Real-Time Head Gesture Recognition on Head- Mounted Displays using Cascaded Hidden Markov Models

Transcription

1 Rea-Time Head Gesture Recognition on Head- Mounted Dispays using Cascaded Hidden Markov Modes Jingbo Zhao Robert S. Aison Department of Eectrica Engineering and Computer Science York University, Toronto, Canada {jingbo, arxiv: [cs.hc] 2 Feb 2018 Abstract Head gesture is a natura means of face-to-face communication between peope but the recognition of head gestures in the context of virtua reaity and use of head gesture as an interface for interacting with virtua avatars and virtua environments have been rarey investigated. In the current study, we present an approach for rea-time head gesture recognition on headmounted dispays using Cascaded Hidden Markov Modes. We conducted two experiments to evauate our proposed approach. In experiment 1, we trained the Cascaded Hidden Markov Modes and assessed the offine cassification performance using coected head motion data. In experiment 2, we characterized the rea-time performance of the approach by estimating the atency to recognize a head gesture with recorded rea-time cassification data. Our resuts show that the proposed approach is effective in recognizing head gestures. The method can be integrated into a virtua reaity system as a head gesture interface for interacting with virtua words. Keywords Virtua and Augmented Reaity System, Human- Computer Interaction I. INTRODUCTION Many different interfaces have been deveoped for interactions between humans and virtua reaity (VR) systems. interfaces [4], hand gesture interfaces [5], haptic interfaces [6] and ofactory dispays [7]. Athough head gesture is a natura and important way for peope to communicate and interact with each other, recognition of head gestures in the context of virtua reaity systems and the use of head gesture recognition to interact with avatars and virtua environments have been argey ignored in previous VR reated research. In the current study, we present an approach for recognizing head gestures using a headmounted dispay (HMD). Enabing rea-time head gesture recognition on HMDs coud be usefu. For instance, in virtua reaity systems, users usuay need to interact with avatars. To answer Yes/No questions asked by avatars, users coud simpy make their responses by nodding and shaking heads through a head gesture interface. One 2017 IEEE. Persona use of this materia is permitted. Permission from IEEE must be obtained for a other uses, in any current or future media, incuding reprinting/repubishing this materia for advertising or promotiona purposes, creating new coective works, for resae or redistribution to servers or ists, or reuse of any copyrighted component of this work in other works. J. Zhao and R. S. Aison, "Rea-time head gesture recognition on headmounted dispays using cascaded hidden Markov modes," 2017 IEEE Internationa Conference on Systems, Man, and Cybernetics (SMC), Banff, AB, 2017, pp doi: /SMC doi: /SMC possibe appication is to use the interface to interact with virtua tour guides in augment reaity (AR) based tours [8]. It is aso very common that a virtua reaity system itsef may raise questions to users and ask users to confirm or reject certain options; in this case, the user can aso respond through a head gesture interface by nodding and shaking. Recenty, there is a growing interest for teeoperation of robots, such as quadcopters, using head motions tracked by HMDs [9-12]. First-person views of robots are often presented directy through the dispay panes of HMDs. In the case of the quadcopter contro, a quadcopter can be maneuvered by head spatia transations [11] or the head orientation can be used to manipuate the attitude of the quadcopter to fy it [12]. Adding a head gesture interface in such appications wi enabe users to perform more compex operations. For exampe, users coud nod their head to perform mode switching to switch fight contro from auto-piot to head motion contro. Head gesture interfaces aso can be appied to (AR) devices, such as the Googe Gass and the Microsoft HooLens. Such an interface woud enabe users to perform actions, such as browsing, with head rotation, head titing, nodding and shaking without touching the gasses. A head gesture recognition method is beneficia to researchers who investigate human ocomotion or driving behaviors etc. In such activities, users usuay have to make head movements for observing the environment around them. A head gesture recognition method can be used to count the number, the type and the duration of each head movement. These parameters may differ significanty given different experimenta conditions, such as restricted fieds of view and compexity of the environment. Without such an approach for recognizing head gestures, researchers have to manuay determine the number, the type and the duration of each head movement from the coected data of head movements. Lasty, a head recognition modue aso can be integrated to driver assistance systems of automobies to monitor the driving behaviors of drivers [13]. Recognizing gestures is usuay considered as the probem of recognizing sequences and HMMs have been widey used for recognizing hand and body gestures [14-17]. Previous studies on head gesture recognitions were mosty computer vision based systems using HMMs [18],[19]. In such systems, a user s face was usuay captured by a camera and computer vision agorithms were used to track a user's face and estimate the user's head orientation from the tracked face.

2 X Morimoto et a. [18] used an optica fow agorithm to estimate the yaw, pitch and ro of a user's head. The estimated anges were quantified into seven observation symbos by threshoding. The threshods were determined by cacuating the energy of a sequence of anges, but the detais of the energy cacuation agorithm were not described. The observation symbos were used to train four HMMs. These HMMs correspond to four gestures: Yes, No, Maybe and Heo. Three gestures Yes, No and Maybe were modeed with fuy connected HMMs whie the gesture Heo was modeed with a eft-right HMM. The cassification performance was evauated offine using sequences of gestures coected from severa participants. Terven et a. [19] used the Supervised Descent Method (SDM) to extract 2-D facia features. The yaw, pitch and ro were estimated by using the POSIT (Pose from Orthography and Scaing with Iterations) method with the extracted 2-D facia features and a 3-D anthropometric head mode. The symbos were generated by comparing changes in yaw and pitch across consecutive frames. To cover six different head gestures: Nodding, Shaking, Left, Right, Up and Down, the researchers trained six HMMs and these modes were fuy connected to form a cycic structure. To evauate the cassification performance, videos that contained different head gestures were recorded and the performance was evauated offine. Athough computer vision based systems using HMMs for head gesture recognition exist, to the authors' knowedge, HMMs have not been used for recognizing head gestures on HMDs. In this paper, we present the Cascaded Hidden Markov Modes (CHMMs) for head gesture recognition on HMDs. Unike existing systems that contain cycic structure, our system is fuy pipeined, making it more suitabe for rea-time appications. The main contributions of the paper are: Y Fig. 1. The coordinate system of the Ocuus Rift DK2 [20]. 1) We proposed the CHMM structure for rea-time head gesture recognition. The system recognizes nine types of head gestures and can be further extended to incude more head gestures; 2) We proposed a training method that maximizes the cassification performance of the CHMM; 3) We presented a procedure for simutaneousy capturing and abeing head gestures using HMDs; The present study focused on a specific HMD -- the Ocuus Rift DK2, but the proposed method is genera and can be adapted to work with VR systems that use head tracking gasses or other types of systems in which a user's head motion is tracked by fast and accurate tracking systems. Z II. METHODS A. The Tracking System of the Ocuus Rift DK2 The Ocuus Rift DK2 uses a six Degree-of-Freedom (DOF) hybrid optica-inertia tracker to track a user's head motion at approximatey 75 Hz. The coordinate system of the Ocuus Rift DK2 is iustrated in Fig. 1. The hybrid tracker consists of an externa camera with an infrared fiter to track the infrared LED array on the front and side panes of the Ocuus Rift DK2 and an embedded inertia measurement unit (IMU) [21],[22]. The tracking data that can be accessed through the VR software WordViz Vizard 5.0 are position, acceeration, Euer anges and anguar veocities. To recognize head gestures, we ony used the anguar veocities of head motions as anguar veocities directy refect whether a user's head is moving and in which direction. We represent the anguar veocity as a 3-D vector ω = (Ψ, Θ, Φ ), where Ψ, Θ, Φ are yaw veocity, pitch veocity and ro veocity, respectivey. A sequence of anguar veocities can be denoted as W = ω 1 ω 2 ω i, where i is the index for a 3-D vector ω. Another option for monitoring head movements is to use quaternions to represent the head anguar veocity but extra time wi be taken to convert head anguar veocities to quaternions. B. Definition of Head Gestures We defined nine casses of head gestures. Seven are simpe gestures: Being Ide (remaining sti), Rotating Left, Rotating Right, Titing Upward, Titing Downward, Leaning Left, Leaning Right. Two are compex head gestures: Shaking, Nodding. The motivation behind the definition of compex gestures is that shaking can be represented a sequence of three simpe head gestures, which are Being Ide, Rotating Left and Rotating Right. Simiary, Nodding can be represented by a sequence of simpe gestures: Being Ide, Titing Upward and Titing Downward. We associated each cass of head gesture with a cass abe, with {1,2,,9}. C. Cascaded Hidden Markov Modes An HMM [23],[24] is governed by the foowing parameters: N the number of hidden states, M the number of observation symbos and the mode parameter λ = (A, B, π), where A is the matrix that represents the transition probabiity between states, B the matrix that represents the emission probabiity of a symbo observed from a specific state and π the initia state probabiities. Simiar to the speech recognition approach [23], we modeed each head gesture with a eft-right HMM λ, where is the cass abe of a head gesture associated with the HMM λ. In a eft-right HMM, ony transitions between adjacent states from eft to right and transitions from a state to itsef are aowed. A set of trained HMMs for the nine casses of head gestures can be represented as Λ, Λ = {λ 1, λ 2,, λ 9 }. The HMMs we used were discrete HMMs and discrete HMMs ony accept discrete observation symbos as inputs. Thus, given an sequence of samped anguar veocities W, before feeding the sequence W into an HMM λ, we used the K-Means agorithm during training step and the minimum distance cassifier [25] during testing step as the vector quantization (VQ) procedure to

3 Being Ide Rotating Left Shaking Vector Quantization Rotating Right Titing Up Titing Down Nodding Output Seection Recognized Gesture Leaning Left Leaning Right Simpe Gestures Compex Gestures Fig. 2. The structure of the CHMMs for rea-time head gesture recognition. quantize the sequence of anguar veocities W into an observation sequence S that consists of discrete observation symbos with S = O 1 O 2 O i, where i is the index of the observation symbo O in the sequence S. To predict how ikey a sequence S beongs to a certain cass of a head gesture, we used a set of trained HMMs Λ and the forward procedure of the HMM to cacuate the posterior probabiities P a for a HMMs, with P a = { P(S λ 1 ), P(S λ 2 ),, P(S λ 9 ) }. An output seection procedure resoves the cass abe of the given observation sequence S from the posterior probabiities P a. As simpe gestures require much ess data to give a reiabe estimate than compex gestures do, to efficienty recognize simpe and compex head gestures, we organized the set of HMMs Λ into the CHMM structure. The structure has two dedicated ayers for recognizing simpe gestures and compex gestures respectivey. For training and testing the CHMM, two HMM agorithms were used. One is the Baum Wech agorithm, which was used to train a eft-right HMM λ. The other is the forward procedure of the HMM, which cacuates the posterior probabiity P(S λ ) of a HMM λ given an observation sequence S. These two agorithms are avaiabe in Matab 2014a as hmmtrain() and hmmdecode(). Detaied descriptions of the agorithms can be found in [23],[24]. In Fig. 2, we present the proposed CHMM structure. Here we describe the rea-time operation of the CHMM and eave the expanation of the training and testing procedures to Section 3. During rea-time operation, the system continuousy reads anguar veocity ω i for processing. A samped anguar veocity ω i is given into the vector quantization modue and the vector quantization modue produces an observation symbo O i based on the Eucidean distance between custer centers C j and the vector ω i using the minimum distance cassifier [25]: the index j of custer center C j that gives the shortest Eucidean distance is assigned as the observation symbo to the vector ω i : O i = argmin ( C j ω i j 2 ) (1) where O i is the assigned observation symbo and custer centers C j were obtained by K-Means during training. The observation symbo O i generated by the vector quantization modue is buffered using an array unti the number of observation symbos O i reaches the ength L B. When the number of observation symbo O i reaches L B, the sequence of observation symbos S is sent to the modues cassifying simpe gestures to cacuate the posterior probabiities P s, with P s = { P(S λ 1 ), P(S λ 2 ),, P(S λ 7 ) }. The buffer is then immediatey ceared and waits for new observation symbos O i. If a recognized simpe gesture is in the category of Being Ide, Rotating Left, Rotating Right, Titing Up or Titing Down, they wi be considered as observation symbos O i for the ayer of compex gestures and wi be further buffered using a queue of a ength L Q. Each time the queue performs a dequeue and an enqueue operation, the buffered sequence in the queue wi be sent to the modue of compex gestures to cacuate the posterior probabiities P c of compex gestures, with P c = {P(S λ 8 ),P(S λ 9 )}. We empiricay set L B = 10 and L Q = 10 to make the array contain 0.13 s of data and the queue contain 1.3 s of data at the samping rate of 75 Hz. We found such choice of vaues gave a reativey fast response for recognizing simpe gestures and reiabe ength of data for recognizing compex gestures. The ast step is to resove the cass abe of the head gesture based on the cacuated posterior probabiities P s and P c. The posterior probabiities of simpe gestures P s and the posterior probabiities of compex gestures P c are not directy comparabe as a compex gesture consists of symbos represented by simpe gestures. To sove this probem, we proposed an output seection procedure to estimate the cass abe from the posterior probabiities of simpe gestures P s and compex gestures P c using the function: = { argmax(p c ) argmax(p s ) if P(S λ 8 ) > τ n or if P(S λ 9 ) > τ s otherwise where τ n and τ s are the threshods that indicate a compex gesture shaking or nodding may exist if either the posterior probabiity P(S λ 8 ) or P(S λ 9 ) is arger than their corresponding threshods τ n or τ s ; is the cass abe associated with a HMM λ of a specific gesture and is the estimated cass abe. (2)

4 TABLE I. THE AVERAGE ACCURACY OF THE SIMPLE GESTURE LAYER FROM A TRAINING SESSION (UNIT: PERCENTAGE, M: THE NUMBER OF DISCRETE SYMBOLS IN HMMS, N: THE NUMBER OF HIDDEN STATES IN HMMS). M N III. EXPERIMENT 1: TRAINING AND TESTING OF THE CHMM In experiment 1, we trained the CHMM and evauated its offine cassification performance. As there was no pubicy avaiabe head gesture dataset for the Ocuus Rift DK2, we deveoped a custom appication, using the Vizard 5.0, that can simutaneousy coect and abe head gestures for training and testing the proposed CHMM structure. Nineteen peope participated in the experiment and informed consents were obtained from a participants in accordance with a protoco approved by the Human Participants Review Subcommittee at York University. In the experiment, the participants wore the Ocuus Rift DK2 and sat approximatey 60 cm in front of the tracking camera of the Ocuus Rift DK2, which was mounted on the monitor of the host machine (Windows 7, an Inte i7 2.8 GHz. CPU, 4 GB memory and an AMD Radeon HD 6850 graphics card). There were nine types (casses) of head gestures that needed to be coected. For each type of head gesture, the researcher pressed the corresponding button on the contro pane (visibe ony to the researcher on the computer monitor) of the custom appication and a prompt (visibe to both participants on the HMD and the researcher on the computer monitor) indicating the type of head gesture that a participant needed to perform was shown in the view of the HMD. A countdown timer was aso started at the same time to count from two to zero by seconds. A participant was expected to compete the head gesture within 2 seconds with their preferred head movement speed. Labeing of the head gesture was done at the same time by the custom appication within the 2-second interva. We coected each type of head gesture twice for each participant. In tota, we had 342 head gesture sampes from nineteen participants. Our training and testing procedures were fuy automated and performed using Matab 2014a. We used muti-cass precision (Precision M ), muti-cass reca (Reca M ) and the average accuracy [26] as the metrics to evauate the performance of training and testing. We divided the training of the CHMM into three phases that trained the modues of vector quantization, simpe gesture ayer and compex gesture ayer separatey. Given the coected head gesture dataset from nineteen participants, we directy ran the K-means agorithm to quantize each anguar veocity vector ω i into an observation symbo O i. The custer centers C j obtained from K-Means training were stored for the testing step for the vector quantization modue. The custer number K of K-Means and the number of observation symbos M of HMMs were equated (M = K). Seven types of observation symbos, with M 7, were needed to represent seven simpe gestures. The head gesture dataset was then divided in haves into a training set and a testing set. For each sequence of head gestures other than being ide, we removed redundant symbos that indicate a user s head is remaining sti as the redundant symbos were not usefu for training. The second step was to train the ayer of simpe gestures and evauate its cassification performance. To represent a simpe head gesture, such as rotating eft, at east two hidden states N are needed, with N 2. The Baum Wech agorithm was used for training. Since we chose L B = 10 for rea-time evauation, we partitioned each sequence in the training set into short sequences of ength 10 and used the short sequences to train its each associated HMM λ, {1, 2,,7}, to obtain the parameters of the transition matrix A and the emission matrix B. The initia state probabiities π were not considered as a head gesture aways starts with the state of the head being remaining sti. The information can be earned during training and stored in the emission matrix B. There were two tunabe parameters: N the number of hidden states in the mode, M the number of observation symbos in an HMM and we knew that N 2 and M 7. To obtain the best cassification performance, we wished to find the optima vaues of N and M that maximize the average accuracy for the ayer of simpe gestures, with smaest M and N possibe. The smaer N and M are, fewer additions and mutipications are invoved in the forward procedure of the HMM; hence the faster rea-time performance for head gesture recognition. The upper bounds for N and M were set to 6 and 25 empiricay. The evauation of the cassification performance of the simpe gesture ayer was conducted after each training cyce with a combination of N and M. We partitioned each sequence in the testing set into the short sequences of ength 10 and used the forward procedure to cacuate the posterior probabiities P s based on short sequences. The cass abe of a simpe head gesture was estimated using the equation: = argmax(p s ) (3) Since the K-means is initiaized randomy, the training resuts may differ even if we run the same agorithm on the same training set. Thus, we ran the training procedure for the ayer of simpe gestures for five different sessions. In each session, the training was performed with different combinations of N and M such that 2 N 6, N Z and 7 M 25, M Z. The highest average accuracy we were abe to obtain from a specific training session was 99.2% when N = 3 and M = 17 (see Tabe I) and the corresponding muti-cass precision and muti-cass reca were 97.9% and 96.6%, respectivey.

5 Gesture Labe Anguar Veocity (Rad / S) TABLE II. HEAD GESTURE RECOGNITION LATENCIES (UNIT: S). RL RR TU TD LL LR S N P P P P P P Mean Std S N LR LL TD TU RR RL BI Yaw Pitch Ro Gesture Frames Fig. 3. An exempar rea-time recognition resut from a participant. To train the compex gesture ayer and evauate the cassification performance, we first partitioned the sequence of nodding and shaking in the head gesture datasets into short sequences of ength 10 (since L Q = 10). We then used the trained HMMs Λ s = {λ 1, λ 2,, λ 7 } of simpe gestures to cassify the short sequences of compex gestures into observation symbos O i that consists of cassified simpe gestures. Specificay, the cassification was done by first using the forward procedure of the HMM to cacuate the posterior probabiities of each short sequence of compex gestures and assigning the short sequence with the cass abe of the simpe gesture with the highest posterior probabiity using equation (3). As a compex gesture is represented by three simpe gestures, we set the number of observation symbos M = 3. The number of hidden states was set as N = 3, which was same as that of the simpe gesture ayer. We then used the Baum Wech agorithm to train the ayer of compex gestures and obtained the parameters for the transition matrix A and the emission matrix B for HMMs of compex gestures Λ c = {λ 8, λ 9 }. As with the training of the simpe gesture ayer, the initia state probabiities π were not considered. To determine the threshods τ n and τ s, we used the forward procedure and the HMMs Λ c = {λ 8, λ 9 } to cacuate the posterior probabiities P c of a sequences of compex head gestures in the training set. We seected the smaest vaues as the threshods τ n and τ s for shaking and nodding, respectivey, with τ n = and τ s = The ast step was to test the ayer of compex gestures. We cacuated the posterior probabiity P c, P c = {P(S λ 8 ), P(S λ 9 )}, of each compex gesture sequence in the testing set, compared the posterior probabiities P c with the threshods τ n and τ s we obtained during the training procedure and estimated the cass abe of each sequence using the equation: = { argmax(p c ) 1 if P(S λ 8 ) > τ n or if P(S λ 9 ) > τ s otherwise where -1 wi be given as an invaid cass abe when both P(S λ 8 ) and P(S λ 9 ) were ower than their corresponding threshods τ n and τ s. The muti-cass precision, the muti-cass reca and the average accuracy were 100%, 96.4% and 98.5%, respectivey, for the ayer of compex gestures. IV. EXPERIMENT 2: REAL-TIME EVALUATION OF THE LATENCY OF THE CHMM For rea-time evauation of the atency of the head gesture recognition framework, we impemented the proposed CHMM structure and the forward procedure of the HMM using python 2.7 in Vizard 5.0. Our goa was to estimate the atency for the agorithm to recognize a head gesture. In practice, we found it necessary to further tune the parameters τ n and τ s such that τ n = 5 and τ s = 4. This heped the CHMM avoid confusing fast head rotation and titing with Shaking and Nodding during rea-time recognition. Each computation cyce of the CHMM takes approximatey 1 ms and thus the proposed approach meets the rea-time requirement of competing a computation cyce within 13 ms at the samping rate of 75 Hz. Nine peope participated in the experiment and none had participated in experiment 1. Informed consents were obtained from a participants in accordance with a protoco approved by the Human Participants Review Subcommittee at York University. In experiment 2, participants aso wore the Ocuus Rift DK2 and sat 60 cm in front of the tracking camera of the Ocuus Rift DK2. Participants were asked to perform the foowing head gesture sequence with their preferred speed three times from the initia neutra position with a head gesture of Being Ide (BI): 1) Rotating Left (RL); 2) Move Back to Neutra Position; 3) Rotating Right (RR); 4) Move Back to Neutra Position; 5) Titing Upward (TU); 6) Move Back to Neutra Position; 7) Titing Downward (TD); 8) Move Back to Neutra Position; 9) Leaning Left (LL); 10) Move Back to Neutra Position; (4)

6 11) Leaning Right (LR); 12) Move Back to Neutra Position; 13) Nodding (N); 14) Move Back to Neutra Position; 15) Shaking (S); 16) Move Back to Neutra Position; The estimated head gesture abes and head anguar veocities during rea-time operation were recorded for atency anaysis (see Fig. 3). We defined the atency for head gestures, except Being Ide, as the time interva between the index of the frame i init when the Eucidean norm of a user's head anguar veocity ω is equa to or higher than a threshod ω init = 0.1 rad / s, to the index of the frame i trigger at which a head gesture was triggered. Then the atency t L can be estimated as: t L = i trigger i init f s (5) where f s is the samping frequency of the Ocuus Rift DK2 and f s = 75 Hz. Three participants were unabe to foow the gesture sequence given above as they did not remember the sequence they needed to perform. Thus, we ony used the data of the remaining six participants and the researcher seected one specific sequence among the three that a participant performed to estimate the atency. The seected sequence was the cearest pattern compared with that of other two sequences. In tabe II, we present the resut of the estimated head gesture recognition atencies of six participants (P1 P6). The mean vaue and the standard deviation of the atencies of a participants were aso cacuated. We can find that for simpe gestures, the atency had a mean vaue of 0.17 s and for compex gestures the atency had a mean vaue of 0.71 s. V. CONCLUSIONS We have presented a CHMM structure for rea-time head gesture recognition. The proposed structure is scaabe and modues for other types of gestures can be added or the existing modues can be removed based on the appication needs. A distinct advantage of the proposed pipeined structure is that the structure can be more easiy impemented on FPGAs (Fied Programmabe Gate Arrays) and ASICs (Appication-specific Integrated Circuits) compared with the cycic structure described in [19], as a cycic structure is iterative and nondeterministic. Such modues can be integrated into head wearabe devices, such as the Googe Gass and Microsoft HooLens, as a dedicated modue for fast head gesture recognition. A imitation with the user-study based rea-time evauation is that it is impossibe to ask the user to perform head motions with precise veocity and duration. Thus, the rea-time cassification performance of the proposed approach needs to be further evauated with a robotic head since the veocity and the duration of a head gesture performed by a robotic head can be easiy controed by programming. This wi enabe us to compare the timings of head movements with that of the reatime cassification resuts and determine the cassification performance. REFERENCES [1] C. Ware, K. Arthur and K. S. Booth. Fish tank virtua reaity. In Proc. INTERACT 1993 and CHI [2] I. Sutherand. A head-mounted three dimensiona dispay. In Proc. FJCC [3] C. Cruz-Neira, D. J. Sandin and T. A. DeFanti. Surround-Screen Projection-Based Virtua Reaity: The Design and Impementation of the CAVE. In Proc. SIGGRAPH [4] J.M. Hoerbach. Locomotion interfaces. In Handbk. of VE: Dsg. Imp. and App., Lawrence Erbaum Associates, Inc., [5] D. Xu. A Neura Network Approach for Hand Gesture Recognition in Virtua Reaity Driving Training System of SPG. In Proc. ICPR [6] I. Choi and S. Fomer. Woverine: A Wearabe Haptic Interface for Grasping in VR. In Proc. UIST [7] H. Matsukura, H. Yoshida, T. Nakamoto and H. Ishida. Synchronized presentation of odor with airfow using ofactory dispay. In Jrn. Mech. Sci. Tech. 24, 1, [8] A.F. Abate, G. Acampora and S. Ricciardi. An interactive virtua guide for the AR based visit of archaeoogica sites. In Jrn. Vi. Lang. & Compu. 22, 6, [9] H. Martins and R. Ventura. Immersive 3-d teeoperation of a search and rescue robot using a head-mounted dispay. In IEEE Conf. ETFA [10] N. Moet and R. Cheai. Virtua and augmented reaity with headtracking for efficient teeoperation of groups of robots. In IEEE Conf. Cyberwords 2008, [11] K. Higuchi, K. Fujii and J. Rekimoto. Fying head: A headsynchronization mechanism for fying teepresence. In Proc. ICAT 2013, [12] C. Pittman and J. J. LaVioa, Jr.. Exporing head tracked head mounted dispays for first person robot teeoperation. In Proc. IUI [13] H. B. Kang. Various Approaches for Driver and Driving Behavior Monitoring: A Review. In IEEE ICCVW [14] K. Liu, C. Chen, R. Jafari and N. Kehtarnavaz. Muti-HMM cassification for hand gesture recognition using two differing modaity sensors. In Proc. IEEE DCAS 2014, 1-4. [15] M. Hossain and M. Jenkin. Recognizing hand-raising gestures using HMM. In Proc. CRV 2005, [16] L. W. Campbe, D. A. Becker. A. Azarbayejani, A. F. Bobick, and A. Pentand. Invariant features for 3-D gesture recognition. In Proc. FG 1996, 157. [17] C. Chen, J. Liang, H. Zhao, H. Hu and J. Tian. Factoria HMM and Parae HMM for Gait Recognition. In IEEE Trans. Syst. Man. Cybern. C (App. and Rev.), 39, 1, [18] C. Morimoto, Y. Yacoob and L. Davis. Recognition of head gestures using hidden Markov modes. In Proc. ICPR , [19] J. R. Terven, J. Saas and B. Raducanu. Robust Head Gestures Recognition for Assistive Technoogy. In Proc. MCPR 2014, [20] ast accessed Dec [21] S. M. LaVae, A. Yershova, M. Katsev and M. Antonov. Head tracking for the Ocuus Rift. In Proc. ICRA 2014, pp [22] Hacking the Ocuus Rift DK2. ast accessed Dec [23] L. R. Rabiner. A tutoria on hidden Markov modes and seected appications in speech recognition. In Proc. IEEE, 77, 2, [24] D. Ramage. Hidden Markov modes fundamentas. Lecture Notes ast accessed Dec [25] H. Lin and A. N. Venetsanopouos. A weighted minimum distance cassifier for pattern recognition, In Proc. Canadian Conf. on ECE, 2, [26] M. Sokoova and G. Lapame. A systematic anaysis of performance measures for cassification tasks. Inf. Process. Manage. 45, 4,