Analysis and Selection of Features for Gesture Recognition Based on a Micro Wearable Device

Transcription

1 (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., 0 Analysis and Selection o Features or Gesture Recognition Based on a Micro Wearable Device inghui hou, Lei Jing, Junbo Wang, ixue Cheng. Graduate School o Computer Science and Engineering. School o Computer Science and Engineering University o Aizu Aizu-Wakamatsu, Japan Abstract More and More researchers concerned about designing a health supporting system or elders that is light weight, no disturbing to user, and low computing complexity. In the paper, we introduced a micro wearable device based on a tri-axis accelerometer, which can detect acceleration change o human body based on the position o the device being set. Considering the lexibility o human inger, we put it on a inger to detect the inger gestures. kinds o one-stroke inger gestures are deined according to the sensing characteristic o the accelerometer. Feature is a paramount actor in the recognition task. In the paper, gestures eatures both in time domain and requency domain are described since eatures decide the recognition accuracy directly. Feature generation method and selection process is analyzed in detail to get the optimal eature subset rom the candidate eature set. Experiment results indicate the eature subset can get satisactory classiication results o 90.08% accuracy using eatures considering the recognition accuracy and dimension o eature set. Keywords Internet o Things; Wearable Computing; Gesture Recognition; Feature analysis and selection; Accelerometer. I. INTRODUCTION Internet o Things (IoTs) has become a hot topic in the computer science ield, which indicates that all objects in the environment like human, home appliances, building, and service equipment can be sensed, identiied, even controlled via the internet. IoTs will promote many development o application system, such as health supporting system or elder. Many countries are acing a serious society issue o population ageing. One common trend is more and more elders living alone and less able to beneit rom the care and supporting that might be available in a large household. Investigation rom World Health Organization indicates, in Japan, the proportion o people living in -generation households has allen rom % in 98 to 0.% in 00. Health care both physical and mental becomes an important problem in current society. Health supporting system has been studied widely in recent years [] []. Two kinds o main supporting way ocus on speech-based communication and activity-based recognition. The ormer provides a direct and eective way to know users intension, which has been used in the hospital and household [] []. However, voice signal is sensitive to environment sound such as a TV being on, so that sometimes hard to pick out useul speech signal made by a user. Even under certain circumstance, the user may too weak to make a voice to call or a help. Activity recognition provides an active and undisturbed way or elderly care. For example, i an elder person alls down, the system o alling recognition can send message automatically asking or help. Among the human activities, inger gestures are the most lexible ones. In our daily lie, most o works are perormed by hands. Gesture recognition is signiicant or learning user behavior, realizing or device control, and getting user intention. In the paper, we designed a wearable device with an accelerometer to detect inger gestures. Based on the accelerometer characteristic, a variety o inger gestures are deined in a D space. Particularly, gestures eatures are studied both in time domain and requency domain considering the paramount importance o eature generation in recognition task. Each kind o candidate eatures and their combination are analyzed based on stepwise regression algorithm to orm a eature vector or accurate gesture classiication and computing complexity control. The paper is arranged as ollows. The section Ⅱ introduces the related work about activity recognition and eature analysis. The section Ⅲ outlines the prototype system o gesture detection, and gives the gestures deinition and data collection. The section Ⅳ describes the eature analysis method in detail including eatures generation process and eatures selection algorithm based on stepwise regression. The section Ⅴ gives the experiment and evaluation on eature generation and selection. Conclusion is given in the last section. II. RELATED WORKS Two types o system are mainly used or activity recognition. One is ixed device-based recognition system, and the other is wearable device-based detection system. Fixed device-based techniques have been applied widely into various ields by dierent devices such as camera, computer vision system, and so on [] []. The kind o system provides an application way o no burden to users. However, some people do not like the way o being supervised, and some private spaces are inconvenient to be set a camera like in bathroom. Moreover, some issues have to be considered including whether environment actor is it or monitoring or P a g e

2 (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., 0 not such as surrounding light and blind corner o a camera, image processing speed and delay, inormation loss o -D object projecting to a -D image, and conusion o multiple users in same background and so on []. With the development o micro-electrical technology, micro wearable devices are penetrating into our lie. They can be attached on human body to obtain user inormation directly, typically using RFID and sensor. RFID technology is attractive or many applications since it can detect user s situations by simple ID and location inormation []. But it is diicult to detect motion. Wearable sensors have shown their capability or activity recognition. Some research set sensors in dierent position o body to detect human daily activities. In [8] the system sets ive accelerometers on hip, wrist, arm, ankle, and thigh or classiying 0 daily activities such as walking, sitting & relaxing, brushing teeth, bicycling, etc, and got the recognition accuracy ranging rom % to 9% or dierent activities. In [9] a large realistic data are collected rom many dierent sensors (accelerometers, physiological sensors, environment sensors, etc.) to recognize activities like lie, row, walk, etc. with accuracy o 80% over. However, to our knowledge, most o the researches seldom ocus on the tiny activity recognition like inger gesture. Finger is one o the most lexible body parts. Most o works in our daily lit are accomplished by it. Thereore, detection o inger gestures not only helps to know user current behavior, but also relect user intension and carry on some operations. Data glove, as an interactive device worn on the hand to sense gestures, has been applied in the environment o virtual reality []. However, the gestural interace required user to wear a cumbersome device to connect with external. It is inlexible and inconvenient or daily operation. Moreover, most o above researches have no explained on the process o eatures generation and selection while the eatures are o paramount importance in any recognition works. Based on dierent sensing device, some researches get eatures directly rom time-varying signal [0] [] and with requency analysis [8] [9]. Some preer to wavelet analysis to obtain both spectral and temporal inormation [] []. However, it is not be illustrated why the eatures are necessary, i they can be substituted on others, what will happen i adding or deleting one o them. In this paper, we introduce a wearable device in our previous research named Magic Ring. It can be used to detect kinds o predeined one-stroke inger gestures based on a -axis accelerometer []. A verity o gesture eatures are extracted or classiication evaluation. However, the process o eature selection is a lack. In this paper, we ocus on the eature analysis method including eature generation, eature selection, and eature evaluation taking the wearable sensor and its receiver as a prototype o the inal target devices. III. PROTOTPE INTRODUCTION AND DATA COLLECTION A. Prototype Structure The system is a ring shape sensing device based on a -trial accelerometer, MMAL rom Freescale Semiconductor, Inc. In its two sensitive scales o ±.g and ±.8g, ±.g is adopted to detect all predeined gestures. Excepting or the sensing unit, data processing unit is used or A/D conversion and simple digital signal processing; and transmitting unit is or acceleration data transmission and communication. The system can be worn a inger with no much disturbing to daily activity as shown in Fig.. Figure. Sensing system on the inger B. Gestures Deinition In the paper, the purpose o gesture recognition is to learn user simple intension and urther to apply the gestures into daily lie like controlling home appliances or calling or help. Thereore the gestures should be easy to be done or reducing physical load and easy to be understood and learned or reducing conscious load. Combining the way o controlling appliances and the characteristic o -axis accelerometer, kinds o dynamic one-stroke gestures are designed. One-stroke reers to dynamic gestures which are perormed no more than one degree o reedom in one direction. For example, pushing a button, turning a knob, and pointing to a picture can be regarded as one-stroke gestures. The tri-axis accelerometer can detect the acceleration change o three directions in space as shown in Fig.. Figure. Sensing direction o a tri-axis accelerometer Considering the tri-axis characteristic o the accelerometer and requirement o inger gestures, kinds o one-stroke inger gestures are deined as shown in Fig.. The kinds o gestures can be divided into pairs o gestures in, and axis and pairs o gestures in, and axial plane. These gestures are named as Crook and Unbend in axis, Finger L-Shit and Finger R-Shit in axis, Finger Up and Finger Down in axis, Wrist L-shit and Wrist R-shit in plane, L-Rotate and R-Rotate in plane, Wrist Up and Wrist Down in plane. The modes o motion or the six pairs o gestures are shown in Fig.. P a g e

3 Crook L-Rotate Wrist R-Shit Unbend Figure. Modes o motion or the six pairs o gestures C. Data Collection The system is attached on the middle phalanx o oreinger since it is the most lexible in all ingers. Gestures data is collected as sampling 0Hz. The digital signal is stored in a PC or data analysis, eatures extraction and gestures classiication. 0 students in the university ( males and emales, average age.8±.8) volunteered or the experiment o data collection under the supervision o a researcher. They are required to perorm the predeined gestures by oreinger in a natural and relax way. The gestures started in horizontal and static state, ended with static state. Each gesture was repeated times per people and 00 times or 0 people totally. IV. FEATURES ANALSIS For a inger gesture, it can be expressed quantitatively as a digital signal based on the sensing inormation. The eatures o the signal can indicate the type o a gesture and is useul or recognizing the gesture. A signal can be identiied with various eatures. Thereore the eatures analysis is vital or identiying the signal. Roughly speaking, the more eatures are used, the higher accuracy may be achieved, but higher complexity the recognition is. However, not each eature can be used or distinguishing the gesture rom others, e.g. or the two signals in Fig., it is the acceleration change in axis rom two dierent kinds o inger gestures. The eature o signal energy or mean can distinguish them, but the peak value as a eature is ailed to recognize one rom the other. Moreover, the number o eatures may not have direct relation with identiication eects. Although it is dierent that each eature and their combinations contribute to recognition accuracy, that do not mean the more eatures are, the better recognition accuracy is. R-Rotate Wrist Up Finger Up Wrist Down Wrist L-Shit (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., 0 Finger Down Finger L-Shit Finger R-Shit gesture gesture Figure. Two signals with same max value Furthermore, the high dimensionality o eatures increases computing cost or some recognition algorithm. Thereore, analysis and selection o proper eatures o gestures to be recognized has to be perormed to get the optimal eature vector/set or the balance between acceptable recognition rate and computing complexity. The eature analysis process is, irst, to extract signals o target gestures; second, based on the signals, generate candidate eatures; and inally, selection o proper eatures rom the candidates to orm eature set or recognition task. The paper, taking an accelerometer as an example, gives the eature analysis and selection procedure o inger gestures with an accelerometer. A. Extraction o Gesture signal Extraction o gesture signal reers to get the section related to the target gesture rom successive signals, namely to detect start and end point o the gesture. Since the gesture signal shows a dynamic change trend rom a static state to a dynamic activity, and then back to static, thereore the short time energy (STE) o signal is considered to distinguish the dierent states. When STE in a sliding window is higher than a level, we think a gesture start to be perormed, until the STE becomes lower than the level. We recorded the duration o each gesture per subject. Results show it roughly ranges rom 00 to 800ms. Due to the sampling rate is 0Hz, the window size is compromised within 0 samples with 0% overlap between two continuous windows. B. Feature Generation Basically, eature can be divided into two types: eatures in time domain and requency domain. Features in time domain show the signal characteristic varying with time. Typical eatures are shown in Fig.. Frequency eatures are used to capture the periodic nature o a sensing signal or distinguishing some repetitive activities like walking and running. The typical requency-domain eature is shown in Fig.. Excepting or the eatures mentioned above, others can be extracted according to dierent classiication objects like Euclidian distances, similarity, and so on. P a g e

4 (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., 0 Time Domain Mathematical Features Statistical Features Max Min Median Dierence Mean Variance Standard Variance Energy Entropy Speciically, let F,,, n be the candidate eature set o using or classiication design. The elements in the i set are descending order by the signiicance level. Let denotes the signiicance level o eature i, or the eature set F S S S n,. We select a classiier as the model, then adding the eature one by one rom to test the classiication accuracy until reaching a satisactory result. S Frequency Domain Others Correlation ero-crossing Rate Figure. Typical time-domain eatures Fourier Transormation Wavelet Transormation DC Component Spectral Peak Spectral Centroid Spectral bands Spectral Energy Spectral Entropy Wavelet Coeicients Figure. Typical requency-domain eatures All above mentioned eatures can be generated based on the corresponding mathematical or statistical method. However, not all eatures are necessary or a classiication system. A direct reason is whether a eature may bring good classiication signiicance. Even though both two eatures have good classiication capability, there is maybe little gain when they are collected into a eature vector due to a high mutual correlation []. Another reason is the computational complexity. The number o eatures directly decides the dimensionality o classiier parameter. Thus a eature vector as small as possible is desired both in training process and in classiying process. C. Feature Selection The eature selection is very crucial, which helps us use as less as possible o eatures to ind out as much as possible o classiication inormation then to get the optimal recognition perormance. Here we try to ind an optimal eature vector to reach the balance between the acceptable recognition rate and computational complexity. In practical application, a satisactory eature vector instead o an optimal vector. In the paper, we adopt the algorithm o eature selection, stepwise regression. It is a greedy algorithm that includes a regression model in which candidate eatures are evaluated automatically. Forward selection and backward elimination are two main approaches to achieve the algorithm. The ormer represents the procedure starting with no eatures in the model, then trying to add one into the eature vector one by one until them reaching a satisactory signiicance. The latter is contrary to the ormer, which including all candidate eatures in the model, and deleting one that is no signiicant. Here, we used the orward selection way. V. EPERIMENT AND EVALUATION A. Features Generation 0 subjects completed the total kinds o inger gestures under the supervision o researcher. Beore each one-stroke inger gesture, it requires inger in a horizontal and static state. When inishing a gesture, inger should maintain ending state and stillness. In other words, the inger is dynamic just during a gesture being perormed. Thereore, it is possible to identiy i a gesture happening using a threshold-based approach. Fig. and Fig. 8 shows two inger gesture signals Finger Let Shit and Finger Up, which is composed o three channels and indicates the acceleration change in three axis. From the extracted the gestures signals, various eatures can be generated both in time domain and requency domain as described in last section. For example the mean and standard deviation (sd) o each axis in Fig. and Fig. 8 can be computed as mean n n i0 x, i sd n n i0 ( x i mean ) where x denotes the sampling, and n denotes the number o sampling in the window. Other eatures can also be achieved by mathematical or statistical way. Although any signal eatures can be regarded as a candidate, in order to reduce the computing load, some obvious insigniicant eatures will be neglected. In our case, the inger gestures are one-stroke type, which means each gesture is aperiodic and instantaneous. Thereore, the eatures in requency domain are neglected. The process o eature generation can be expressed as: () Observe the signals o one gesture rom dierent subjects and try to describe it using some typical eatures. For example, in Fig., we may consider mean, energy, etc. () Observe the signals o variety o kinds o gestures to ind out eatures with the capability o distinguishing with others like sd o axis in above two igures. () Abandon some insigniicant eatures like amplitude o each axis in our case, because even i a gesture is perormed by same person, amplitude o each time will be great dierent due to the dierence o perormance speed. () Collect the eatures to orm a candidate eature set or eature selection. P a g e

5 (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., Figure. Acceleration curve o a inger gesture Finger Let Shit Finger Let Shit Finger Up Figure 8. Acceleration curve o a inger gesture Finger Up In our case, time-domain eatures o each axis are calculated including mean, standard deviation, energy, entropy, correlation o any two axes, dierence o peak and valley, and position o peak and valley in the time axis, totally 8 kinds o eatures to compose a candidate eature set. Each eature in the set consists o three elements in,, and axis, such as {mean, mean, mean}. B. Features Selection Features selection is to ind a satisactory eature subset rom the candidate eature set, so that to reach an optimal classiication accuracy and computing complexity control. It is crucial since it decides the classiication result directly. Forward selection algorithm o stepwise regression is adopted to test each eature and their combinations one by one. In the algorithm, a model is required to evaluate the eatures. For obtaining an objective evaluation results, here we select three basic classiiers o machine learning, C. decision tree (C.), Nearest Neighbor (NN), and Naïve Bayes (NB), as three test models to calculate the classiication accuracy o kinds o one-stroke inger gestures mentioned above, and the testing average o three classiiers (Avg) is employed as the inal evaluation result. First, each one in the candidate eature set is tested by three models, the evaluation results are ranged in a descending order as shown in Table. 8 TABLE. THE RESULTS OF SINGLE FEATURE EVALUATION Features C. NN NB Avg mean, mean,.%.%.08%.% mean energy, energy, 9.8%.9% 9.% 0.% energy sd, sd, 8.%.%.0% 9.% sd di, di, 8.9%.8%.9% 8.% di col(,), col(,),.%.8% 8.08%.% col(,) posval, posval,.8% 8.8% 8.8%.8% posval entropy, entropy,.8%.% 9.%.% entropy pospeak, pospeak,.%.0%.%.% pospeak Second, selecting the optimal eature rom the Table,, and combining with other eatures, the combinations will be recomputed based on the models. The results are shown in Table. TABLE. THE EVALUATION RESULTS OF COMBINING TWO FEATURES Features C. NN NB Avg.8%.%.9%.% 8.%.8% % 0.%.%.0% 8.9%.0% 9.9% 9.8% 9%.8%.0% 8.8%.%.8% 9.%.% 8.% 0.% 8.08% 8.%.%.% Third, repeating the above process o selection to get the best one,, as a basic eature combination, and then combining with rest eatures to reevaluate the signiicance o combined eatures. From Table to Table show the evaluation results with dierent combination. Each optimal combination in each Table will be the basic o next combination. TABLE. THE EVALUATION RESULTS OF COMBINING THREE FEATURES Features C. NN NB Avg.9%.8% % 9.% 8 8.% 9.%.% 0.%.0% 8.9%.%.9% 8.0% 8.0% 8.% 80.%.% 0.%.% 8.%.8% 8.08%.8% 9.% P a g e

6 (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., 0 TABLE. THE EVALUATION RESULTS OF COMBINING FOUR FEATURES Features C. NN NB Avg.9% 8.%.% 9.% 9.% 8.%.% 80.% % 88.% 8.% 8.% 9.% 8.8% 8.8% 8.8% 8.08% 8.% 8.% 8.% represented as pospeak>posval, pospeak>posval, and pospeak>posval % 80.00% 0.00% 0.00% TABLE. THE EVALUATION RESULTS OF COMBINING FIVE FEATURES 0.00% Features C. NN NB Avg 80.0% 8.00% 9.% 8.% % 8.9% 9.08% 8.% 8.% 88.9% 8.0% 8.9% 8 8.9% 8.% 8.9% 8.% 0.00% TABLE. THE EVALUATION RESULTS OF COMBINING SI FEATURES Features C. NN NB Avg % % % % % % % % % % % % 8 TABLE. THE EVALUATION RESULTS OF COMBINING SEVEN FEATURES Features C. NN NB Avg % % % % % 90.8 % 8. % 8.8 % Finally, the average classiication accuracy o combining all eatures, 8, is 8.%. Fig.9 shows the evaluation results under each kinds o combination. It can be seen rom the Fig. 9, with the increasing o the number o eatures, the classiication accuracy will increase. However, when eature combination reaches to some extent, such as ater, the accuracy has no obvious change. That indicates not the more eatures are, the better classiication accuracy is. Even under certain circumstance, large number o eatures will reduce the accuracy, which is the peaking phenomenon occurring or larger eatures. We can ind rom the Fig. 9, the best classiication result is about 8%. To improve the classiication, other eatures are considered to add into the candidate set. By observing the acceleration signals o each kinds o inger gesture, we ind, to a gesture signal, the acceleration change in each axis shows great dierence. For example the gesture Finger Let Shit in Fig., acceleration in axis shows intense change than and axis. While or Finger Up in Fig. 8, and axes show obvious luctuation. I taking sd as the luctuation level, the comparison o sd between two axis is generated adding into the candidate eature, which is sd>sd, sd>sd, and sd>sd. we name the eatures as relative eatures. In addition, or identiying the gestures with opposite direction, coming sequence o peak and valley in single signal is adopted, which is Figure 9. Recognition accuracy under dierent eature combinations Using the relative eatures, candidate eature set is regression evaluated again. Result show the recognition accuracy reach to 90.08% with the eature subset {mean, mean, mean, sd, sd, sd, sd>sd, sd>sd, sd>sd, pospeak>posval, pospeak>posval, pospeak>posval}. These relative eatures not only improve the recognition accuracy, but also reduce the computing complexity because o their alternative Boolean value. Besides, the system robust can be increased since a relative relationship can prevent classiier rom acting o the initial state. The inal recognition matrix is shown in Table 8 based on Nearest Neighbor classiier. VI. CONCLUSION In the recognition task, eatures are o paramount importance. In the paper, we ocus on the research o eature analysis and selection, which includes how to generate the candidate eature set based on the sensing inormation, how to evaluate each eature and their combinations, and how to select the optimal eature subset. Feature generation process need to observe the activity signal and initially select some eatures and neglect insigniicant one. Based on the candidate eature set, orward selection algorithm o stepwise regression is adopted to evaluate each eature and their combinations. The inal combination with good classiication signiicance is selected as eature subset or gesture recognition. Experiment result indicates the process o eature analysis and selection is easible to most o activity recognition research. In the uture, we plan to use our method to evaluate other kinds o sensing device and activity data. ACKNOWLEDGMENT We would like to thank the subjects who help us with the experiment in the research. P a g e

7 (IJACSA) International Journal o Advanced Computer Science and Applications, Vol., No., 0 TABLE 8. RECOGNITION MATRI FOR KINDS OF FINGER GESTURES BASED ON THE SELECTED FEATURE SUBSET Classiied as a b c d e g h i j k l 90.08% a=crook % b=unbend % c=finger Down % d=wrist Down % e=l-rotate % =Finger L-Shit % g=wrist L-Shit % h=r-rotate % i=finger R-Shit % j=wrist R-Shit % k=finger Up % l=wrist Up % REFERENCES [] S. Consolvo, P. Roessler, et al. Technology or Care Networks o Elders, IEEE Perv. Comp., vol., no., pp. -9, 00. [] Stanord, V. Using Pervasive Computing To Deliver Elder Care, IEEE Perv. Comp., vol., no., pp. 0-, 00. [] P. W. Jusczyk, and P. A. Luce, Speech Perception and Spoken Word Recognition: Past and Present, Ear Hear, vol., no., pp. -0, Feb. 00. [] A. aar, J. Overhage and C. McDonald, Continuous Speech Recognition or Clinicians, J. Am. Med. Inorm. Assoc., vol., no., pp. 9-0, 999. [] S. Mitra, and A. Tinku, Gesture Recognition: A Survey, IEEE Trans. on System, Man, and Cybernetics Part C: Applications and Reviews, vol., no., pp. -, 00. [] S. Helal, B. Winkler, et al. Enabling Location-aware Pervasive Computing Applications or the Elderly, In Proc. o First IEEE International Conerence on Pervasive Computing and Communications (PerCom 0), 00, Fort Worth, Texas, March, pp. -. [] J. R. Smith, K. P. Fishkin, et al. RFID-based Techniques For Human-activity Detection, Communication o ACM, vol. 8, no. 9, pp. 9-, 00. [8] L. Bao, S. S. Intille, Activity recognition rom user-annotated acceleration data, in Pervasive 00, Springer: Linz, Vienna. pp. -. [9] P. Juha, M. Ermes, et al., Activity Classiication Using Realistic Data rom Wearable Sensors. IEEE Trans. on Inormation Technology in Biomedicine, vol. 0, no., pp. 9-8, 00. [0] U. Maurer, A. Rowe, A. Smailagic, and D. Siewiorek,, Location and activity recognition using ewatch: A wearable sensor platorm, Ambient Intelligence in Everyday Lie, Lecture Notes in Computer Science, vol. 8, pp. 8-0, 00. []. hang,. Chen,. Li, and etc. A Framework or Hand Gesture Recognition Based on Accelerometer and EMG Sensors, IEEE Trans. on Systems, Man and Cybernetics, Part A: Systems and Humans, Vol., No., pp. 0-0, 0. [] T. G. immerman, J. Lanier, C. Blanchard, and etc. A hand Gesture Interace Divice, in Proc. o the SIGCHI/GI conerence on Human actors in computing systems and graphics interace, New ork, USA, 98. [] M. Sekine, T. Tamura, T. Togawa, and etc. Classiication o waistacceleration signals in a continuous walking record, Med Eng Phys. 000 May;():8-9. [] N. Wang, E. Ambikairajah, N. H. Lovell, Accelerometry Based Classiication o Walking Patterns Using Time-requency Analysis, in Proc. o 9th Annual Conerence o IEEE Engineering in Medicine and Biology Sciety, pp , Lyon, France, 00. [] L. Jing,. hou,. Cheng, and J. Wang, A Recognition Method or One-Stroke Finger Gestures Using a MEMS D Accelerometer, IEICE Trans. on Inormation and Systems, E9.D(), pp. 0-0, 0. [] S. Theodoridis and K. Koutroumbas, Pattern Recogniton, st ed. U.K. Elsevier, 009, ch., pp. -. AUTHORS PROFILE inghui hou received her B.E. degree and M.E. degree in Computer Science and Engineering rom Jiamusi University and anshan University, China in 00 and 00 respectively. Now she is pursuing her ph.d. Degree in University o Aizu. Her research is concerned with ubiquitous learning, wearable computing and pattern recognition. Lei Jing received his B.Eng. degree rom Dalian University o Technology, China, in 000, M.Eng. degree rom the anshan University, China, in 00, and Ph.D rom University o Aizu, Japan, in 008. Currently he is a special researcher at the University o Aizu. His research interests include sensor networks, wearable computing, and ubiquitous learning. Junbo Wang received his B.E. in Electrical Engineering and Automation and M.E. in Electric circuits & systems in 00 and 00, rom the anshan University, China, and received a Ph.D. degree in Computer Science at the University o AIU, Japan in 0. Currently, he is a visiting researcher in the University o Aizu. His current research interests include IoT, ubiquitous computing, context/situation awareness, and WSN. ixue Cheng received his B.Eng. degree rom Northeast Institute o Heavy Machinery in 98, his Master degree and Ph.D degree rom Tohoku University, Japan in 990 and 99, respectively. Currently, he is a ull proessor the School o Computer Science and Engineering, the University o Aizu, Japan. His current interests include distributed algorithms, ubiquitous learning, context-aware platorms, and unctional saety or embedded systems P a g e