2011 IEEE 15th International Symposium on Consumer Electronics
`ACCELEROMETER-BASED SWINGING GESTURE DETECTION FOR AN
`ELECTRONIC HANDBELL
`
`Liyanaarachchi Lekamalage Chamara Kasun, Wooi-Boon GOH
`School of Computer Engineering, Nanyang Technological University, Singapore 639798
`CHAMARAKASUN@ntu.edu.sg: ASWBGOH@ntu.edu.sg
`
`ABSTRACT
`
`This paper tackles the problem of detecting the swinging action of
`an electronic handbell. It describes a threshold based algorithm
`that is able to detect an orientation free swinging motion using
`only the X and Y axis signals of an accelerometer that is mounted
`at the end of a handle. Equations governing the accelerations of the
`accelerometer are defined. The equations are used to select the
`appropriate accelerometer axis for swing motion detection, which
`were X axis and Y axis. The characteristics of the swing motion
`are identified empirically and incorporated to the swing detection
`algorithm. The experimental
`results
`for swinging motion
`performed by 4 users on the electronic bell show that the accuracy
`of swing motion detection is 95.3%.
`1. INTRODUCTION
`Handbells are often use by pre-school educators to teach
`children musical concepts such as rhythm, note sequences
`and timing. Figure l a shows a physical handbell, which is
`usually one of a set of many; each carefully tuned to ring at
`a specific frequency so that it plays a particular note when
`swung. They are generally costly and can only ring out the
`sound of one specific note. We are interested in developing
`an electronic equivalent of such handbells, which can be
`easily reprogrammed to ring a synthetic bell sound of any
`note of choice when swung in the same physical manner.
`Figure 1 b shows the physical construction of such a device
`and a schematic block diagram of its major components,
`which consist of a MCU, audio module and motion sensing
`module, which comprises of a 3-axis accelerometer. The
`main issue addressed by this paper is the development of an
`appropriate algorithm that would allow this swinging
`motion gesture to be detected in an orientation free manner
`signals
`from
`the
`single 3-axis
`using appropriate
`accelerometer.
`
`Accelerometer
`
`Handle
`
`(a)
`
`(b)
`
`Figure 1. (a) A physical handbell. (b) The electronic
`handbell and its various component modules
`
`Gestures are expressive means of communicating between
`consumer devices and human. Gesture recognition is
`performed using video [1, 2, 3], touch screen [4] or by using
`micro electro mechanical systems (MEMS) such as
`accelerometers [5]. Accelerometers are devices which
`measures acceleration along a predefined axis. It is capable
`of measuring static (acceleration created by gravity) and
`dynamic acceleration (acceleration created by movement).
`Static acceleration is used to calculate the tilt or orientation
`of a device, while dynamic acceleration is used for gesture
`recognition and fall down detection. The advantage of using
`video based systems for gesture recognition is that the user
`does not need to wear any sensors or hold any device.
`However, vision based gesture recognition are complicated
`and prone to the problem of self-occlusion and fast gesture
`speeds due to the limited frame rates of the video camera.
`
`Wearable devices for gesture recognition are easier to
`implement because the orientation of device can fixed by
`the way it is secured on the human body, for example with
`the use of gloves or strap-on embedded sensors. Near
`symmetrical hand held devices such as that shown in Figure
`1 b are more problematic for accelerometer-based gesture
`recognition since the orientation of the device can be rather
`arbitrary during motion execution. One way to avoid this
`problem is to define the physical dimensions of the hand
`held device in a manner that allows human ergonomics or
`device function to persuade the user to hold it in a consistent
`orientation. Some examples include gesture-capable devices
`such as the iPhone [6] from Apple and the Wii remote [7]
`from Nintendo. Phones have a screen and are usually
`designed to be flat, which guide users to instinctively hold
`them in a specific orientation that is comfortable and permit
`good view of the screen. Similarly, the Wii remote models
`after typical remote controller-type interfaces, which have
`buttons on only one face of an elongated device, which
`encourage handling in a pre-deterministic orientation. In our
`case, in mimicking the symmetrical form factor of typical
`handbells, the orientation constraint required to make
`accelerometer-based gesture recognition easier has been
`compromised. This paper presents an algorithm that is able
`to overcome this limitation.
`2. EXISTING APPROACHES
`Existing gesture recognition algorithms using signals from
`accelerometer-based motion sensors can be categorized
`based on the following gesture classification methods:
`
`• Hidden Markov Models (HMM)
`• Dynamic Time Warping (DTW)
`
`978-1-61 284-842-6/11/$26.00 ©2011 IEEE
`
`272
`
`Authorized licensed use limited to: University of Southern California. Downloaded on October 14,2023 at 00:53:08 UTC from IEEE Xplore. Restrictions apply.
`
`Petitioner Samsung Ex-1049, 0001
`
`

`

`logic and
`
`• Support Vector Machine (SVM)
`• Machine Learning Algorithms (Fuzzy
`Artificial Neural Networks)
`• Threshold-based
`Hidden Markov Models (HMM)
`Joselli and Clua [8] presented a HMM based gesture
`recognition method for mobile phones. The authors propose
`that the 3–axis accelerometer values to be processed in 5
`stages. The initial stage will find the start and the end of the
`motion. The second stage is used to reduce noise via a low
`pass filter. The third stage reduces the amount of data sent
`to the HMM by using a k-mean algorithm. The fourth stage
`uses an existing proven 8 states HMM for gesture
`identification. In the final stage a Bayesian classifier is used
`to remove non gesture movements. Dictionary of 10
`gestures are created by the authors and the gesture
`recognition accuracy for the set of gestures are between
`75% - 97.5%. If the HMM parameters; the number of states
`of HMM, number of distinct observation symbols per state,
`state transition probability, observation symbol probability
`distribution in a state and initial state distribution is known,
`then a gesture could be modeled by HMM. Kong et al. [12]
`proposes an algorithm to capture these parameters, allowing
`to model gestures using HMM. The authors however have
`not determined the gesture recognition accuracy of the
`proposed gesture model.
`Dynamic Time Warping (DTW)
`The same gesture performed by various users is not exactly
`same. Hence gesture recognition algorithms should be
`capable of detecting gestures with minimal user
`dependence. In order to tackle this issue Akl and Valaee [9]
`proposed a gesture recognition method based on DTW. The
`Wii remote is used as the hardware platform and the authors
`have
`created
`a dictionary of 18 gestures. The
`experimentation results show that the gesture recognition
`accuracy of the proposed method is between 90% - 99.79%.
`Same gesture performed by the same user on different days
`tends to change. In order to tackle this problem Liu et al.
`[10] proposed user personalized gesture recognition method
`based on DTW called μwave. For the same gesture the
`proposed method in contrast to [9] uses different gesture
`templates for each user, which changes daily. A dictionary
`of 8 gestures has been defined. They implemented the
`algorithm using Wii remote and the gesture recognition
`accuracy is between 98.6%-98.9%.
`Support Vector Machine (SVM)
`The effectiveness of any gesture recognition method
`depends upon the features extracted. He et al. [11] proposed
`3 methods for feature extraction namely, discrete cosine
`transform (DCT), Fast Fourier Transform (FFT) and a
`hybrid method which uses wavelet packet decomposition
`(WPD) with FFT. Gesture recognition is performed by
`using SVM in all 3 cases. A dictionary of 17 gestures is
`defined and a mobile phone with a 3-axis accelerometer is
`used for implementation. DCT, FFT and the hybrid method
`
`respectively produces gesture recognition accuracy of
`64.51% - 95.49%, 70.44% - 94.29% and 71.98% - 95.49%.
`Machine Learning Algorithms
`Gesture is nebulous by nature hence gesture recognition can
`be based on highly adaptable algorithms such as fuzzy logic
`and artificial neural networks. Helmi and Helmi [13]
`propose a gesture recognition method based on fuzzy logic
`and neural networks. Dictionary of 25 gestures are defined
`and implemented on a wireless accelerometer device which
`transmits the acceleration data to a computer. The authors
`showed that machine learning algorithms are the ideal
`solution
`for gesture detection because
`the gesture
`recognition accuracy is 100%. But the drawback is the
`required huge computational resources to perform gesture
`detection, which is not appropriate in an embedded
`processing situation like the electronic bell. Bailador et al.
`[14] used a Continuous Time Recurrent Neural Network
`(CRTNN)
`for gesture
`recognition. As CRTNN
`is
`computationally less expensive, it can perform gesture
`recognition in real-time. Dictionary of 8 gestures are
`defined and gesture recognition accuracy is 64%. But when
`the gestures are done in controlled manner (user sitting
`while performing gestures, gestures are performed non-
`continuously) recognition accuracy of up to 94% could be
`obtained.
`Threshold based
`Parsani and Singh [15] proposed using the running variance
`to find activity in the accelerometer. Once the running
`variance is higher than a predetermined threshold, a sub
`gesture detection algorithm is executed to find the specific
`gesture. Dictionary of 6 gestures are defined by the authors
`and implemented on a programmable system on chip which
`communicates with a PC using Bluetooth. However the
`authors haven’t performed any experiments to determined
`the gesture recognition accuracy of their propose method.
`Many of the gesture recognition methods reviewed require
`training phase in which templates of gestures to be
`classified are generated when sufficient exemplar gestures
`are supplied. The requirement of training can make the
`method somewhat inconvenient. Most importantly, of the
`methods reviewed, the threshold-based method requires the
`least computational resources such as memory and
`processing power, making it an ideal approach for an
`embedded solution which has constrained resources. In this
`paper we propose a novel threshold-based gesture detection
`algorithm to detect the orientation-free swinging gestures on
`a physical construction such as that shown in Figure 1b.
`3. SWINGING MOTION
`The physical construction of the electronic bell and the
`respective three axes of the embedded accelerometer are
`shown
`in Figure 2. For
`the ease of discussion, 3
`perpendicular reference axes called X, Y and Z are created
`with respect to the user. Any forward-backward movement
`performed by the user is called X-axis movement, left-right
`
`Authorized licensed use limited to: University of Southern California. Downloaded on October 14,2023 at 00:53:08 UTC from IEEE Xplore. Restrictions apply.
`
`273
`
`Petitioner Samsung Ex-1049, 0002
`
`

`

`movement performed by the user is called Y-axis movement
`and any up-down or rotation movement performed by the
`user is called Z-axis movement. The red dot shown in
`Figure 2 describes the orientation of the accelerometer X, Y,
`Z axis. If the user holds the electronic bell as shown in
`figure 2 where the red dot is in the upper right corner, then
`the accelerometer axes and reference axes coincide and
`called as the ideal position. Because X, Y, Z acceleration
`created by movement correlates to the X, Y, Z accelerations
`measured by the accelerometer.
`
`(1)
`
`Where,
`a; is the instantaneous linear tangential acceleration,
`r is the radius of the swing motion and
`Bis the instantaneous angular acceleration.
`
`The instantaneous linear tangential acceleration is equivalent
`to the dynamic acceleration detected by the accelerometer
`when a swinging motion is performed.
`
`X,YandZ
`Accelerometer
`Axes
`
`X,YandZ
`Reference Axes
`
`Figure 2. Physical construction of the electronic bell and the
`directions of the various accelerometer axes.
`
`When the electronic bell is rotated around the Z reference
`axis, angle µ defines the rotation from the ideal position as
`show in figure 4a. The swing angle a defines the movement
`of the electronic bell when swinging motion is performed
`from the pivot point as shown in Figure 4b. This electronic
`bell is used to analyze the static and dynamic acceleration
`created along the 3 reference axis when a swinging motion
`is performed. The swinging motion is essentially an X axis
`and/or Y axis movement, but not a Z axis movement due to
`the physical characteristics of the electronic bell as shown in
`Figure 3.
`
`Z axis Movement (not a swing motion)
`
`X axis Movement (swing motion)
`
`Y axis Movement (swing motion)
`
`Figure 3. Valid swinging motions in the X and Y directions
`
`The user can perform the swing motion by using the wrist or
`the elbow and the pivot point will be respectively wrist or
`elbow. It is most unlikely that a person will use both the
`elbow and wrist simultaneously to perform a swing motion.
`Hence the person will create a graceful arc when the swing
`motion is performed.
`
`The instantaneous tangential acceleration of a circular
`motion can be described by using equation 1.
`
`(b)
`(a)
`Figure 4. (a) Rotation along the Z axis given byµ and (b)
`the swing angle a when a swinging motion performed
`The total accelerations detected by the accelerometer along
`the 3 axes X, Y, Z are respectively described by equations
`2, 3 and 4.
`arx = adY sinµ+ a 5y sinµ + adx cos µ + a 5x cos µ
`ary = adX sinµ + a 5x sin µ + adY cos µ + a 5 y cosµ
`
`(2)
`
`(3)
`
`Where,
`arx :- Total acceleration detected by the accelerometer along
`the accelerometer X axis.
`ary :- Total acceleration detected by the accelerometer along
`the accelerometer Y axis.
`arz :- Total acceleration detected by the accelerometer along
`the accelerometer Z axis.
`adx:- Dynamic acceleration along the reference X axis.
`ady:- Dynamic acceleration along the reference Y axis.
`a 5x:- Static acceleration along the reference X axis.
`a 5y:- Static acceleration along the reference Y axis.
`a 5z:- Static acceleration along the reference Z axis.
`a:- The angle of the swing motion performed.
`µ:- The angle rotated along the Z axis of the reference axis
`
`By analyzing equations 2, 3 and 4, it can be inferred that the
`easiest way to detect a swinging motion is to use equation 4.
`According to equation 4, when the electronic bell is kept in
`the upright position (as shown in Figure 2), the Z axis of the
`accelerometer will detect - 1 g and when horizontal, it will
`measure Og. When a forward-backward swinging
`is
`performed with a less than ,r/18 radians (10 degrees) andµ
`equal to O radians, the acceleration generated in the Z axis
`of the accelerometer will be insignificant compared to the X
`
`Authorized licensed use limited to: University of Southern California. Downloaded on October 14,2023 at 00:53:08 UTC from IEEE Xplore. Restrictions apply.
`
`274
`
`Petitioner Samsung Ex-1049, 0003
`
`

`

`axis accelerometer values as shown in Figure 5. This is
`because according to equation 2, the X axis accelerometer
`output is affected by dynamic acceleration unlike the Z axis.
`Hence the solution is to use X axis and Y axis accelerometer
`values to detect the swinging motion instead of the Z axis
`accelerometer values.
`
`Forward-backward movement when µ=O a
`<n/18
`
`As illustrated in Figure 6, even though the Y axis of the
`accelerometer must be Og, the values are observed to be
`fluctuating.
`
`Swinging Motion Characteristics
`A breakdown of the forward-backward swinging motion
`(with µ equal to O radians and a is less than 1t/2 radians)
`signal is show in Figure 7.
`
`Break down of non·continuous forward-backward
`motion
`2.5 ~ - - - - - - - - - - - - - -
`
`"' QI
`::s
`ni
`>
`
`QI)
`
`1.5
`
`1
`
`0.5
`
`0
`
`-0.5
`
`-1
`
`-1.5
`
`-
`
`-
`
`X-axis
`
`z-axis
`
`1.5
`
`Ill
`Ill
`J
`iii
`) 0.5
`00
`0
`
`-0.5
`
`·1
`
`Seconds
`
`-
`
`X-axis
`
`Figure 5. Z axis and X axis accelerometer g values when
`forward-backward swing is performed
`
`3.1.
`
`CHARACTERISTICS OF SWING MOTION
`
`The developed algorithm utilizes the following basic
`characteristics of the swinging motion for identification.
`
`Accelerometer Signal Noise
`The augmented signal of the X and Y axis accelerometer
`will be twice as noisy as the original individual signals,
`because the accelerometer is not noise free. Hence the X
`and Y axis of the accelerometer must be evaluated
`individually. Figure 6 shows X axis and Y axis
`accelerometer g values when forward-backward swing
`motion is performed with µ equal to O radians and a is less
`than 1t/2 radians (90 degrees).
`
`Wave motion
`performed
`
`Idle
`
`Wave motion
`performed, opposite to
`the previous motion.
`
`Figure 7. Break down of a non-continuous forward(cid:173)
`backward swing motion with µ equal to O radians and a less
`than 1t/2.
`According to Figure 7, the circled acclivity shows the swing
`motion occurring to one side while the circled declivity
`shows the swing motion occurring to the opposite side of
`the initial swing motion. When µ is between O and 21t, X
`and Y axis signal values can merely infer the direction of
`the swinging motion relative to the previous swing.
`
`Y axis noise when forward-backward
`movement is performed
`
`Continuous forward-backward movment
`
`1.5
`
`1.5
`
`1
`
`-0.5
`
`-1
`
`-
`
`-
`
`X-axis
`
`Y-axis
`
`Seconds
`
`1
`"' 0.5
`QI .a
`"' >
`
`0
`
`QI)
`
`-0.5
`
`-1
`
`-
`
`X-axis
`
`-"
`
`c:i l ~ 1
`
`0 H
`
`,- N
`
`~ I•
`C
`
`"
`
`ti:
`
`I•
`
`")
`
`iJ
`
`Seconds
`
`Figure 6. Noise generated when forward-backward
`movement swing motion with µ equal to O radians and a less
`than 1t/2 radians.
`
`Figure 8. Continuous forward-backward swing motion with
`µ equal to O radians and a less than 1t/2.
`In order to find the exact direction of the swinging motion,
`the relationship between the reference axes and the
`
`Authorized licensed use limited to: University of Southern California. Downloaded on October 14,2023 at 00:53:08 UTC from IEEE Xplore. Restrictions apply.
`
`275
`
`Petitioner Samsung Ex-1049, 0004
`
`

`

`accelerometer axes must be established by using the value µ
`and the starting position of the electronic bell. When the
`swinging motion is performed continuously, the circled
`acclivity and circled declivity in Figure 7 merges together to
`create a sinusoidal swing as shown in Figure 8.
`
`Different values ofµ (rotation around reference Z)
`When the value of µ is not 0, 1t/2, 1t and 31t/2 accelerations
`are measured in both X and Y accelerometer axes. In this
`case the axis which produces the largest acceleration values
`is used to decide the swinging motion, as shown in Figure 9.
`
`Continuous forward-backward µ=n/18 a<n/2
`
`2.5
`2
`1.5
`1
`"'
`~ 0.5
`~ 0
`-0.5
`-1
`-1.5
`-2
`
`QI)
`
`..
`...
`0 ( ..
`
`I
`
`' :) s:t
`
`ftj n ..
`
`--
`
`lo
`
`~
`
`1•
`
`(
`
`,..
`
`r. ~ Ir
`ti'
`
`('I
`
`~
`
`-
`
`-
`
`X-axis
`
`Y-axis
`
`Seconds
`
`Figure 9. Continuous forward-backward swing motion with
`µ equal to 1t/18 radians and a less than 1t/2.
`But when µ is 1t/4, 31t/4, 51t/4 and 71t/4 both X and Y
`accelerometer axes will generate acceleration values which
`are equal. In this case the signal of the X axis is used to
`decide the swing motion as shown in Figure 10.
`
`Continuous forward-backward µ=n/4 a<n/2
`
`2
`
`1.5
`
`1
`
`~ 0.5
`ni
`~ 0
`
`-0.5
`
`-1
`
`-1.5
`
`-
`
`-
`
`X-axis
`
`Y-axis
`
`Seconds
`
`Figure 10. Continuous forward-backward swing motion
`with µ equal to 1t/4 radians and a less than 1t/2.
`
`SWING DETECTION ALGORITHM
`
`3.2.
`The algorithm developed is capable of detecting swinging
`motion when µ is between 0 and 21t radians (rotation along
`the reference Z axis from the ideal position).
`
`Pseudo Code
`
`2.
`
`3.
`
`1. Perform median filtering (windows size is 3) on X and
`Y axis accelerometer values.
`ls Acclivity or declivity detected on X or Y axis? lf yes
`go to step 3 else go to step 5.
`ls the acclivity or declivity detected on X and/or Y axis
`larger than threshold? If yes got to step 4 else go to step
`5.
`ls the amplitude of the acclivity or declivity detected on
`X and/or Y axis not twice smaller than the amplitude of
`the previous acclivity or declivity in the respective
`axis? lf yes go to step 5 else go to step 5.
`5. Record the activity (acclivity, declivity or no activity)
`occurred in X and Y axis.
`6. Cross compare the amplitude of the acclivity or
`declivity occurred in X and Y axis. Choose the axis
`showing the highest amplitude to decide the swing
`motion.
`
`4.
`
`Median filtering is used to reduce the noise generated by the
`accelerometer and also to remove noise due to slight
`unconscious hand movements. Acclivity or declivity
`generated on an axis means a swinging motion has
`occurred. Intra axis amplitude comparison in step 4 of the
`pseudo code is done to reduce false triggers. Finally a cross
`axis amplitude comparison of the acclivity or declivity
`generated is performed in order to negate the effects caused
`by rotation along the Z reference axis (µ is not equal to 0).
`
`4. EXPERIMENTATION RESULTS
`
`The recognition accuracy of the proposed swinging motion
`detection algorithm was tested on an embedded processor
`system consisting of a microcontroller (Texas Instrument
`CC2510 (16]) powered by a 7. 7V Lithium-ion battery that is
`voltage regulated to 3.3V operating voltage. The embedded
`processor system was used to build a wireless portable
`interactive device (PID) featuring various input and output
`modalities. The
`input peripherals
`include a 3-axis
`accelerometer (MMA7260 from Freescale) and a serial(cid:173)
`accessed micro-SD memory card. The output peripherals
`include six tri-color LEDs and a mini vibration speaker. The
`PID activates these peripherals when required by enabling
`the appropriate CC2510 component (USART, DSM, ADC).
`The hardware components of the PID were fitted to the
`cuboid acrylic housing section (see Figure 2) of the
`electronic bell.
`
`Four users tested the swing motion recognition accuracy
`when µ is 0, 1t/4, 1t/2, 31t/4, 1t, 51t/4, 31t/2 and 71t/4. The PIO
`was programmed to generate a sound and light up the LEDs
`when swing motion performed. According to the direction
`of the swinging motion (relative to the previous swing
`
`Authorized licensed use limited to: University of Southern California. Downloaded on October 14,2023 at 00:53:08 UTC from IEEE Xplore. Restrictions apply.
`
`276
`
`Petitioner Samsung Ex-1049, 0005
`
`

`

`motion) a different sound and a different LED was light up.
`Table 1 show the number of swinging motion recognized
`accurately out of 10 swinging motions performed.
`
`User2
`
`User3
`
`User4
`
`Table 1: Number of swinging motions recognized accurately
`for different µ values
`User 1
`10
`
`10
`
`10
`
`10
`
`µ = 0
`µ = ,r/4
`
`µ = ,r/2
`
`µ = 3,r/4
`
`µ = 1t
`µ = 5,r/4
`
`µ = 3,r/2
`
`µ = 7,r/4
`
`9
`10
`
`10
`
`7
`
`9
`10
`
`9
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`8
`10
`
`7
`6
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`
`10
`
`User 2 and User 4 achieve 100% swinging motion detection
`accuracy because the swing motion is done rapidly. User 1
`and User 3 performed the swinging motion slowly causing
`detection misses to occur. The average swinging recognition
`accuracy is 95.3% for all four users.
`5. CONCLUSION AND FUTURE WORK
`There are many accelerometer gesture classification
`methods namely, HMM, DTW, SVM, machine learning
`algorithms
`and
`threshold. Threshold-based
`gesture
`classification methods require Jess computational resources
`and do not require a training phase. We have proposed a
`novel orientation-free
`threshold-based swing detection
`algorithm for a physical construction which is similar to a
`mechanical bell. Four users tested for swinging motion
`recognition accuracy for different µ (rotation around the Z
`axis) values and the average recognition accuracy is 95.3%.
`
`incorporate a gyroscope or a
`to
`is
`The next step
`magnetometer in order to find the value of µ. This will
`enable the swing motion detection algorithm to differentiate
`the X and Y axis movements. Being able to differentiate the
`different swing gestures will permit the electronic bell to
`generate different notes or timbre depending on the manner
`the bell is swung.
`
`6. REFERENCES
`[1] Rokade R., Doye, D., and Kokare, M. "Hand Gesture
`Recognition by Thinning Method", International Conference on
`Digital Image Processing, pages 284 287, 7 9 Mar 2009.
`[2] Mariappan
`R.
`"Video
`Gesture
`Recognition
`System"International Conference
`on
`Conference
`on
`Computational Intelligence and Multimedia Applications, 2007,
`pages 519 521
`[3] Raheja J.L., Shyam R., Kumar, U., and Prasad, P.B.
`"Real Time Robotic Hand Control Using Hand Gestures Machine",
`Second International Conference on Learning and Computing
`(ICMLC), pages 12 16, 2010.
`
`[Online]
`
`Inc
`
`[Online]
`
`[4] Jangwoon Kim., Jaewan Park., HyungKwan Kim., and Chilwoo
`Lee. "HCI(Human Computer Interaction) Using Multi touch
`Tabletop Display",
`IEEE Pacific Rim Conference
`on
`Communications, Computers and Signal Processing, pages 391
`394, 2007.
`[5] Jing Yang., Eun Sook Choi., Wook Chang., Won Chui Bang.,
`Sung Jung Cho., Jong Koo Oh., Joon Kee Cho., and Dong Yoon
`Kim. "A novel hand gesture input device based on inertial sensing
`technique", 30th Annual Conference of IEEE Industrial Electronics
`Society, pages 2786 2791, vol 3, 2004.
`[6]
`Apple,
`(2011)
`Apple
`http://www.apple.com/iphone/
`Nintendo
`[7]
`Nintendo,
`(2011)
`http://www.nintendo.com/wii/console/controllers
`[8]Joselli, M., and Clua, E. "gRmobile: A Framework for Touch
`and Accelerometer Gesture Recognition for Mobile" 8'h Symposium
`on Games and Digital Entertainment (SBGAMES), Brazil , pages
`141 150, 8 10 Oct 2009,
`[9] Aki, A., and Valaee, S. "Accelerometer based gesture
`recognition via dynamic time warping, affinity propagation, &
`compressive sensing ", IEEE International Conference on
`Acoustics Speech and Signal Processing (ICASSP), Dallas, Texas,
`USA, pages 2270 2273, 14 19 March 2010.
`[10] Jiayang Liu., Zhen Wang., Lin Zhong., Wickramasuriya, J.,
`and Vasudevan, V. "uWave: Accelerometer based personalized
`gesture recognition and its applications", IEEE International
`Conference on Pervasive Computing and Communications, pages
`1 9, 13 March 2009.
`[11] Zhenyu He., Lianwen Jin., Loon Zhen., and Jiancheng Huang.
`"Gesture recognition based on 3D accelerometer for cell phones
`interaction", IEEE Asia Pacific Conference on Circuits and
`Systems (APCCAS) , pages 217 220, 30 Nov 3 Dec 2008.
`[12] Jun qi Kong., Hui Wang., and Guang quan Zhang. "Gesture
`recognition model based on 3D accelerations", 4th International
`Conference on Computer Science & Education ICCSE '09, pages
`66 70, 25 28 Jul 20()1).
`[13] Helmi, N., and Helmi, M. "Applying a neuro fuzzy classifier
`for gesture based control using a
`single wrist mounted
`accelerometer", IEEE International Symposium on Computational
`Intelligence in Robotics and Automation (CIRA), pages 216 221,
`15 18 Dec 2009.
`[14] Gonzalo Bailador., Daniel Roggen., Gerhard Treister., and
`Gracian Trivino ;."Real time gesture recognition using continuous
`time recurrent neural networks", Proceedings of the /CST 2nd
`international conference on Body area nenvorks, June 2007.
`
`"A Single
`[15] Rahul Parsani., and Karandeep Singh.
`Accelerometer based Wireless Embedded System for Predefined
`Dynamic Gesture Recognition", Proceedings of the First
`International Conference on
`Intelligent Human Computer
`Interaction, 2009.
`Instrument.
`Texas
`[16]
`Texas
`Instrument.
`http://focus.ti .com/lit/ml/swrb007 /swrb007 .pdf
`
`[Online].
`
`Authorized licensed use limited to: University of Southern California. Downloaded on October 14,2023 at 00:53:08 UTC from IEEE Xplore. Restrictions apply.
`
`277
`
`Petitioner Samsung Ex-1049, 0006
`
`