onset of testing, all subjects were new to the particular virtual environment used in this experiment and had not experienced any head-coupled virtual interface in the previous 30 days (seven subjects had never before been exposed to head-coupled virtual interfaces).
The overall experiment was conducted as a 3 x 2 x 3 within subjects design. Three levels of image scale (SCALE: MIN [0.5X], NEU [1.0X], and MAG [2.0X]) were crossed with two levels of testing time (TESTTIME: PRE-exposure, POST-exposure) and three levels of oscillation frequency (FREQ; 0.2, 0.4, 0.8 Hz). Additionally, VOR measurements were recorded 10 minutes post-exposure on several subjects. These data, obtained only on a subset of subjects, were analyzed separately. Each subject participated in a total of three, 1.5 hour sessions (i.e., one session per SCALE condition). Each session was separated by a minimum of 6 days to minimize carryover effects. Sessions were counterbalanced across SCALE and SUBJECT.
The dependent variables for VOR gain were averaged VOR gain estimates and percent gain adaptation. Percent gain adaptation was a derived variable of adaptation using the formula: ((POST-PRE)/PRE)*100. VOR phase data were also collected, averaged and analyzed.
Sickness reports were also collected before, during, and after VE exposure. These dependent variables included: 1) oral reports of sickness during exposure and 2) SSQ scores. The oral reports were collected at specified times during VE exposure (10, 20, and 30 minutes). Each report was a number on a scale of 0 to 3, in which a ‘0’ indicated no discomfort, feeling fine, ‘1’ indicated slight but noticeable discomfort, ‘2’ indicated moderate discomfort, and ‘3’ indicated very strong discomfort such that the subject wished to take a break or end the experiment. SSQ data were collected pre-exposure, immediately post-exposure, and 20 minutes post-exposure.
In addition, postural stability was measured pre- and post-exposure as a safety check prior to releasing the subject. The metric of postural stability utilized in this experiment was the number of stance breaks by the subject over a 30 second period while in a Sharpened Romberg stance with eyes open.
5.4 Experimental Set-up and Apparatus
This experiment was conducted in the VR Effects Laboratory, a component of the HITL on the University of Washington campus. The apparatus and experimental set-up was as described in Section 4.3. Head and eye position data were collected at 165 Hz. The minimum system time latency (mean: 48 ms; SD: 8 ms) was fixed for this experiment, as was DFOV (25 deg horizontal by 19 deg vertical). Only GFOV was systematically varied. GFOV was set by adjusting the appropriate variable in the WARP TV software.
This experiment required two experimenters, myself and an assistant. I was the main communicator with the subject, controlled all equipment, conducted all eye tracking calibrations and VOR testing, collected sickness data, called out search targets during VE exposure, and managed the conduct and flow of each session. As I was sealed in a light-tight shed during VOR testing and VE exposure, the assistant was by the subject’s side during this time. The assistant aligned the chair, issued mental tasks for the subject to perform during VOR testing, administered the first post-exposure SSQ as well as the controlled post-exposure eye-head movement task, acted as a safety monitor and handled miscellaneous problems that would occasionally arise during VOR testing or the exposure period. A copy of the experimental protocol appears in Appendix B which more explicitly details this teaming arrangement.
5.5.1 Preliminaries, Calibrations, and Baseline Measures
Each experimental session began with a pre-briefing, subject consent (Appendix C), a pre-exposure SSQ (Appendix D), and a test of visual acuity (using a Snellen Chart). Two pre-exposure balance trials were then conducted to obtain baseline values for use in assessing the subject’s post-exposure postural stability (a safety precaution prior to releasing the subject after the experiment). In each balance trial, the subject stood on the lab floor in a Sharpened Romberg position with eyes open for 30 seconds. The number of stance breaks per trial was recorded and averaged across trials for comparison with post-exposure values.
After the balance tests, the subject was seated in the rotating chair and secured with a 5-point safety harness and foot straps. The HMD/Eye tracker unit was positioned on the subject’s head such that the subject felt comfortable and a optimum eye image was obtained by the VOG camera (see Figure 41). In actuality this required some effort because in order to prevent slippage the HMD had to be snug, but not so tight as to cause subject discomfort. Lastly, a chair headrest was positioned to minimize potential decoupling of head from chair during chair oscillations.
Once subject comfort and eye imaging were maximized, the eye tracking system was calibrated for horizontal eye movements. Subjects were instructed sit erect, place his/her head against the headrest (to prevent any head movements), and to only move the eyes towards each of three calibration dots. These dots (located at approximate eye-height on a wall 4.3 meters from the subject) consisted of a center dot directly in front of the subject, along with a left and a right dot which were each horizontally separated from the center dot by 10 deg visual angle from the subject’s position. Calibration was accomplished manually with the aid of the MacEyeball data acquisition system. The results from a minimum of three successful calibration sequences were averaged to obtain the final calibration factor for horizontal eye position data.
After calibration of eye movements, the subject’s VVOR in the real world was measured. The subject oscillated at 0.8 Hz (32 deg/sec peak velocity) while he/she attempted to fixate on the center calibration dot. Two separate trials (8 sinusoid cycles each) were recorded. Upon completion, the room was darkened for dark VOR baseline measurements. Final room darkening consisted of sealing the shed entrance, turning off all lights and lowering occluding blinds around the subject area.
Baseline VOR data were collected at each sinusoidal test frequency (0.2, 0.4, & 0.8 Hz), with a peak velocity of approximately 50 deg/sec. Two trials were collected at each test frequency. For each trial, four sinusoidal cycles were collected at 0.2 Hz, and eight cycles were collected at 0.4 and 0.8 Hz The subject was given mental alerting tasks to perform during the oscillations. The subject was not instructed to fixate on an imaginary point in the distance, as that behavior (EVOR) may involve contributions from a separate visual fixation system (Fuchs, personal communication, 1997).
After collecting baseline VOR data, the subject’s VVOR in the VE was recorded. These virtual VVOR data provided information on the saliency of the VE for stimulating oculomotor adaptation processes. The HMD was turned on so that the subject could view the virtual image (at the specified image scale for that session) and the subject was instructed to fixate on a point within the VE near the center of the display. The subject was then oscillated at 0.8 Hz (approximately 30 deg/sec peak velocity) while he/she maintained fixation on this point in the VE. Data from two trials were collected, with eight sinusoidal cycles recorded per trial.
Upon completion of the virtual VVOR tests, the subject began 30 minutes of active, task-driven interaction with the VE. A total of five different QuickTime VR 360 degree cylindrical images were presented to the subject during the exposure period (approximately six minutes of exposure to each image). For each VE image, the first minute of the exposure was designated as exploration time for the subject to memorize the spatial arrangement of objects within the scene. The subject performed several head rotations during this time in order to view as much of the visual scene as possible. Since all scenes were images from around the Seattle area, subjects enjoyed guessing ‘where they were at’ during this exploration time. The remaining five minutes consisted of a series of visual search tasks. The subject began each search task looking directly ahead. The experimenter then called out a target and the subject responded by finding the target, fixating on it, and verbally identifying that the search was completed. The experimenter verified the response by monitoring a display that mirrored the subject’s visual scene. After verbal confirmation by the experimenter that the target was located, the subject would again look straight ahead and a new search target was identified. The main purpose of the search task was to continually have the subject interact with the VE using active, unrestricted head movements. Therefore, performance metrics on the task were not recorded. After this five-minute search period, the subject was presented with a new VE image. This procedure was repeated until all five images had been presented, resulting in a 30-minute exposure duration.
It is important to note that during the VE exposure period, no specific instructions were given to the subject regarding the speed or type of head movements. The subject moved his/her head as desired to explore the scene and successfully complete the search tasks. In order to gain knowledge on particular head movement patterns, head position epochs (yaw angle only) were collected. These epochs, 80 seconds in length, were collected at the beginning of the VE exposure, halfway through the exposure, and near the end of the exposure (they corresponded with the presentation of the first, third, and fifth VE images).
Simulator sickness data were collected during the exposure period through oral reports. At 10, 20, and 30 minutes, the subject was prompted to report his/her comfort level on a scale of 0 to 3, as described in Section 5.3. If a subject reported a ‘3’, the display was immediately turned off and the subject was prompted to describe his/her symptoms. The subject was also prompted (and often encouraged) to end the experiment. Only those subjects who insisted on continuing were allowed to do so, under more stringent observation and with breaks taken as often as desired by the subject.
5.5.3 Post-Exposure Testing and Re-adaptation Protocol
At the end of the 30-minute exposure period, the subject’s VOR response was again tested at 0.2, 0.4 and 0.8 Hz. The procedure was the same as for the baseline VOR tests. In some cases the HMD had to be adjusted to re-acquire a valid eye image. This readjustment was accounted for by a post-test eye tracking calibration.
At the end of these post-exposure VOR measurements, the room lights were turned on and the subject completed the first post-exposure SSQ (Appendix D). The subject was then asked to remain in the chair so that VOR data could be collected at 10 minutes post-exposure. Between the completion of the SSQ and the ’10 minutes after’ VOR tests, the subject performed controlled eye-head gaze shifts in the real world by searching for letters and numbers on a large white poster paper that was placed approximately one meter in front of him/her on the occluding blind. The subject kept the HMD on during this entire period, viewing the real world through the now transparent display optics. Although the subject remained seated, there was no restriction on head movements. At 10 minutes post-exposure, the room was once again darkened and VOR data were collected in the same manner as before. Lastly, the lights were turned on and a post-calibration of the eye tracking system was accomplished.
At this point the HMD was removed, the chair’s harnesses were released, and the subject got out of the chair. The subject was provided with refreshments (soda and cookies) as he/she completed a general post-test questionnaire (Appendix F). The subject then completed 2 balance trials (in the same manner as pre-exposure) to verify that no obvious balance instability was created by the environment. If any was noted, the subject was required to wait 5 minutes and perform the balance tests again. Lastly, at approximately 20 minutes post-exposure (5 minutes post-release from the chair) the subject filled out a second SSQ. Then, upon assurances that the subject was ‘feeling fine’ with no observable aftereffects, the subject was released.
Specific details regarding the analysis of VOR and VVOR data is presented first. Simulator sickness data and head position analyses are then described.
The following is a summary of: 1) how the MacEyeball Analysis Program (MAP) obtained VOR gain and phase estimations and 2) how the final estimates used in the analyses were determined. A more detailed description of the complete MAP can be found though Demer, et al. (1989) or Demer (1992). This research utilized only a portion of the capabilities of the MAP.
The head and eye position signals were digitally low-pass filtered (8 pole Bessel, 0 to 22 Hz) and differentiated using the two-point central difference method. Large quick phases were removed using conventional velocity and duration criteria. The de-saccading algorithm did not attempt to remove all saccades, only those that were large and well defined. Smaller saccades and artifacts were removed though the elimination of statistical outlying velocity points during the regressions. The data were further analyzed by two separate methods, both of which used Fourier analysis techniques.
In the first method, termed ‘Varant’, the remaining data (after removal of quick phases) were fit cycle-by-cycle to a sinusoidal function using the method of least squares. The first response cycle was always discarded due to the potential presence of transient components. A Fourier transform was then performed at the fundamental frequency. Head velocity, eye velocity, phase and gain were computed for each cycle, and outlying cycles indicating low gain artifacts were automatically removed The gain and phase estimates recorded were the averages of the remaining cycles from that trial. This method also provided an estimate of within-trial variability by calculating VOR data for each valid cycle in a trial.
In the second method, termed ‘Fourier’, the data points removed by the saccadic removal algorithm were replaced by velocity values computed from linear regressions and instantaneous head velocities. Fourier spectral analysis was then performed to compute gain and phase values at each of a range of frequencies. Phase lags were assumed negative, phase leads positive. This analyses was only accepted as reliable if the coherence at the frequency in question exceeded 0.80. Typical coherence values in this research were above 0.98 and nearly all of the trials were above 0.95. Gain and phase estimates at the test (i.e., peak) frequency were computed.
The above describes the automated determination of VOR gain and phase by two different methods. However, a few additional (manual) steps were required to determine the final values recorded first for each trial and then for each cell. These steps are discussed below.
Since MacEyeball calculates gain and phase estimates using two different methods (Fourier and Varant), the final gain and phase values recorded per trial were obtained by averaging these two estimates. These estimates were in most cases within 1 to 3 percent of each other for gain estimates and for phase estimates were within a degree of each other. In addition, two trials were collected for each cell. For cells with 2 valid trials (144 of 162; 89%) the two estimates were averaged to get the final VOR value for that cell. This was the case for most cells, but a few cells had trial values that were deemed unacceptable due to excessive noise/distortion in the eye position data (i.e., lack of coherence between eye and head velocity data as determined by MAP), incomplete data collection, unwanted head movements by the subject, spurious head tracker output, etc. For cells where there was only one acceptable trial (14 of 162; 9%) that trial was used as the final value. For the cells that had invalid values for both trials (2 of 162; 1%), there was no meaningful information for what occurred in that cell (both of these cells occurred during POST tests of the MAG condition). I omitted those 2 cells and their associated PRE cells from the analysis, reducing the N from 27 to 25 for the MAG condition.
In most cases (22/27), the post-exposure eye calibration value was within two percent of the pre-exposure value. However, if the post-exposure value deviated by three percent or more, the post-exposure and ‘after ten-minute’ VOR gain values were adjusted using the appropriate correction ratio. The highest deviation found between pre- and post-exposure calibration values was seven percent (one case).
A note about the experimental design: though the experiment was originally designed to be analyzed using one omnibus ANOVA, I decided to break the design into smaller sub-experiments for the following reasons. The primary question of this experiment was whether or not gain adaptation occurred, pre- vs. post- exposure, for each image scale. The levels of TESTTIME (PRE, POST) were individually meaningless but could be reduced to a single, meaningful ‘difference’ statistic. Additionally, the effects of FREQ, though interesting, were decidedly less pertinent to the focus of this dissertation. It is quite possible, given that subject attrition occurred between sessions (discussed below) along with the small number of subjects involved, that a three-factor between-subjects analysis of variance statistical procedure could mask true effects of adaptation. Therefore, it was deemed more reasonable to consider this experiment as consisting of three separate sub-experiments, one for each level of SCALE. Identification of adaptive effects is made much more directly and clearly with the sub-experiment approach while the risk of increases in family-wise error is no greater, given the multitude of post-hoc analyses required to glean specific findings from the omnibus ANOVA alternative. Finally, this approach to statistical analysis is consistent with much of the VOR adaptation literature. It should also be stated that this decision was made prior to any statistical analyses being performed, and no omnibus procedure was ever attempted.
Since the dependent variables utilized (e.g., VOR gain estimates, percent gain adaptation) were found to be normally distributed with equal variances, parametric statistics were used in the majority of cases. However, when these assumptions were not met, appropriate non-parametric statistics were employed.
Almost all SS data are positively skewed (i.e., not normally distributed). This is true with the data collected in this experiment as well. In addition, the sickness reports during exposure were ordinal, not interval, data. Therefore, only non-parametric tests were used in the analyses which in turn obviates the need for normality or homogeneity of variance testing.
As discussed below, 8 of 9 subjects participated in all sessions. The remaining subject position was filled by 3 different subjects, one for each SCALE. Because of this change, no ‘related’ or ‘paired’ statistics were used. The result is that all tests are correspondingly more conservative than would normally be expected for a design so heavily weighted towards pure repeated measures.
5.6.3 Head Position Data Analysis
Head position epochs (80 seconds each, yaw angle only) were collected at the beginning, middle, and end of each VE exposure session. Using LABVIEW software, each position epoch was passed through a autopower spectrum function to get the power spectral content of the head movements during that epoch. Data recorded from each spectral distribution included peak frequency, peak amplitude (in RMS deg), and the cutoff frequencies at which the power falls below 10 deg RMS and below 5 deg RMS (for estimates of bandwidth). Since the precise conditions under which these epochs were collected were not controlled, only descriptive statistics were calculated.
5.7 Results
Eight subjects successfully completed the three 1.5 hour sessions. Two subjects withdrew after the first session: a female after becoming sick in the MIN condition and a male after the NEU condition, claiming disinterest. Rather than discard their data, one subject (a male) was run only in the MAG condition so that each of the three SCALE conditions would have a within-subjects design with nine subjects each.
This section addresses VVOR data, adaptive changes in VOR response, changes in sickness reports, head position analyses as related to changes in the VOR and sickness reports, the relationship between VOR changes and sickness reports, and re-adaptation data. To facilitate matters, some subsections will be parsed into the relevant sub-experiment involved (MIN, NEU, MAG).
The balance data collected pre- and post-exposure were not analyzed due to the complete lack of variance in the data.
To verify that the VE stimuli were appropriate to promote VOR gain adaptation, average real-world VVOR was compared to average virtual VVOR. These data are shown in Table 2 and graphically in Figure 42 and Figure 43.
|
Condition |
Gain (SD) |
Phase (SD) |
|
Real |
0.94 (0.05) |
-1.1 (2.0) |
|
MIN (VR) |
0.56 (0.04) |
-8.5 (5.5) |
|
NEU (VR) |
0.90 (0.06) |
-14.0 (3.0) |
|
MAG (VR) |
1.45 (0.06) |
-20.3 (4.0) |
The VVOR gain data suggests that each condition provided the expected gain adaptation stimulus. Real-world VVOR and NEU VVOR data were near the perfectly compensatory gain of 1.0. Phase data
indicated an increasing VVOR phase lag (lag is denoted by negative values of phase) with increasing OKN stimulation.A summary of the VOR gain adaptation data across the three sub-experiments is shown in Table 3 and Figure 44. Each sub-experiment’s statistical results are then individually presented.
5.7.2.1 MIN Sub-Experiment
Summary data from the MIN sub-experiment are shown in Table 4. The main focus was to determine if a VOR gain reduction occurred after a 30-minute exposure to a VE with a image scale of 0.5X. Both the PRE and POST data were normally distributed (Shapiro-Wilks) with equal variances (F-test) and the PRE/POST data were significantly correlated (r = 0.77; p > 0.001). A paired t-test (POST - PRE) revealed a significant reduction of gain in the POST tests (t (26) = 6.19; p < 0.0001).
|
Test |
N |
Mean |
Median |
Variance |
SD |
SE |
% Change |
|
Pre |
27 |
0.635 |
0.65 |
0.016 |
0.12 |
0.02 |
|
|
Post |
27 |
0.539 |
0.55 |
0.010 |
0.10 |
0.02 |
-15.1 |
Percent adaptation was also found to be normally distributed (Shapiro-Wilks). A one-way ANOVA indicated no effect of FREQ on percent gain adaptation (Figure 45), and a visual inspection of percent gain adaptation by session indicated no systematic effects of repeated exposure such as learning (Figure 46).
5.7.2.2 NEU Sub-Experiment
Summary data from the NEU sub-experiment are shown in Table 5. Both the PRE and POST data were normally distributed (Shapiro-Wilks) with equal variances (F-test) and the PRE/POST data were significantly correlated (r = 0.85; p > 0.001). A paired t-test (POST - PRE) was nonsignificant (t (26) = -1.18; p > 0.24).
|
Test |
N |
Mean |
Median |
Variance |
SD |
SE |
% Change |
|
Pre |
27 |
0.602 |
0.62 |
0.023 |
0.15 |
0.03 |
|
|
Post |
27 |
0.585 |
0.59 |
0.016 |
0.13 |
0.02 |
-3.0 |
Percent adaptation was also found to be normally distributed (Shapiro-Wilks). A one-way ANOVA indicated no statistically significant effect of FREQ on percent gain adaptation (Figure 47) and a visual inspection of percent gain adaptation by session indicated no systematic effects of repeated exposure.
5.7.2.3 MAG Sub-Experiment
Summary data from the MAG sub-experiment are shown in Table 6. PRE data were normally distributed (Shapiro-Wilks) but the POST data were only marginally so (passing the K-S test but failing the Shapiro-Wilks). Invoking the Central Limit Theorem, a test of normality was conducted on the (POST - PRE) data. These data were normally distributed, allowing for the paired t statistic to be used. The PRE/POST data were significantly correlated (r = 0.68; p > 0.001). A paired t-test (POST - PRE) was significant (t (26) = 2.24; p > 0.04), indicating an increase in VOR gain as a result of VE exposure.
|
Test |
N |
Mean |
Median |
Variance |
SD |
SE |
% Change |
|
Pre |
27 |
0.608 |
0.61 |
0.01 |
0.10 |
0.02 |
|
|
Post |
27 |
0.644 |
0.66 |
0.01 |
0.10 |
0.02 |
+5.9 |
Percent adaptation was also found to be normally distributed (Shapiro-Wilks). A one-way ANOVA indicated no main effect of FREQ on percent gain adaptation (Figure 48), and a visual inspection of percent gain adaptation by session indicated no systematic effects of repeated exposure such as learning, though there was no net adaptation in the third session.
As stated earlier, only nonparametric statistics were used for the sickness data given the lack of normality involved. Two of nine subjects had to terminate a VE exposure session early. These subjects both withdrew from the MIN condition after approximately 15 minutes due to sickness. However, both of these subjects completed post-exposure VOR testing, completed the remaining two SSQs, and one returned for her remaining sessions. However, due to the remaining subject attrition, only between-subject statistics were employed. First SSQ results are presented, followed by the sickness ratings during exposure.
5.7.3.1 SSQ Data
Mean and median SSQ total scores over time are shown in Figure 49. A Kruskal-Wallis test revealed that simulator sickness was induced as a result of VE exposure (collapsed across SCALE), (X2 (2) = 23.07; p < 0.001). Nonparametric Scheffe post-hoc tests indicated that POST1 sickness reports (those collected immediately after VE exposure) were significantly higher than PRE sickness reports (S2 = 4.73; p < 0.001). POST2 reports (collected 20 minutes post-exposure) were also higher than PRE reports (S2 = 3.14; p < 0.01), indicating that sickness aftereffects, though reduced, were still present after 20 minutes.
Mean and median POST1 SSQ total scores by SCALE are presented in Figure 50. A Kruskal-Wallis test indicated a significant effect of SCALE (X2(2) = 5.99; p < 0.05). Nonparametric Scheffe post-hoc tests indicated that sickness reporting was less in the NEU condition than in the altered scale conditions (MIN & MAG) (S2 = 2.45; p < 0.05). There was no statistical difference between the MIN and the MAG condition.
Mean and median POST2 SSQ total scores by SCALE are presented in Figure 51. A Kruskal-Wallis test was not significant but indicated a slight trend towards an effect of SCALE (X2 (2) = 4.29; p < 0.12).
Figure 52 presents POST1 SSQ data across sessions. A Kruskal-Wallis was not significant (X2 (2) = 3.04; p > 0.22), indicating that subject’s reports of sickness were not effected by prior exposures. POST2 SSQ data were also not significantly different across sessions.
Though mean SSQ scores were lower for males than females, a Mann-Whitney nonparametric test indicated that there was no significant effect of Gender on either POST1 SSQ scores (U = 88.5; p > 0.90) or POST2 SSQ scores (U = 91; p = 1.0).
5.7.3.2 Sickness Reports During Exposure
Reported sickness during the exposure were influenced by SCALE (X2 = 9.48; p < 0.009) and a nonparametric Scheffe revealed that reports in the NEU condition were lower than in the deviated conditions (MIN and MAG) (S2 = 3.07; p < 0.01). Figure 53 presents average reports of sickness during exposure by SCALE. In addition, there was a trend for sickness to increase with exposure time (X
2 = 4.85; p < 0.09) (Figure 54). There was no effect of Gender on sickness during (U = 65; p > 0.22).
A total of 68 of 81 potential head position epochs were collected during the experiment (nine subjects, three sessions each, with three epochs per session). Each head position epoch (80 s) was analyzed using LABVIEW software’s auto-power amplitude spectral analysis program to obtain a range of frequencies over which most head movements occurred and the associated power (deg RMS) at those frequencies. A typical example of an epoch’s autopower spectrum is shown in Figure 55.
Four metrics were derived from each epoch: peak amplitude, frequency at which the peak amplitude occurred, frequency at which the power dropped below 10 deg RMS, and frequency at which the power dropped below 5 deg RMS. The last two metrics were rough estimates of bandwidth. Table 7 summarizes data across all conditions.
Table 8 compares the head movement data for each scale condition. There does not appear to be any meaningful difference in head movements across scale. Table 9 presents head movement data collapsed across scale but divided by time of recording (beginning, middle, or end of a VE exposure). Head movements did not seem to change appreciably from the beginning to end of the session. The head data were also no different between sessions.
5.7.5 Gain Adaptation / Sickness Relationship
Correlation tests were performed between percent VOR gain adaptation and each of three sickness variables: final sickness during rating (at 30 minutes), the POST1 SSQ total score, and POST2 SSQ total score. A total of 27 pairs of data were compared (nine subjects with three sessions each) in each correlation. Due to the ordinal nature of the sickness data, Spearman’s rho statistic was chosen over the Pearson statistic (which requires interval data). All correlations were examined using two-tailed tests.
The results of all correlation pairings indicate only a weak association between percent VOR gain adaptation and sickness reports. With POST1 SSQ, r = +0.283 (p < 0.15), with POST2 SSQ, r = +0.24 (p < 0.25), and with the third sickness during rating, r = +0.282 (p < 0.15). Figure 56 shows the scatterplot of percent gain adaptation and final sickness rating during exposure. Note the slight but noticeable positive relationship between these variables. The other two scatterplots also did not reveal any nonlinear relationships.
Figure 56: Scatterplot VOR Adaptation/Sickness During
Gain re-adaptation data are presented separately because not all cells in the PRE-POST analysis contained re-adaptation information and because the question of re-adaptation time-course is only relevant after establishing and detailing the nature of any occurring adaptation. Incomplete data were a result of subjects exiting early due to sickness and cells that had two unacceptable trials at a particular frequency. A total of nine additional data lines were removed from the analysis, leaving 70 analyzable data lines.
Figure 57 shows all VOR gain data (PRE, POST, and AFT10: after 10 minutes) for each SCALE. In the deviated scale conditions, VOR re-adaptation was incomplete after 10 minutes of re-exposure to the real world. Only 20% of the gain change dissipated in the MIN condition and only 50% of the gain change in the MAG condition. The PRE and AFT10 gain data were almost identical in the NEU condition.
Phase data were also collected and analyzed with the results shown in Figure 58, Table 10, and Table 11. A paired t-test revealed that POST phase showed slightly more lag (1.6 deg) than PRE phase, when collapsed across conditions and frequencies (t = 2.37; p > 0.03). Much of this phase adaptation dissipated within 10 minutes post-exposure. The most substantial phase changes occurred in the NEU condition. Also note that the phase change at 0.2 Hz (2.6 deg phase lag) was 80% of the required change at that frequency given the minimum system time delay used in this experiment (48 ms).
5.8 Discussion
This section first addresses the findings regarding the VOR including VVOR, VOR adaptation, and re-adaptation data. Simulator sickness issues are then discussed, followed by the head movement information that was obtained. Lastly, the weak correlation between VOR gain adaptation and simulator sickness is considered.
VVOR gain data indicate that the expected VOR gain direction demand existed in each visual condition. Relative to the real world condition, eye movement amplitude increased in the MAG condition, decreased in the MIN condition, and was essentially unchanged in the NEU condition. This confirms a fundamental premise that magnification and minification effects can be simulated in virtual interfaces.
VVOR gain did not fully match the stimulus demand in the MAG condition however. This finding has also occurred in previous research (Demer, et al., 1987), though an explanation remains elusive. One obvious possibility, that the visual target may have moved outside of the DFOV during the VVOR test, did not occur in this experiment and I would expect that most experimenters would also prevent this error from occurring.
Another possible explanation involves a differential ability to track virtual targets across image scale conditions. The VVOR task in the MIN and MAG condition required a continual sinusoidal displacement of gaze (eye in space) in order to perfectly track the virtual target. Given that the amplitude of the gaze stimulus varied with SCALE and that head movements were identical across all conditions, differing gaze requirements brought about by the different scale factors must be accomplished solely through altered eye movements. Eye movements have vestibular (VOR) and visual (OKN, smooth pursuit, saccadic) components, though it is reasonable to assume that during these initial VVOR tests the visual components provided the majority of the additional compensation in the MIN and MAG conditions. Pursuit movements likely dominate the compensation due to the nature of the task. Failures in smooth pursuit need to be corrected for by saccadic movements towards the target. For instance, a lag in tracking (which would most likely occur in the MAG condition) would result in corrective saccades to ‘catch up’ to the target. Since these saccades were eliminated from the eye velocity data automatically in this experiment, the remaining slow phase eye movement velocity was reduced below that required. Therefore, incomplete compensation for gain demand occurred and this effect was more pronounced in the upward direction (i.e., with increasing optic flow).
The VVOR phase data support this interpretation. In the NEU condition, the obtained phase delay was as predicted by the minimum time delay in the experiment (at 48 ms time delay, the theorized lag at 0.8 Hz is 13.8 deg and the actual phase lag of eye movements averaged 14.0 deg). However, there was less phase lag in the MIN condition and more phase lag in the MAG condition, even though WARP TV responds a bit more quickly in the MAG condition than the MIN condition (by approximately 5-8 ms). Previous research on smooth pursuit explains this outcome (Matin, 1986). Pursuit movements often begin to lag appreciably behind the stimulus at around 40 deg/s. Though the VVOR tests were performed at approximately 30 deg/s, in the MAG condition the optic flow moved at 60 deg/s due to the 2X magnification. At such speeds, there is often an inability to perfectly track the target combined with an increasing phase lag, both of which occurred in this experiment. Therefore, this increased phase delay in the MAG condition indicates difficulty in maintaining perfect pursuit. In the MIN condition, however, it is easier to maintain fixation because the target velocity is reduced in response to the head motion to approximately 15 deg/s. It is possible that if subjects were given a period of practice on the tracking task, the resulting gain and phase would approach calculated values (due to the activation of prediction mechanisms).
This experiment demonstrated that VOR gain adaptation (and to a small extent phase adaptation) can occur during natural interaction with head-coupled virtual environments, confirming Hypothesis 1. This is quite an achievement for the oculomotor system given the method and duration of exposure. Subjects moved their heads in a natural, unrestricted fashion as they performed visual search activities for 30 minutes. At no time were the subjects instructed to move their heads more quickly or at specific frequencies; in fact the power spectral distributions of head movements clearly indicate that the largest concentration of head movement power was at or below 0.2 Hz (in most cases below 0.1 Hz). However, VOR tests conducted at only three pre-defined frequencies captured the existence of VOR gain adaptation in the MIN and MAG conditions. Active head movements during the exposure period may have supported VOR gain adaptation (Demer, Oas, & Baloh, 1993).
VOR gain changes were not fully compensatory for stimulus demand. This was expected given previous research (Collewijn, et al., 1983). The adapting mechanism may have limits that have been ecologically determined over the millennia, and seldom have there been such artificial visual-vestibular demands occurring in nature. Smaller demands, which are more likely to appear inadvertently in VEs, are much more likely to be fully compensated for in a short period of time (Collewijn, et al., 1983).
VOR gain adaptation direction and magnitude were clearly influenced by changing GFOV. This is an important finding because virtual interface systems are often designed with little or no regard to the proper setting of this variable. There are potentially many virtual interface systems that are inadvertently stimulating VOR gain adaptation through DFOV/GFOVs that do not equal 1.0. To reduce the potential for unwanted oculomotor adaptation, designers should attempt to equate GFOV with the DFOV used, as in the NEU condition where no significant adaptation took place. If that is not possible, it is better to error on the side of image magnification, as these results indicate that gain adaptation may be less modifiable in the upward direction.
Why was VOR gain more readily reduced than increased? There are five potential explanations: 1) VOR adaptation processes have direction asymmetries, 2) influence of spectacles to correct for myopia, 3) influence of the system time delay, 4) perceptual effects of visual displays, and 5) subject fatigue. These explanations are clarified below.
Past VOR adaptation studies have indicated that VOR adaptation processes may be biased towards gain reduction (Bello, Paige, & Highstein, 1991). This could be due to evolutionary conditioning, though it seems more intuitive that VOR gain increases would be likely to be demanded naturally due to the effects of disease, trauma, or age. Second, myopics who wear corrective spectacles are as a result pre-conditioned to reduce their VOR gain (Cannon, et al., 1985). One diopter of visual correction in spectacle optics results in approximately a 2-3% VOR gain change demand (Cannon et al., 1985; Collewijn, et al., 1983). Since the two subjects who wore spectacles for visual correction were both myopic, there may have been an increased propensity, based upon past experience, for those subjects to reduce VOR gain. The third potential explanation, further discussed in Section 6.8, is that the minimum time delay inherent in the virtual interface produced a gain reduction stimulus across conditions. This time delay may have further reduced the gain in the MIN while inhibiting gain increases in the MAG condition. Fourth, previous research suggests there is often a perceived minimization of a spatial environment when viewed through a HMD (Roscoe, 1991). Roscoe argued that this minimization of the virtual image is due in part to ‘positive miss-accommodation’ of the eyes, i.e., the eyes do not focus at infinity when viewing collimated virtual images, but rather lapse towards dark focus (which is around an arm’s length). This perceptual minification could contribute to the resulting VOR gain adaptation downward bias through cognitive inputs to adaptation processes. Lastly, subject fatigue may have played a part in driving VOR gain downward over the three conditions (Demer, et al., 1987).
As a general note, it is not image scale per se that stimulated VOR gain adaptation but relative velocity of optic flow in response to head movements in concert with other hypothesized factors (Shelhamer, et al., 1994). As discussed earlier, changing image scale is simply one direct way to modulate optic flow velocity with regard to head movements. Another way would be to decrease or increase the gain on head tracker rotations. Reducing rotational gain on a head tracking sensor would result in less OKN in the visual response to head motion (a MIN condition) while increasing the gain would result in increasing OKN (a MAG condition). In either case, it is the amount of retinal slip detected during head movements that contributes to the error signal. Thus, additional design guidance for the reduction of unwanted oculomotor adaptation must include the importance of accurately calibrating the head tracker’s recording of rotational head motion.
When collapsed across SCALE, VOR gain adaptation did not appear to systematically change over the three sessions. This implies that the session separation was adequate to prevent long term sensory-motor learning between sessions. It would be interesting to examine the effects of more tightly spaced sessions, as some research indicates that VOR gain adaptation carryover effects can occur (Gonshor & Melvill Jones, 1976a). Multiple exposures within a limited period of time could result in the long term storage of gain change parameters for that stimulus (Shelhamer, et al., 1992). This has been termed ‘adaptation set’ by Parker (personal communication, 1997).
Lastly, a statistically significant phase lag adaptation occurred in this experiment. The phase lag increased by 1.6 deg as a result of the 30 minute exposure to the VE (collapsed across conditions). The most significant phase lag adaptation (2.6 deg) occurred at 0.2 Hz, with less adaptation occurring at the higher test frequencies. This could be interpreted as minor support for frequency specificity in phase adaptation, especially given that the phase lag change at 0.2 Hz was nearly fully compensatory for phase change demand (3.2 deg at 0.2 Hz). However, given the small adaptation magnitudes involved, it seems a somewhat tenuous conclusion.
5.8.3 Frequency-Specific vs. Generalized VOR Gain Adaptation
An interesting finding resulted from the comparison of VOR gain change with test frequency. There has been some debate in the VOR literature on the issue of frequency specificity and VOR adaptation. Many researchers have argued that VOR adaptation is frequency specific (i.e., tuned to the peak frequency experienced during the stimulus exposure period) while others have found that it may be more generalized across head movement frequencies. These data support generalized adaptation across frequencies, given that VOR gain changes did not vary across test frequency while the head movements performed by subjects in the VE had peak power below 0.2 Hz.
These results suggest the importance of active, unrestricted head movements in generalized adaptation of VOR gain. The concept of frequency specificity was developed using passive head oscillations at one set frequency (Lisberger, et al., 1983). It is quite possible that more natural head movements provide an effective stimulus for the adaptation processes to extrapolate gain compensation beyond dominant frequencies of motion during exposure. This issue is further discussed in Chapters 6 and 9.
The results of this experiment suggest that VOR gain re-adaptation following VE exposure is incomplete after 10 minutes. Though gain values recovered 20% to 50% of changes within the 10 minute period, the gains were still off pre-exposure levels. This finding was not expected, given that adaptation processes often follow a decaying exponential with the majority of adaptation (or in this case, re-adaptation) occurring quickly (Collewijn, et al., 1983; Welch, 1986). This, combined with the magnitude of gain changes experienced in this experiment and the hypothesized existence of adaptation sets, all point to more complete re-adaptation after 10 minutes. Some explanations for the obtained results follow.
One explanation involves the influence of subject fatigue or lack of alertness. In the MAG condition, 50% of the gain change was recovered in 10 minutes whereas only 20% percent of the gain change was recovered in the MIN condition over the same time period. Given the observation by experimenters that subjects quickly grew tired of the post-exposure controlled eye-head movement task and often showed signs of boredom during this recovery period, perhaps this contributed to the directional recovery asymmetry. Since the gain levels after 10 minutes recovery from the MAG condition were within 3% of pre-exposure levels, the remaining difference could very well be within the signal variance of the reflex.
A second explanation involves a lack of interaction with the real world. Since subjects were still strapped in the chair and looking through the HMD during this period, full interaction with the real world was not possible by the subject. The remaining contextual information provided by the chair and HMD may have contributed towards either a decrease in rate of re-adaptation or perhaps a minor relapse towards the post-exposure gain levels once the lights were turned off. This relapse is not unheard of in the literature (Gauthier & Robinson, 1975; Shelhamer, et al., 1992).
The overall findings regarding simulator sickness are discussed, followed by the influence of uncontrolled factors, re-adaptation, and miscellaneous issues.
5.8.5.1 Overall Findings
This experiment induced significant levels of simulator sickness as a result of a 30 minute exposure to the virtual interface. However, the magnitude of simulator sickness was related to the image scale of the VE such that the NEU condition was approximately half as provocative as either the MIN or MAG condition. The data were consistent whether the metric examined was sickness reports during, SSQ total score immediately post-exposure, or SSQ total score 20 minutes post-exposure. This finding also reinforced what the experimenters had informally noticed throughout the experiment: the NEU condition least adversely effected the health of the subjects.
These results support the sensory rearrangement theory of motion sickness which argues that sensory rearrangements provoke the occurrence of simulator sickness. Both the MAG and MIN conditions had visual-vestibular sensory rearrangements involving visual motion velocity in response to head movement velocity, whereas the NEU condition did not (save the minimum time delay in the system which was shared across all conditions).
Other theories of motion sickness were not supported by these data. The emerging subjective vertical theory (de Graf, Bles, & Groen, 1997; de Graf & Bos, 1997; Bles, de Graf, Bos, & Groen, 1997) is a modification of the sensory rearrangement theory. It argues that only those sensory rearrangements that stimulate otolith responses are provocative. However, this experiment involved very few, small-amplitude head movements in pitch and none in roll, yet significant simulator sickness was elicited. Furthermore, the subjective vertical theory cannot explain the differential sickness experienced across conditions, given that the tasking across exposures were identical.
The proponents of an optic flow model of simulator sickness will also be disappointed by the outcome of this experiment. These researchers hypothesize that increased optic flow is a prime factor in the generation of sickness symptoms. However, both the MIN and MAG conditions resulted in similar amounts of sickness even though the optic flow differed by a factor of four between these conditions.
Lastly, though this experiment does not specifically address the postural control theory of motion sickness, it is evident that although the subjects were firmly strapped into a chair using a five-point harness and Velcro straps for the feet, simulator sickness symptoms were still elicited.
5.8.5.2. Other Influencing Factors
It is important to note that significant simulator sickness was reported in the NEU condition, only less so (by a factor of two) than in the MIN and MAG condition. This indicates that there are other provocative factors of motion sickness that were present but not controlled for in this experiment. A prime suspect, system time delay, is investigated in Chapter 6. Other potential contributors include accommodation-vergence mismatch, reduced resolution, display flicker, optical distortions, and form and fit of the HMD.
In fact, those who totally discount the potential contribution of poor HMD fit on SSQ ratings have probably been out of an HMD for too long. Current technology HMDs can be heavy, hot and have pressure points where they contact the surface of the head. Though the specific HMD used in these experiments was the lightest and most comfortable on the market, there was likely a contribution of poor HMD fit on SSQ results (most likely in the ‘headache’ category). However, this contribution was stable across sessions and does not account for the obtained variance in sickness reporting.
5.8.5.3 Simulator Sickness Recovery
Simulator sickness was still present in the MIN and MAG condition after 20 minutes post-exposure, though its magnitude was decreased. Evidently the time constant of re-adaptation is not much shorter than the time constant of adaptation, at least for this virtual interface-tasking combination. The sickness stemming from the NEU condition, however, was completely absent after 20 minutes. This further indicates that GFOV setting is an important modulating factor in the occurrence of simulator sickness.
5.8.5.4 Miscellaneous Issues
Two more issues merit discussion. First, simulator sickness increased throughout the exposure period. This corresponds with other research, indicating a build-up of effect over time (Regan, 1995). This also correlates with ataxic effects experienced in the postural ataxia studies (Chapter 4) which also increased over exposure time. Though data were not collected on VOR adaptation onset time-course, such data would allow for time-course comparisons between these two processes. Second, there was no effect of gender on simulator sickness incidence. This is most readily explained by the small sample size involved (six males and five females), as there is a preponderance of data in the literature supporting the claim that females are more susceptible than males. Even in this study, the only subject to discontinue due to sickness was a female.
The most striking aspect of the head movement data was the lack of variance involved. Subjects did not appreciably change head movement frequencies or amplitudes with the changing scale conditions. Evidently, the task involved was the major determinant of head motion and it remained consistent across conditions. Subjects were required to search for a target, fixate on it, then return to the starting head position. There was no time pressure on the subject to find objects as quickly as possible so subjects often moved at a leisurely pace. Also, it is important to keep in mind that changing the image scale does not change the angular movement required to center each target in the display. Therefore, subjects were required to make similar head movements during the exposure periods regardless of the scale condition, which may also account for the lack of variance obtained.
However, a common finding in simulator and motion sickness is that head movements are often curtailed when subjects begin to experience simulator sickness. This would predict a reduced set of head movements in the MIN and MAG condition versus the NEU condition, which did not occur. In addition, there was no effect of exposure time on head movement amplitude or frequency, which does not correspond with the increase in simulators sickness reported with time exposed. There are two possible explanations for this discrepancy. The most likely reason is that the power of head motion was concentrated at such low frequencies due to task requirements (discussed further below) that a floor effect existed. Subjects may not have been able to move their heads more slowly and still complete the search tasks in a reasonable length of time. Another explanation points to the lack of control surrounding the collection of these data. There was no rigorous attempt to standardize the collection periods across subjects other than to test each at the beginning, middle, and end of the 30 min exposure period. Subjects were not given the exact same items to look for in the exact same order, for instance. This potentially allowed extra variance in the data that could mask small effects, though the 80 second duration of data collection probably minimized these small differences between subjects.
These data indicate surprisingly low peak head movement frequencies. As stated above, this can partially be explained by a lack of a time constraint in task performance. In addition, during each 80 second epoch, there were several periods of time where the subject did not move his/her head at all (i.e., while awaiting a new target to search for or for confirmation that a target was correct). This stationary head data served to pull the peak power frequency towards 0 Hz. Finally, virtual interfaces have been shown to decrease head movements of VR participants (Pausch, et al., 1996).
5.8.7 VOR - Sickness Relationship
At first glance there appeared to be a rather significant relationship between VOR gain adaptation and simulator sickness. VOR gained changed significantly in the MIN and MAG conditions but not in the NEU condition. Simulator sickness was significantly worse in the MAG and MIN conditions versus the NEU condition. However, the correlations between VOR gain change and simulator sickness metrics were never larger than 0.30. This implies that maximally only 9% of the variance in the simulator sickness data could be explained by VOR gain changes. This low correlation confirms some previous research (Bouyer & Watt, 1996; Watt, 1987) but it refutes the findings of a relationship between VOR adaptation and sickness found by others (Gordon et al., 1996). It would seem that Gordon’s work may be a special case; it involves real motion (on seafaring ships) and vestibular habituation vs. VOR adaptation per se. This issue is explored further in the next experiment.