Preventing Imminent Collisions between Co-LocatedUsers in HMD-Based VR in Non-Shared Scenarios

Iana PodkosovaTU Wien

yana.podkosova@tuwien.ac.at

Hannes KaufmannTU Wien

hannes.kaufmann@tuwien.ac.at

AbstractThis paper presents two experiments set in amulti-user HMD-based VR system where usersnavigate by real walking in a large real and vir-tual area. We investigate a case that could beused in a multi-user VR game or a training ap-plication: several users are walking in the samephysical space without seeing each other in thevirtual environment. Such a scenario involvesthe risk of collisions between users. In the firstexperiment, we investigate the strategy of stop-ping a walking user in a dangerous situation. Inparticular, we compare the effectiveness and theperceived difficulty of two visual and two audi-tory stopping signals. The results of this com-parison show that the tested visual and auditorysignals are equally effective in stopping users.With both visual and auditory signals, partici-pants prefer the signal to contain a ”stop” com-mand. In the second experiment, avatars aredisplayed at users’ positions if the distance be-tween users is dangerously small. The methodis tested with four avatars of various degrees ofanthropomorphism and in two different appli-cation scenarios. Our results suggest that thetype of scenario influences users’ preference ofa notification avatar. It is sufficient to displayan area occupied by other users in scenarioswith specific goals and interactive content. Ifusers are exploring a virtual world without hav-ing any other goal, they prefer to see human-like avatars as a possible collision notification.

Keywords: immersive virtual environments,multi-user Virtual Reality, collision prevention

1 Introduction

In recent years, HMD-based VR systems havebeen developed that allow their users to explorevirtual environments (VEs) by physically walk-ing in tracked spaces of large scale, of the sizeof a gym [1, 2], or even a football field [3]. In-deed, navigation by real walking has been ex-tensively studied and proved to be beneficial foruser experience. It has been shown to improveusers’ immersion [4] and to positively contributeto task performance [5, 6]. A number of VRapplications require a multi-user setup in whichnavigation by walking is possible [2]. Systemshave been developed that satisfy these require-ments [1, 2, 7]. Our research concerns preciselythese systems: multi-user, where users are fullyimmersed in a VE with HMDs and can walk oreven run in a large physical area.

When building an application based on alarge-scale setup with co-located users, scenar-ios are possible where VEs or at least parts ofthem are not shared by users. An example is alarge collaborative scenario that takes place in abuilding with multiple floors. Such scenario canbe used in a game or in a team rescue trainingapplication, for example for fire-fighters. Play-ers start the game on the ground floor and spreadout in different directions to explore the virtualbuilding. Some of them go to different floorsand therefore are no longer visible to the play-ers who stayed on the ground floor. Meanwhile,all players stay within borders of the same track-ing space. Since players do not see or hear whatis happening in the real space around them [8],there emerges a danger that they walk or run intoeach other.

CASA 2017 37

We focus on situations where walking play-ers are dangerously close to each other and acollision is likely to occur within a short pe-riod of time. In such circumstances, a robustmethod is needed that would allow to preventan imminent collision. One possible way to pre-vent an immediately approaching collision is tostop involved players. However, having to stopsuddenly during the VR exposure can provokebreaks in presence and negatively affect user ex-perience. Therefore, we suggest an alternativestrategy: to allow the involved users avoid a col-lision on their own by temporarily displayingavatars at users’ positions.

The effectiveness and plausibility of neitherof these two approaches to the imminent col-lision prevention has been investigated before.This paper reports two experiments set up to ex-plore the potential of both strategies. In Exper-iment 1, we use four different notification sig-nals to stop users walking in VR. We then com-pare the signals in terms of their effectivenessand perceived difficulty. In Experiment 2, weuse avatars to notify two users simultaneouslyexploring a VE about a possible collision at anear distance. We investigate how the visual ap-pearance of notification avatars affects user ex-perience. Four avatars reflecting various degreesof anthropomorphism are used to test our mainrationale: a notification avatar should be distinctenough to allow successful collision avoidanceyet it should not disrupt users’ experience of in-dividually exploring a virtual world.

The contribution of this work is the first toour knowledge experiment on imminent colli-sion prevention conducted with pairs of walkingin VR users. Its findings can be helpful for thedevelopment of multi-user VR applications withnon-shared VEs.

2 Related Work

2.1 Collisions in VR

A large body of research on collisions in VR in-vestigates collisions with static virtual objects.Blom and Beckhaus investigate the influence ofthe collision response on the perception of a VE[9]. Their results show that the stop handlingmethod, coupled with feedback fitting the colli-sion context, positively influences the perceived

realism of collisions. Based on this work, Al-fonso and Beckhaus suggest an aural notificationmethod to prevent collisions with virtual geome-try. In their study, constant directed sound feed-back notifies users of the proximity of the wallsin a virtual maze. Provided with such notifica-tion feedback, users collided with walls less fre-quently than without it. Both studies were per-formed in an immersive projective display (IPS)system with a wand used to navigate through themaze. However, virtual trajectories of users nav-igating in a VE highly depend on input and out-put devices [10]. Thus, a combination of a wandas navigation input and an IPS as an output de-vice is likely to produce different results than acombination of real walking and an HMD.

Several studies compare collision avoidancebehaviour of humans physically walking in thereal world and in VEs. The results show adecrease of walking speed and an increase ofthe clearance distance when avoiding a collisionwith a static virtual object compared to its realcounterpart, tested in an IPS environment [11]and with an HMD [12]. Furthermore, there isevidence of users keeping more distance to ananthropomorphic obstacle than to an inanimateobject [11]. However, situations where physi-cally walking users avoid moving obstacles orother users in immersive VR have not been stud-ied. Research of Olivier et al. on pairwise inter-actions between walkers in the real world showsthat walkers adapt their movement trajectoriesonly when required [13]. It is proposed thatwalkers’ collision avoidance behaviour can bedescribed by a mutual function of their statesnoted as minimal predicted distance (the an-ticipated crossing distance at every moment oftime). In the carried out user study, participantsadapted their walking trajectories to avoid a pos-sible collision only when the minimal predicteddistance was lower than 1m. There is evidencethat users are also able to predict whether theywould collide with walking virtual humans ornot and correctly choose their collision avoid-ance strategy [14], the study however being con-ducted in a desktop setup.

Situations with two users simultaneouslywalking in the same physical space while be-ing immersed with HMDs in individual, non-shared VEs are addressed only by a few stud-ies. An experiment with two users walking side-

CASA 2017 38

Figure 1: From left to right: the stop sign used in Exp.1, the photo-realistic human avatar used as thefigure signal in Exp.1 and Human avatar in Exp.2, Spaceman, Robot and Shape avatars usedin Exp.2, a pair of users in full gear during the Game task of Exp.2.

by-side with HMDs and headphones investigateswhether users who do not see and hear their testpartners in a VE could notice their proximity inthe real world [8]. The results show that usersdo not notice each other at distances as smallas 1m, especially if they are not actively tryingto localize each other. Another work presents acollision prevention algorithm for several userswalking in an immersive VE [15]. The algorithmis developed for a specific case where users arebeing steered away from the walls in the realroom by a redirected walking method [16]. Themethod consists of two parts: collision predic-tion based on the redirected walking assump-tion and collision avoidance. Collision avoid-ance is performed by either steering users awayfrom each other or by stopping them to avoid animminent collision. The method is tested in asimulation on recorded paths of individual userswalking in the tracking laboratory. Research oncollision avoidance strategies used by walkers inthe real world shows that collision avoidance isperformed collaboratively and includes complexinteraction between subjects [17]. Mutual col-lision avoidance behaviour of users walking ina VE has not been tested. However, it is likelyto be based on similar principles. Therefore, webelieve that it is crucial for an imminent collisionprevention strategy to be tested with real walkingusers and not in a simulation.

2.2 Stopping Walking Users in VR

As to our knowledge, no previous studies inHMD-based VR investigate directly the task ofstopping walking users. The aforementioned re-search on mutual proximity awareness [8] in-volves situations where users are stopped, withthe use of a stop sign rendered in front of them.The effectiveness of this method is not reported.

However, users are walking at controlled andlow speeds of 0.3 m/s and 0.7 m/s. It is to ex-pect that users are able to stop easily in such testconditions. Another work on non-shared VEssuggests to stop users when necessary by blank-ing their HMDs and displaying the word ”stop”[15]. The approach assumes the reaction time of1 sec, however only used in a simulation. Nodifferent signals for stopping users are tested inany of the above work. Neither is it investigatedhow fast users can actually stop while walkingin VR. Our intention is to fill this gap in Exper-iment 1 where we separately test two visual andtwo auditory stopping signals.

2.3 Avatars in Immersive VR

Virtual humans can be used as agents steered byartificial intelligence and as avatars representingusers in VEs. Effects of both agents’ and avatars’appearance and responsiveness on presence, vir-tual body ownership and co-presence are wellstudied in immersive VR as well as in collabora-tive virtual spaces. Here, we only discuss workthat is especially relevant to our research.

More elaborate avatars have higher successrate in social encounters [18]. Similarly, agentswith higher responsiveness contribute to a higherdegree of co-presence, or social presence per-ceived by users [19]. However, avatar realismdoes not seem to have a strong impact on the il-lusion of the virtual body ownership provided animmersive VE and full freedom of movement, inan experiment comparing a photo-realistic hu-man avatar with a cartoon-like and a machine-like anthropomorphic avatars [20].

In our notification method of Experiment 2,the purpose of avatars is to merely make col-lision avoidance easy for users without invit-ing them to interact or inducing the sense of

CASA 2017 39

social presence. A user seeing an appearingavatar should be able to remember that it repre-sents another user and colliding with it should beavoided at any time. Ideally, the sense of beingalone in the VE should not be lost even whenusers actively avoid collisions with each other.Following the approach presented in [20], weuse the degree of anthropomorphism as a classi-fication for notification avatars to test the aboverequirements.

3 System and Method

3.1 Experiment 1: Stopping Users

The aim of this experiment is to test whether atype of signal used to stop a user walking in VRhas effect on how quickly the user stops. Ad-ditionally, we collect information on perceiveddifficulty of signals. This experiment has 4 × 1within subject design, the stopping signal beingthe within group factor.

3.1.1 Stopping Signals

Two visual and two auditory signals are tested.The visual signals are a human figure and a stopsign depicted in Figure 1. The auditory signalsare a voice staying ”stop” and a short clap sound.

The motivation behind using a stop sign anda stop sound is to give a short and simple com-mand known to everybody. In a casual situation,it is natural to stop a walking person by talkingto them. However, the visual system is likely todominate other senses during a VR experience.We test a visual and an auditory stop commandseparately to investigate this reasoning. The fig-ure of a virtual human used as the second vi-sual signal is meant to remind of another userco-located in the shared tracking area. We as-sume that seeing such a reminder could makestopping easier. The position at which the fig-ure is displayed should identify the location of apossible collision. Finally, the clap sound is usedas a comparison for the auditory signal contain-ing the stop command semantics (stop sound).

The position of the human figure is fixed inthe VE whereas the stop sign is always shown at1.5m distance in front of a user. This way, it isclearly visible even when the user keeps walkingfor some time after receiving the stopping signal.Both sound signals are non-spatialized.

Figure 2: The VEs used in Exp.1 (top) and Exp.2(top for Walk and bottom for Game).

3.1.2 Environment and Tasks

The used VE is shown in Figure 2. The VE isdesigned to look appealing enough to not dis-appoint participants but at the same time to besimple and non-distracting. The size of the VEis 8x11 m, corridor width is 1.2m.

The main task is to walk in a virtual corridor.Participants are instructed to stop as quickly aspossible if they see objects appearing in front ofthem or hear a sound. Both sounds were playedfor each participant before the experiment. Thetask is performed by one participant at a time.They are asked to walk at a comfortable pace andto not slow down in the anticipation of a signal.

The signals are invoked by invisible triggersplaced at different locations in the middle ofthe virtual corridor where participants are walk-ing. When a participant reaches a pre-defineddistance towards each trigger, a signal is given.Thus, the triggers represent a second user thecollision with whom would need to be preventedby stopping the walking participant.

Each participant performed two walk-throughsequences with all four stopping signals beinggiven in each of them. The order of the signalswas different between the sequences. The dis-tance at which the triggers invoked the stoppingsignals was 2m in the first walk-through and 1m in the second walk-through. Effectively, thisdifference only affects the human figure signal.This signal is implemented by displaying the fig-ure at the location of the corresponding trigger,at 2m distance in front of a participant in the firstwalk-through and at 1m in the second one.

CASA 2017 40

3.1.3 Measures

The main quantitative measure is the distancethat a user walked before the full stop after astopping signal had been issued (further ”stop-ping distance”). The stopping distance is mea-sured as a difference between the tracked posi-tion of a user when the stopping signal is givenand the position of the user registered when theuser have fully stopped. Subjective data on thesignal perception is also collected. In a post-testquestionnaire, participants rate how difficult itwas to stop following each signal. We call thismeasure ”stopping difficulty”. Participants alsoindicated their preferred signal, if any.

3.2 Experiment 2: Displaying Avatars

The goal of this experiment is to explore thestrategy of displaying notification avatars atusers’ mutual positions if the distance betweenwalking users is dangerously small. This exper-iment has 4 × 2 mixed design, with the withinfactor being the notification avatar and the be-tween factor the type of the VE. The experimentis conducted with two participants at a time.

3.2.1 Avatars

Avoiding collisions with others while walkingis a natural task daily performed by everybody.A human figure displayed at the position of anapproaching co-located user would probably al-low to avoid a collision in a most intuitive way.However, a human avatar is likely to producean illusion of another person being in the sameVE. A different solution would be to use a verygeneric non-human shape that would only iden-tify a currently occupied volume in the VE. Suchan avatar is unlikely to induce a feeling of theVE being shared with somebody else, althoughit would be less familiar to users and might causeconfusion. To explore the range between thesetwo limit cases, we introduce four levels of an-thropomorphism and realism of the avatar, rang-ing from a photo-realistic human to a completelynon-human shape.

We use two human avatars: a photo-realisticavatar from Experiment 1 (further ”Human”)and a detailed but cartoon-styled and less re-alistic avatar (”Spaceman”). A lower level ofavatar anthropomorphism is represented by a

robot (”Robot”), obviously non-human, facelessand having simplified geometry. Finally, we usea semi-transparent cylindrical shape (”Shape”)that does not resemble a human at all. All avatarsare shown in Figure 1.

Avatars are displayed at respective users’ po-sitions when the distance between two usersreaches the minimum threshold set to 2 m. Theavatar’s position and rotation is set according tothe position and head rotation of the user whoseproximity the avatar notifies. Avatars’ limbs arenot animated. Participants did not have any self-avatars. Each participant tested all four avatars.

3.2.2 Environments and Tasks

Initially, we believed that a featureless avatar(such as Shape) would be the most appropriatefor our notification strategy as it would not dis-tract users from their activity in VR. However,preliminary test sequences showed that partici-pants did not like this type of avatar at all.

We hypothesized that the suitability of a spe-cific avatar type could be affected by the struc-ture of the VE and by the task that users areaccomplishing. Therefore, we test our notifica-tion strategy in two scenarios. In the first one(”Walk” task), users simply walk in a rather con-strained VE. In the second scenario (”Game”task), the VE consists of a larger area withoutobstacles and users fulfil a specific goal.

Walk The Walk task is set in the same VE thatis used in Experiment 1. Each participant hasto move on the path indicated by arrows insidethe virtual corridor. The paths for both partic-ipants from a pair have several common seg-ments where they are likely to collide. If thedistance between the participants gets as low asthe threshold of 2m, each of them sees an avatardisplayed at the other participant’s position. Thisavatar disappears when the participants walk far-ther apart again. The task is completed when aparticipant reaches the end point of the path.

Each pair of participants from the Walk groupperformed four walk-through sequences, with adifferent notification avatar used in each.

Game The Game task is set in the VE shownin Figure 2, 8x8 meters large. The VE is filledwith blue spheres. At each moment, one of these

CASA 2017 41

spheres is coloured in pink. Participants are in-structed to collect spheres by walking into them.However, only the pink sphere can be collected.When a participant walks into it, it disappearsand a new sphere turns pink. The game is fin-ished when all spheres are collected.

The spheres are not shared between the play-ers. Each participant sees the same initial set ofspheres, and the same sphere is pink in the be-ginning. The next sphere turning pink for eachparticipant is chosen in the following way. A po-sition is calculated that is equidistant from bothplayers. Then, all remaining spheres are foundthat are located within 2m from this equidis-tant point. One of these spheres is then cho-sen to change its colour. If there are no spheresfound in this area, one of the remaining spheresis turned pink at random. With this approach,players are drawn to walk towards each otherwhen trying to collect the next sphere but donot stay too close to each other during the wholegame. As in the Walk task, players see avatars ateach others’ respective positions if the distancebetween them gets lower than 2 m.

Each pair of participants from the Game groupperformed one game sequence in which all fouravatars were used. The notification avatar waschosen in a pseudo-random manner each timewhen the participants from a pair came closer toeach other.

3.2.3 Measures

In Experiment 2, the primary measures are sub-jective questions. In a post-test questionnaire,participants are asked to rate how surprised,scared and disturbed they felt when they saweach avatar, as well as how well they could as-sociate each avatar with their test partner. Thesequestions are used to assess how disruptive theappearance of each avatar is for user experience.In addition, an in-depth discussion with eachpair of participants took place.

3.3 Technical Setup

The VR application for each participant (imple-mented in Unity version 5.3.2) is run on a Win-dows 7 laptop with an Intel Core i7 processorand an Nvidia GTX 980M graphics card, at anupdate rate of 45 fps. The used HMDs are Ocu-lus Rift DK2. We use an inside-out optical track-

ing system. A camera attached to a user’s HMDtracks markers on the ceiling of the laboratory.The size of the tracking area is 14x9 m. Thecamera position is calculated on the user’s lap-top. In Experiment 2 that uses a multi-user setupit is distributed to a server machine over a wire-less network. The equipment worn by partici-pants is shown in Figure 1.

Participants did not wear headphones, so theycould hear each other’s and the test coordinator’ssteps and eventually voices. We were aware thatsuch a setup would possibly lower participantsimmersion. Nevertheless, we decided that par-ticipants would feel safer and more confident ifthey could easily communicate with the experi-ment coordinator, and possibly also hear steps,especially in Experiment 2.

3.4 Participants

29 volunteers (10 female and 19 male) in the agefrom 21 to 48 years (median 25) took part in thestudy. 14 participants had no previous experi-ence of VR. The remaining 15 had either triedOculus Rift before or taken part in a differentuser study in VR, including the ones with realwalking. We did not intend to make a distinc-tion between naive and experienced users. How-ever, we assessed the possible influence of pre-vious experience of VR on the results. All 29participants accomplished Experiment 1. 17 par-ticipants accomplished the Walk task and 12 theGame task of Experiment 2. Experience of VR(p = 0.419) and playing video games (p = 0.845)does not differ between these two groups in theMann-Whitney test.

3.5 Procedure

Participants came to the laboratory in pairs. Oneparticipant did not have a test partner, so he com-pleted Experiment 2 together with the experi-ment coordinator (the Walk task). The study pro-cedure had the following order:

1. General introduction and explanation of theprocedure.

2. Free walking in the VE (the same as used inExperiment 1) to accommodate to the setup,separately for each participant from a pair.

3. Instructions for Experiment 1.

CASA 2017 42

4. Experiment 1 for the first participant fromthe pair.

5. Experiment 1 for the second participant.

6. Questionnaire for Experiment 1.

7. Explanation of the general purpose of thestudy and a detailed explanation of the no-tification strategy for Experiment 2.

8. Instructions for Experiment 2 (either Walkor Game).

9. Experiment 2 with both participants (Walkor Game).

10. Questionnaire for Experiment 2.

11. Discussion about the experience in Experi-ment 2 with both participants.

4 Results

4.1 Experiment 1

The results are obtained from the data of 24 par-ticipants. Data of the remaining 5 participantsare excluded due to the missing values. Nei-ther distance nor subjective difficulty data is nor-mally distributed (in Shapiro-Wilk test). There-fore, we use non-parametric tests. The signifi-cance level used for all tests is α = 0.05. Illus-trating box-plots are shown in Figure 3.

Friedman’s ANOVA does not show signifi-cant differences in the stopping distances forall four signals in the first walk-through (withthe trigger distance 2m), p = 0.192. However,the differences are significant for the secondwalk-through (with the trigger distance 1m), p<0.001. In the follow-up pairwise comparisonswith the Wilcoxon signed-rank test, the stoppingdistance for the figure signal (Median M = 0.5m) is significantly shorter than for the stop soundsignal (M = 0.78 m) (p <0.001, effect size r =0.27). The difference in the stopping distancefor the figure and the stop sign (M = 0.7 m) ismarginally insignificant (p = 0.052, r = 0.15).We believe that the significance in the ANOVAresult for the second walk-through is induced bythe fact that the figure is displayed at a closerdistance compared to the first walk-through. In-deed, the stopping distance for the figure is sig-nificantly shorter in the second walk-through (M= 0.5 m) than in the first walk-through (M = 0.7m) (Wilcoxon signed-rank test, p <0.001, r =

Figure 3: Results of Exp.1: stopping distance(left) and stopping difficulty (right).Median values, inter-quartile rangesand full ranges are shown.

0.51). Between-walk-through comparisons forthe other signals are not significant, as expected.The tests are conducted on data without outliers.

Significant differences are found in the scoresof the stopping difficulty of the signals (Fried-man’s ANOVA, p <0.001). The stopping dif-ficulty is estimated on a 7-point Likert scale,where 1 indicates that a signal is very easy and 7- very difficult. The follow-up pairwise compar-isons with the Wilcoxon signed-rank test showthat the clap sound (Median score M = 3.0) isperceived as significantly more difficult than thestop sign (M = 1.0) (p <0.001, effect size r =0.21) and the figure (M = 2.0) (p = 0.037, r =0.15). The actual stopping distance is howevernot larger for the clap sound than for any othersignal as seen from the stopping distance com-parison.

Neither VR experience no gender is signifi-cant for the stopping distance or difficulty scoresof any of the signals (in Mann-Whitney U tests).

13 participants indicated the stop sign and 8the stop sound as the preferred stopping sig-nal. Three participants found the human figureto be the easiest signal. However, participantsreported to have often ”walked into” the figureshown in front of them (at 1m distance) in thesecond walk-through and judged it as being tooclose. Clap sound was not preferred by any ofthe participants, a result that well coincides withit being found significantly more difficult thanthe visual signals. Seven participants indicatedthat they found the visual signals generally eas-ier, five said the same about the auditory signals.

CASA 2017 43

Figure 4: Box-plots of the results of Exp.2. Medians, inter-quartile ranges and full ranges are shown.

4.1.1 Discussion and Limitations

The results of Experiment 1 show that auditoryand visual (displayed at 1.5 to 2 m distance froma user) signals are on average equally effective instopping walking users in immersive VR. A vi-sual signal displayed as close as at 1 m distanceis likely to make users stop faster. On average,participants walked between 0.6 m and 0.8 m be-fore the full stop after a stopping signal had beengiven. Thus, in a situation where users are walk-ing without seeing each other on paths that po-tentially lead to a frontal collision, they shouldbe stopped no later than when the distance be-tween them reaches 2m. Our test is limited to animmersive but not interactive VE where the onlytask is to walk and stop when a signal is given.Therefore, our results are likely to indicate users’fastest possible reaction. Even though the testedsignals are similarly effective, they are differ-ently perceived by users. A sound signal not car-rying any command (the clap sound) is found tobe more difficult than visuals. The absolute ma-jority of participants prefer a stop command tobe used, although their preferences are dividedbetween this command being visual or auditory.In a VE filled with dynamic content and variousgame sounds, the best practice for quickly stop-ping users might be a combination of a visualand an auditory signal. However, a follow-upstudy in such a sensory-intense VE might be re-quired to confirm this conclusion.

4.2 Experiment 2

4.2.1 Questionnaire Responses

Figure 4 shows box-plots of responses on ques-tions to which extent each avatar surprised,scared and disturbed participants and how wellthey could associate it with their test part-ner. All answers are given on a 7-point Lik-

ert scale where 1 is ”not at all” and 7 is ”verymuch”. Scores are not normally distributed(in the Shapiro-Wilk test), except for the Asso-ciate question for Human, Robot and Spaceman.Non-parametric tests are thus chosen. The anal-ysis is performed on data without outliers. Thesignificance level used for all tests is α = 0.05.

In the Walk group, Friedman’s ANOVA indi-cates significance of the avatar type in the scoresfor Surprised (p = 0.04), Disturbed (p = 0.003)and Associate (p = 0.006). The follow-up pair-wise comparisons performed with the Wilcoxonsigned-rank test do not show significant differ-ence for all pairs. However, the effect sizescalculated based on the pairwise comparisonsrange from medium to large in many cases. Forthe Surprised question, the largest effects arefor Human-Shape (r = 0.46), Spaceman-Shape(r = 0.39) and Robot-Shape (r = 0.32). Forthe Disturbed question, again, the largest ef-fects are produced in comparisons with Shape: Human-Shape (r = 0.43), Spaceman-Shape (r= 0.33) and Robot-Shape (r = 0.5). For theAssociate question, effect sizes grow with theincrease of the difference in the degree of an-thropomorphism of the compared avatars: fromsmall for Human - Spaceman (r = 0.16) , Space-man - Robot (r = 0.18), Robot - Shape (r =0.19), to mediam for Human-Robot (r = 0.34),Spaceman-Shape (r = 0.37), to large for Human-Shape (r = 0.54). In the Game group, Fried-man’s ANOVA shows significant differences inthe scores for Scared (p = 0.036), with no signif-icant differences in the pairwise comparisons.

In the between-group analysis, the only sig-nificant result is Shape being 2 points less dis-turbing for the Game group than for the Walkgroup (Mann-Whitney U test, p = 0.011, r =0.53). Between-group comparisons for Shapeshow medium effects for the Surprised (r = 0.34)

CASA 2017 44

and Scared (r = 0.39) questions.

4.2.2 Interview Data

In the Walk group, 8 out of 17 participantsstrongly preferred to see Human avatar notify-ing them about the proximity of their test part-ner. Another 4 participants preferred any human-shaped avatar. Only one of the participantspreferred to see Shape, whereas 5 participantsstrongly disliked it.

In the Game group, 9 out of 12 participantspreferred to be notified by Shape. However,none of them disliked any other avatar.

The participants who preferred human andhuman-like avatars said that human figureslooked natural and expected. They also helpedto avoid the person whose location they notifiedas they indicated the direction of movement. Thenon-human shape was described as ”unnatural”.

Those participants who preferred the Shapeavatar felt less distracted from fulfilling theirtask and ”did not want to look” at Shape for toolong while the visuals of human-like avatars at-tracted their attention. They said that Shape’ssemi-transparency allowed to see what was go-ing on in the game in the area behind it.

4.2.3 Discussion and Limitations

The notification strategy used in Experiment 2proved successful for imminent collision pre-vention: all participants could avoid their testpartners when they saw notification avatars forshort periods of time. The results show that thetype of VE and tasks performed in it strongly af-fect users’ preferences for a notification avatar.A human avatar is desired in an explorationtask whereas a featureless non-anthropomorphicshape is better suited for a scenario with spe-cific goals. This observation follows from inter-views with participants as well as between-groupcomparisons of test scores for the Shape avatar.The core idea of our notification strategy is toprovide a non-distracting yet effective guidancefor collision avoidance. The results suggest thatthe strategy might be better suited for scenariossimilar to our Game task, where users’ attentionis concentrated on tasks but not the VE itself.While notification avatars are effective for col-lision prevention in exploration scenarios theyinevitably attract users’ attention. This fact is

reflected in the significant differences found byFriedman’s ANOVA for avatar comparisons inthe Walk group. In situations where users’ atten-tion is attracted to a notification avatar they pre-fer it to have a familiar human appearance. Inscenarios with other goals than exploration, onthe contrary, participants preferred the avatar tohave as few visual details as possible, followingour initial hypothesis.

The limitation of our experiment is the size ofthe participants’ sample. Especially in the Walkgroup, we believe that more data would allowto find statistical significance in post-ANOVApairwise comparisons that produce effects ofmedium to large size. The impact of the degreeof avatar anthropomorphism could then be estab-lished in a more detailed manner.

5 Conclusion

In the user study reported in this paper, we ex-amine two strategies for the prevention of im-minent collisions in large-scale multi-user VEsnavigated by real walking: stopping users andnotifying them of each other’s presence by dis-playing avatars. In the stopping strategy, vi-sual and auditory signals are found to not dif-fer in their effectiveness. Nevertheless, users re-port differences in the perceived difficulty andthe preferences for the stopping signal. A stopcommand in either visual or auditory form ispreferred by the majority of participants. Ourresults for the notification strategy suggest thatthe application scenario should be taken into ac-count in the choice of notification avatars. Morespecifically, interactive and goal-oriented sce-narios require abstract and featureless notifica-tion avatars, while human-like avatars are prefer-able for exploration scenarios.

The findings of both experiments can serveas guidelines for the development of scenarioswith non-shared VEs in HMD-based multi-usersetups. The development of an unnoticeable forusers collision prevention algorithm remains anavenue for future work. We believe that such analgorithm would be especially beneficial for ex-ploration scenarios in constrained VEs where asimple notification method has a larger impacton user experience.

CASA 2017 45

References

[1] D. Waller, E. Bachmann, E. Hodgson, andA. Beall. The hive: A huge immersivevirtual environment for research in spatialcognition. Behavior Research Methods,39(4):835–843, 2007.

[2] I. Podkosova, K. Vasylevska, C. Schoe-nauer, E. Vonach, P. Fikar, E. Broneder,and H. Kaufmann. Immersivedeck: Alarge-scale wireless vr system for multipleusers. In SEARIS 2016, 2016.

[3] E. Hodgson, E. Bachmann, D. Vincent,M. Zmuda, D. Waller, and J. Calus-dian. Weavr: a self-contained and wear-able immersive virtual environment simu-lation system. Behavior research methods,47(1):296–307, 2015.

[4] M. Slater, M. Usoh, and A. Steed. Tak-ing steps: The influence of a walking tech-nique on presence in virtual reality. ACMTrans. Comput.-Hum. Interact., 2(3):201–219, 1995.

[5] R. Ruddle and S. Lessels. The benefits ofusing a walking interface to navigate vir-tual environments. ACM Trans. Comput.-Hum. Interact., 16(1):5:1–5:18, 2009.

[6] E.A. Suma, S. Babu, and L. Hodges. Com-parison of travel techniques in a complex,multi-level 3d environment. In 3DUI 2007.

[7] J. Thomas, R. Bashyal, S. Goldstein, andE. Suma. Muvr: A multi-user virtual real-ity platform. In IEEE VR 2014.

[8] I. Podkosova and H. Kaufmann. MutualProximity Awareness in Immersive Multi-User Virtual Environments with RealWalking. In ICAT-EGVE 2015.

[9] K. Blom and S. Beckhaus. Virtual travelcollisions: Response method influencesperceived realism of virtual environments.ACM Trans. Appl. Percept., 10(4):25:1–25:19, 2013.

[10] G. Cirio, A. H. Olivier, M. Marchal, andJ. Pettre. Kinematic evaluation of vir-tual walking trajectories. IEEE TVCG,19(4):671–680, 2013.

[11] F. Argelaguet Sanz, A.-H. Olivier,G. Bruder, J. Pettre, and A. Lecuyer.Virtual Proxemics: Locomotion in thePresence of Obstacles in Large ImmersiveProjection Environments. In IEEE VR2015, pages 75 – 80.

[12] P. Fink, P. Foo, and W. Warren. Obstacleavoidance during walking in real and vir-tual environments. ACM Trans. Appl. Per-cept., 4(1), 2007.

[13] A.-H. Olivier, A. Marin, A. Cretual, andJ. Pettre. Minimal predicted distance:acommon metric for collision avoidanceduring pairwise interactions between walk-ers. Gait & Posture, 36(3):399 – 404, 2012.

[14] A.-H. Olivier, J. Ondrej, R. Pettre,J.and Kulpa, and A. Cretual. Interactionbetween real and virtual humans duringwalking: Perceptual evaluation of a simpledevice. In APGV, pages 117–124, 2010.

[15] J. Holm. Collision prediction and preven-tion in a simultaneous multi-user immer-sive virtual environment. Master’s thesis,Miami University, 2012.

[16] E. Hodgson, E. Bachmann, and D. Waller.Redirected walking to explore virtual envi-ronments: Assessing the potential for spa-tial interference. ACM Trans. Appl. Per-cept., 8(4):22:1–22:22, 2008.

[17] A.-H. Olivier, A. Marin, A. Cretual,A. Berthoz, and J.Pettre. Collisionavoidance between two walkers: Role-dependent strategies. Gait & Posture,38(4):751 – 756, 2013.

[18] D. Banakou, K. Chorianopoulos, andK. Anagnostou. Avatars’ appearance andsocial behavior in online virtual worlds. InPCI 2009, pages 207–211, 2009.

[19] M. Garau, M. Slater, D. Pertaub, andS. Razzaque. The responses of people tovirtual humans in an immersive virtual en-vironment. Presence: Teleoper. Virtual En-viron., 14(1):104–116, 2005.

[20] J.-L. Lugrin, J. Latt, and M. Latoschik.Anthropomorphism and illusion of virtualbody ownership. In ICAT-EGVE 2015.

CASA 2017 46