Probabilistic Graphical Model based Personal Route Prediction in Mobile Environment

Appl. Math. Inf. Sc. 6 No. 2S pp. 651S-659S (2012) Appled Mathematcs & Informaton Scences An Internatonal Journal @ 2012 NSP Natural Scences Publshng Cor. Probablstc Graphcal Model based Personal Route Predcton n Moble Envronment Je-Mn Km 1, Haejung Baek 1 and Young-Tack Park 1 1 School of Computng, Soongsl Unversty, 1-1, Sangdo-dong, Dongjak-Gu, Seoul, 156-743, Korea Correspondng author: Emal:kmjemns@hotmal.com, baekhj@gmal.com, park@ssu.ac.kr Receved July 15, 2011; Revsed Aug. 15, 2011; Accepted Sep. 2, 2011 Publshed onlne: 1 January 2012 Abstract: Indvduals tend to follow ther own preferred paths when travelng to specfc places. Informaton on these routes could be utlzed to buld varous ntellgent LBSs. In order to predct a current user s route, varous approaches have been researched. In ths paper, we suggest a practcal approach to learnng users' route patterns from ther hstores and usng that nformaton to predct specfc routes. In cases where exstng routes overlap,.e., where parts of routes are the same, n a user's route model, t s dffcult to dentfy the user's ntended path. For more accurate predcton, frstly, we extract route patterns by adoptng mage processng. Secondly, we buld a stateobservaton model reflectng users' ntentons, based on route patterns, temporal features and weather nformaton. Our approach consst of four steps: recognzng regons for splttng routes nto trp segments, route pattern mnng, learnng users' route models and trp route predcton. Our method acheved a predcton accuracy of 96.4% n tests performed wth 15 smartphone users. Keywords: Data Mnng, Temporal Probablstc Model, Route Predcton, Locaton based Servce. 1 Introducton Smartphones ncorporate many dverse and powerful sensors. In partcular, the global postonng system (GPS) can easly observe and collect the trajectores of a person. When people travel n the real world, they leave record of ther locaton hstores n the form of GPS logs. The GPS logs provde a useful bass to learn the route patterns of people. Most people have personal route patterns, whch are comprsed of journeys of a repettve nature. Personal routes would be extremely useful n the doman of locaton-aware ntellgence servces, partcularly n applcatons such as group socal networkng based on the future locaton of users, predcton of personal behavors, predcton of ndvduals arrval tmes at specfc places and so on. These servces have motvated researchers to explore the possbltes of route predcton. Currently, there are many dfferent approaches to route predcton. A personal route predcton system proposed n [1] predcts the route that people wll take, adaptng a basc Markov model and 2nd Markov model. However, ths system only use a GPS coordnates sequence to learn route patterns. In the same manner, most exstng work does not ncorporate all avalable facts that could provde clues to dentfyng a route.. In ths paper, we propose a practcal approach to predctng the current transt route of a user, usng a probablstc graphcal model bult from hstorcal data. Our approach conssts of four sequental parts: decdng upon RSPs (Route Separaton Ponts), route pattern mnng, personal route modelng and route predcton. In the RSP-decson step, we fnd sgnfcant places that separate a route nto segments. Ths s done usng a heurstc algorthm, adaptng four flters to consder the velocty and densty of a GPS sequence, WF access ponts, and user actvty data. In the route pattern mnng, usng an mage-processng algorthm, we abstract users' route patterns from personal GPS hstores and detected RSPs. In the personal route modelng step,

256 Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... Fgure 1: The personal route predcton archtecture we buld a probablstc model that s used n route predcton. In cases where segments overlap (.e., some parts of routes are the same n a user's route patterns), t s dffcult to dentfy the user's ntended route. In general, when a person vsts a specfc place, routng s nfluenced by the user s envronment. Therefore, we consder usng not only coordnate sequences from users' GPS hstores, but also aspects of ther envronment hstores, ncludng temporal nformaton (day of the week, travel tme), users' actvtes and weather nformaton. In order to buld a model, we consder the relatonshp among varables. There s a transton relaton between values of segments because each route conssts of an ordered segment sequence. On the other hand, there are condtonal dependences between segments and other envronmental facts. In ths paper, we adapt the State-Observaton model [2]. In the personal route predcton step, we predct a user's current transt route, based on ther route patterns and probablstc models. Ths step ncludes four components: the canddate segment decson process, whch dentfes target segments to compare wth a user s current locaton, the smlarty calculaton between canddate segments and a user locaton, the segment valdaton process, whch decdes whether the most smlar segment s a user's current movng segment, and the route decson, based on current envronmental facts. In ths paper, we make an mportant assumpton. Many studes of users' actvty recognton have been conducted [3-5]. Usng dverse sensors offered by moble devces, these studes nfer users' actvtes, such as takng a bus, rdng the subway, walkng, runnng or standng. Therefore, the results of ths pror work are ncorporated nto our approach to buld a probablstc model of users' route patterns. 2 Personal Route Modelng and Predcton Our route predcton approach conssts of four major steps: the RSP decson, whch fnds sgnfcant places that separate a route nto segments; the route pattern mnng step, whch extracts the complete route patterns of user usng an mage processng algorthm [6,7]; the personal route modelng step, whch entals learnng the transton probablty relatonshps among segments, as well as the dependences between segments and envronmental facts, usng a State-Observaton model; the personal route predcton step, whch chooses the most sutable route based on a probablty model. To reduce the computatonal load on the smartphone, our approach has a clent-server structure. Fg. 1 presents the clent-server archtecture for personal route predcton. The clent sde (smart phone) gathers recorded data n order to buld a behavor model and route model. Then, the clent sends the collected data to the server. The server learns personal behavors, fnds RSPs usng four flters, extracts route patterns by employng an mage-processng algorthm and bulds a personal route model, based on a probablstc analyss. The learned behavors and personal route models are transmtted to the clent. Thus, the clent nfers a

Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... 256 user s behavor and predcts ther current transt route. The route pattern extracton and model buldng process consumes a large amount of computng tme and resources, and, further, t should be executed only rarely, thus these steps are performed on the server. On the other hand, the route predcton process should be executed n real-tme, thus ths step s performed on the clent (smartphone). 3 Route Separaton Ponts Recognton and Route Pattern Mnng In order to predct users' current transt routes, frst, we perform the route recognton process to separate a personal route nto segments. Adaptng four flters, each regon s separated by passng t through these flters. In ths step, route separaton ponts (RSPs) are decded upon by comparng a sngle parameter, assocated wth each flter, to velocty and densty values, whch are based on detected GPS coordnates, detected WF access ponts and a user's behavor durng an arbtrary perod of tme, t+. A RSP s categorzed nto a fxed area and a separaton area. A fxed area s a regon that a user remans n for a long tme, such as a home or offce. A separaton area s a space wthn whch a person's behavoral changes occur, such as a bus staton or subway. When a person reaches a separaton area, there are dfferences n ther velocty and actvty. Ths occurs as they transton between route segments that are dvded by ths area because they may board publc transt. When a person stays n a fxed area, ths area shows a hgh densty of GPS pont records and, further, the same WF access ponts are detected repeatedly. Therefore, separaton areas could be dentfed by the change n velocty and behavor, whle fxed areas could be dentfed by the densty of GPS ponts and detected WF access ponts n the area. Velocty Flter: It s assumed that ndvduals move at a constant speed between two consecutve GPS ponts, and that there s a reasonable speed range for ndvduals. The parameter values of the velocty flter nclude vel non, vel walk, and vel transt. Densty Flter: Ths s desgned to check for the presence of redundant poston data, whch s recorded when users are nsde specfc areas, such as buldngs. Gven a wndow sze, d, we frst calculate the centrod for each d-sequental-poston sequence n a regon. Then, the maxmum dstance between these centrods s calculated to estmate whether the regon contrbutes to a reasonable movement dstance. The parameter values of the densty flter nclude den non, den sep, and den fx. Behavor Flter: Ths flter assumes that ndvduals exhbt a sngle type of behavor at a gven pont n tme. The parameter values of the behavor flter nclude act stay, act walk, act bus, act car and act subway. WF Flter: The value of the WF flter reflects the comparablty of WF access ponts detected n a specfc regon. Therefore, ths flter ncludes the wf mat parameter value. As a result of testng, we concluded that the best precson for fxed area recognton s observed when we assgn wf mat a value of 0.85. The route pattern extracton process s dvded nto two parts n whch spatal nformaton and temporal nformaton are separately recognzed. The frst part entals representng locatons on a grd space and learnng only the geospatal nformaton of GPS logs usng mage-processng. The second part entals makng path graphs based on the learned spatal models and, further, learnng route patterns usng temporal factors of the GPS data. 3.1. Extractng Route Lne Grd mappng and lne generaton: Consder Even though a user follows the same routes, the produced logs rarely show the same values because of GPS's naccuracy (about 50m). In addton, GPS ponts n a real scale are sporadcally dstrbuted, whch makes t more dffcult to extract any pattern and also hghly ncreases the complexty of dong so. To solve these problems and obtan GPS error tolerances, we generalze GPS measurements usng a regular grd. The transformaton functon for GPS measurements from a sngle 2D Grd s defned as T : H 2 N(R 2 ), where H 2 s the real world space usng Haversne dstance and R 2 s the whole 2D Grd usng Eucldean dstance. We represent a space as a 400*400 twodmensonal, one-level grd, where each cell s less than 50 meters on a sde. Decdng upon the dmensons of a grd space s a crtcal ssue for the accuracy of trajectory representaton. The basc dea s to cluster trajectores wth smlar start and end ponts, and to remove trajectory regons that are not clustered; those regons that are seldom vsted by the user. We represent trajectores usng smple lnes, whch connect start ponts wth end ponts.

256 Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... Fgure 2: The process of lne generaton Then, each trajectory s smply characterzed by an angle, a start pont, an end pont, and the mdpont. If the angle and the three ponts (start, end and md) between trajectores are smlar, we consder the trajectores to be n smlar regons. The functon s desgned to evaluate the closeness of four features when comparng two trajectores. dst( I, I j ) d ( I, I j ) w dhcenter ( I. sm, I j. s jm' ) (1) dh ( I. s, I. s ) dh ( I. s, I. s ) start 1 Here, angular dstance, d θ, s the gap between the bearng angle of the startng and endng ponts of I and the bearng angle of I j, whle dh s the Harvesne dstance. Even though the GPS ponts project onto a much smaller grd space, the grd ponts are not connected, whch can be seen n panels (a) and (c) of Fg. 2. We make a connecton by constructng ntermedate ponts between the two ntal ponts, followng an nterpolaton process. For the connecton, we adapt the Brensenham lne algorthm, whch s an effcent and fast algorthm, to draw a lne on the grd space, graphcally. Panels (b) and (c) of Fg. 2 depct the results of generatng lnes between ponts based on the Brensenham method [8]. These lnes are called GPS lnes, n order to dfferentate them from route lnes, whch are learned from each GPS lne. Lne ntegraton and lne thnnng: To abstract the accumulated GPS lnes, we have adapted the thnnng approach n computer vson. We restate our problem as a skeletonzaton of j j1 end e j je' routes from an mage, whch s generated by ntegratng the GPS lnes of trajectores on a 2D grd. When we ntegrate the lnes, we use a thckenng technque, whch adds addtonal pxels to the orgnal lne. Ths smple technque allows our system to overcome the naccuracy of GPS logs, stemmng from GPS errors or readng ntervals. From the ntegrated mage, we extract pxel-wde route lnes usng a thnnng algorthm. A smple example s shown n Fg. 3. The accumulated mage could be consdered the result of summng mages of the trajectores. The pxels, (x, y j ), n an ntegrated mage, are counted usng both those trajectores that have a pxel (x, y j ) that s taken as a GPS lne as well as those trajectores that have a pxel (x, y j ) that s the neghbor of a GPS lne, even though the latter pxels are not actually on GPS lnes. Panel (a) of Fg. 3 smply shows ths mechansm, n whch the pxels surroundng the target pxel are on GPS lnes. Even though the target pxel s not on the lnes, we consder these pxels as components of user paths, check ther valdty and count them. We use the Zhang-Suen thnnng algorthm [7] for the skeletonzaton, to extract route lnes from an ntegrated mage. The Zhang-Suen algorthm has the advantage of processng speed as t uses a parallel method. However, ths algorthm has some weaknesses, such as the fact that t produces that skeletons contan artfacts, lke neckng, tal and lne fuzz. Further, the skeletons are not one pxel n wdth. To solve these problems, Parker has ntroduced a hybrd thnnng algorthm [6]. The Fgure 3: Pxel wde route lne extracton example

Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... 255 algorthm merges three methods: Stentford's preprocessng for reducton of defects, Zhang- Suen s thnnng algorthm and Holt's starcase removal as a post-processng step, to produce a onepxel-wde skeleton. Panel (b) of Fg. 3 shows the result of thnnng, whch also reveals starcase problems. To extract one-pxel skeletons, we perform Holt's star removal method. The one-pxel skeleton produced by Holt's algorthm s presented n panel (c) of Fg. 3. In route pattern learnng, t s mportant to mantan topologes on the mage, however, smplcty s more mportant. As such, we add the post processng algorthm to remove relatvely short lne fuzz as well as crculatons that have a short dstance, below a threshold. 3.2. Learnng Route Pattern We construct route patterns as graphs from the produced lnes (skeletons) and RSPs. Frstly, structural features of skeletons become nodes, such as termnal ponts, turnng ponts, whch have less than two adjacent pxels or more, and RSPs. The connected pxels between nodes become edges. A lne graph, LG, s a undrected graph that s a par, (V, E), where V = {v : v P} and E {{, j}:, j V and {,j} route lnes}, whch s a set of unordered pars. Fg. 4 shows a graph generated from a thnned mage. We adapt a breadth-frst search algorthm to make a lne graph. The basc dea s to recognze nodes, choose a start node, s, explore every edge of the nodes, put the encountered nodes nto a queue, Q, and to then repeat ths routne untl every node and edge has been checked. Usng the produced lne graph, we learn the trajectores of a user wth temporal nformaton. We project each trajectory wth a tmestamp nto a lne graph, LG, usng the smlartes between the pxels of each trajectory and the pxels of the LG. Ths method of comparng pxels s relatvely easy, but also produces a local mnmum problem; a few pxels of a gven trajectory may be smlar. To avod the local mnmum problem, we perform a hybrd approach. In cases where the pxels of trajectores correspond to nodes of a LG that has at least three alternatves, we use a more sophstcated smlarty functon, whch s explaned n chapter 5. In the other case, where the pxels are not near nodes and have less than two alternatves, we just use the smlarty between pxels. Fgure 4: An example of a lne graph 4 Buldng Personal Route Model In cases where overlappng segments exst (.e., where parts of routes are the same n a user's route patterns), t makes dffcult to dentfy the user's ntended route. In order to solve ths problem, we adapt a probablstc analyss to consder facts reflectng a user's ntenton. Consderng two relatons, we buld the personal route model based on a State-Observaton model. Frst, we buld a transton model for the transton probablty between segments. Generally, the probablty of a person travels n a segment s nfluenced by the prevous segment because each route conssts of an ordered segment sequence. Therefore, the transton model s defned by the followng equaton n P( s) P( s s ) P( s Prev s ) (2) 1 Gven a user s current segment, s, or a seres of segments, (s 1, s 2, s 3,..., s k ), the predcton of the next segment that the user wll vst, s 0, s determned by ths jont probablty. P(s, s ) s the probablty that s and s occur. Ths s the probablty that a person vsts s and s. Second, we buld an observaton model for the condtonal probablty between a segment and set of envronment facts. Generally, the probablty that a person travels n a segment s nfluenced by the partcular tme, day of the week and weather. Furthermore, users' behavors are dfferent n dfferent segments. Therefore, these facts could provde a bass for understandng a user's ntended route. The observaton model s defned by the followng equaton. n 1,..., on ) P( s) P( o s) 1 P( s, o (3) Gven a user s current segment, s, and a seres of observed values reflectng envronmental facts, (o 1, o 2, o 3, o 4 ), the predcton of the next segment, s 0, s determned by the probablty of each o and the condtonal probablty, P(s o ), for each o. P(s o ) s the probablty that s and o occur. Ths refers to the chances of t beng a partcular tme, partcular

252 Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... day, the weather beng n a partcular state and the users performng a partcular behavor when they vst s. The followng table shows the values of observed varables. Table 1: The values of each envronmental varable Envronment Value varable tme of day mornng, noon, nght day of week weekday, holday weather sunny, rany, snowy behavor stay, walk, nbus, nsubway, ncar 5 Personal Route Predcton The route predcton process s composed of canddate segments selecton from a set of target segments, a smlarty calculaton, whch fnds the best connected segment (BCS), a segment valdaton process, whch decdes whether the best smlarty segment s a predctable segment, and a route decson based on current envronmental facts. In route decson step, consderng the user's ntentons, we adapt a probablstc method to resolve the overlappng routes problem. We use both a segment of GPS logs as well as the average travel tme, travelng days, user actvty and weather nformaton, n order to determne the user's ntenton. Because recordng a segment of GPS logs for use n route predcton consumes sgnfcant smartphone battery lfe, we set a tme nterval of ten seconds for the recodng of GPS ponts. A segment of GPS logs, G', s a sequence of locatons, from g 1 to g m. Fndng a current route wthn a route model consumes much computaton tme n the smlarty calculaton. Therefore, t s mportant to reduce the number of segments that are to be compared, whch we call target segments. The canddate segment decson process fnds target segments that exst wthn a bounded segment of GPS logs. The boundary threshold, whch we call the MB (Mnmum Bound), s a boundng box that reduces the search space of the route model. Pvotng on each pont n a segment of GPS logs, MBs are calculated by λ. Consderng our context of locaton-aware moble servces, we determne λ usng GPS error rates and the walkng speed of each user. Therefore, λ s determned by the maxmum devaton between each trajectory that s used n buldng a route model, a learned route pattern and the average dstance per perod (10 seconds). Fgure 5: An example of canddate segments selecton 2 2 (max_ devaton ) ( Ave _ dst / tme) (4) The smlarty calculaton process decdes the BCS that has the hghest smlarty wthn a canddate set, for a segment of GPS logs. The smlarty s calculated by a functon that consders the dstance from the ponts of each segment n the route model to each pont n the segment of GPS logs. We defne dst, the dstance between a pont of GPS log, g, and a segment, s m = {s 1, s 2,..., s j }, as dst g, s ) ( dst( g, s )) (5) ( m mn j s j where dst(g, s j ) s the Haversne formula based on the dstance between g and s j. So, n actualty, dst(g, s m ) s the shortest dstance from g to any pont on s m. We defne the smlarty functon, Sm(G', s m ), between G' and s m, based on the dstance of each matched par. The matched par, <g, s j >, conssts of g and the nearest pont to g, s j. We use the exponental functon, e, to measure the contrbuton of each matched par to Sm(G', s m ). Ths s because we would lke to assgn a larger contrbuton to a closer matched par of ponts, whle gvng a much lower value to those pars that are dstant. Ths results n an exponentally decreasng contrbuton, as dst(g, s m ) lnearly ncreases. Sm G T ( m dst( g, sm ) ', sm) e (6) 1 The segment valdaton process decdes whether the BCS s the segment that should be used n a predcton. To mantan effcency, we adapt a lower bound (LB). The LB s a standard value that accumulates the allowable error dstance for each s j. In order to select a proper LB, we consder the maxmum devaton and average devaton between each trajectory used n buldng a route model, a learned route pattern, as well as the margn of GPS error (20, 30, 50 meter). Accordng to the results of a test, usng the average devaton n the valdaton

Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... 256 process was found to produce the best accuracy for route predcton. If the smlarty of the BCS s greater than the LB, the BCS s a predctable segment. LB m 1 The route decson process nfers a user's current transt route based on a predctable segment and a route probablty model. As Fg. 6 shows, routes r 1 and r 2 have a common segment, whch s the nearest path (from v 1 to v 2 ) to a segment of GPS logs. Thus, r 1 and r 2 have equal smlarty to a segment of GPS logs. In order to resolve ths problem, we adapt a State-Observaton model to consder other facts that reflect a user's ntenton. In order to decde a user's ntended route, we use envronmental nformaton, such as the varables n table 1. For nstance, our method calculates the probablty of segment v 2 ~v 4 and the probablty of segment v 2 ~v 3 for a gven State-Observaton model, n order to determne whether the predctable segment s r 1 or r 2. Therefore, a route ncludng a segment that has the hghest probablty s determned to be the user's current transt route. current route (arg max p( seg )) r p( seg ) n 1 p( o Prev seg) p(preg seg) e rs p( seg Prev seg ) p( o Prev seg ) n 1 p( o Preg seg) (7) (8) developed by Kyunghee Unv. Ths module could dstngush 5 behavors (stay, walk, run, bus, subway) based on the smartphone s accelerometer and GPS. The test data consst of 46 routes of 15 users (tester), generated by route model learnng. In Korean ctes, n order to get on a bus or a subway, people commonly walk (about 5~10mntes) from specfc places (home, offce and so on) to bus stops or subway statons. Therefore, t s proper that a segment of GPS logs has a length of 1~5 mnutes, to mantan the practcalty of predcton results. Through the frst test, we decded the approprate length of a segment of GPS logs. Each tester dentfed whether a predcted route s a current route on whch he/she s movng. Ths was done usng our plot applcaton, whch predcted the user's current route. Ths work s performed over one month wthout the trajectory valdaton step. For the detaled analyss of test results, the ffteen testers were dvded nto fve groups and each tester dentfed whether a predcted route was ther current transt route. As shown n Fg. 7, the accuraces of route predctons, based on 3-mnute segments of GPS logs, are 92~94%. On the other hand, the accuraces for 1 or 2-mnute segments of GPS logs are low, overall. The accuraces for 4 or 5-mnute segments of GPS logs are no better than those obtaned wth 3-mnute segments of GPS logs. Fgure 6: An example of an overlappng route problem 6 Experment We descrbe the results of tests for our approach. For the tests, we have collected real data sets from 15 users for roughly 60 days, n Korean ctes. We performed three tests n order to prove the practcalty of our approach for route predcton. Our approach s mplemented n Java and examned on a androd phone wth QSD 8250 and 512MB Memory. In order to detect users' behavor, we use an actvty recognton module, whch was Fgure 7: The accuracy by the length of a GPS log In order to decde a threshold (for the LB) to valdate whether a BCS s a predctable segment, we performed a second test (usng 3 mnute segments of GPS logs). We consder the maxmum devaton and average devaton between the GPS ponts of a route and GPS ponts of each trajectory used n buldng the route, n addton to GPS error margns of 20, 30, and 50m (because detected GPS ponts have an error rate of about 50m, at most)

256 Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... partcular, the proposed probablstc approach, whch s based on a State-Observaton model, helps to solve problems wth overlappng routes that reflect a user's ntentons. Then, we suggest the average devaton as a useful threshold for valdaton tests. Based on experments conducted wth 15 smartphone users, our approach shows 96.4% accuracy. In route pattern learnng, we have found several ponts that can be mproved, such as parameter -senstvty or troubles wth routes over short dstances. In the future, we wll focus on these mprovements. Fgure 8: The valdaton accuraces for dfferent threshold factors of the LB As shown n Fg. 8, the accuraces of the valdaton results are the hghest when the average devaton s appled to a threshold to calculate the LB. Because the overall valdaton accuraces are 92~95%, the average devaton s a good factor n determnng the threshold for valdaton. In the fnal test, we measured the accuracy of our route predcton approach, adaptng a probablstc approach wth a segment of GPS logs (3 mnutes n length - test 1) and the average devaton (test 2). We consder senstvty and specfcty n computng accuracy. Senstvty refers to whether a predcted route correctly dentfes the user's current ntended route and specfcty refers to whether non-current routes are correctly fltered. Table 2 shows the result for our fnal test. Table 2: The test results of route predcton Predctable Unpredctable Total Current transt 536 14 550 route No current 22 428 450 transt route Total 541 459 1000 Senstvty 0.975 Specfcty 0.951 Accuracy 0.964 7 Concluson In ths paper, we propose a practcal approach for personal route modelng and predcton. For effcency and performance modelng, we suggest a learnng approach usng mage-processng and a probablstc method. In order to perform effcent predcton, we focus on three parts - the approprate length of a segment of GPS logs, the decson of users' ntended route, adaptng a probablstc method, and the threshold for valdaton. In Acknowledgements Ths work was supported by the Industral Strategc Technology Development Program (10035348, Development of a Cogntve Plannng and Learnng Model for Moble Platforms) funded by the Mnstry of Knowledge Economy(MKE, Korea). References [1] L. Chen, M. Lv, Q. Ye, G. Chen, and J. Woodward. A personal route predcton system based on trajectory data mnng. Informaton Scence, Vol.181, No.7, (2011), 1264-1284. [2] D. Koller and N. Fredman. Probablstc Graphcal Models: Prncples and Technques (The MIT press, Cambrdge, USA, 2009). [3] E.M. Tapa, S.S. Intlle, W. Haskell, K. Larson, J. Wrght, A. Kng, and R. Fredman. Real-tme recognton of physcal actvtes and ther ntenstes usng wreless accelerometers and a heart rate montor. Proceedngs of the Internatonal Symposum on Wearable Computers, (2007), 1-4. [4] J. Lester, T. Choudhury and G. Borrello. A practcal approach to recognzng physcal actvty. Proceedngs of the Internatonal Conference on Pervasve Computng, Vol.3968, (2006), 1-16. [5] Y. Zheng, Y.K. Chen, Q. L, X. Xe and W.Y. Ma. Understandng transportaton modes based on GPS data for web applcatons, ACM Trans on the Web, Vol.4, No.1, (2010), 1-36. [6] J.R. Parker. Algorthm for mage processng and computer Vson (Wley, Hoboken, USA, 2010). [7] S. Zhang and K.S. Fu. A thnnng algorthm for dscrete bnary mage, Computer Graphcs and Image Processng, Vol.13, No.2, (1980), 142-157. [8] J.E. Bresenham, Algorthm for computer control of a dgtal plotter, IBM systems journal, Vol.4, No.1, (1965), 25-30.

Je-Mn Km et. al.: Probablstc Graphcal Model based Personal Route... 256 Je-Mn Km s a Ph.D. student at School of computng, Soongsl Unversty. In 2004 he receved hs M.Sc. degree n computer scence. He worked prevously at the Kangnam Unversty as a teacher n the fe ld of compute r scence. He s the author and co-author of several scentfc papers, and partcpates n semantc web, ubqutous computng and moble computng project. He has strong scentfc and development expertse n ontology modelng, ontology reasonng and machne learnng. Haejung Baek s a researcher at School of computng, Soongsl Unversty. She receved her M.Sc. degree n computer scence n 1998. In 2004, she receved her Ph.D degree n Artfcal Intellgence; wth thess focused on actve learnng of robots usng symbolc and non-symbolc learnng. She studed on robots at KIST(Korea Insttute Scence Technology) and on object recognton for robots at CMU(Carneg Mellon Unversty n USA). She s also a adjunct professor n Kjeon Unversty n Jeonju. She s nterested n machne learnng, robot(vson and cogntve learnng) and moble computng. Young-Tack Park s a professor at School of computng, Soongsl Unversty. He receved hs M.Sc. degree n computer scence n 1980. In 1992, he receved hs Ph.D. degree n Artfcal Intellgence; wth thess focused on black board scheduler control knowledge for heurstc classfcaton. He teaches A.I at the Faculty of Computer Scence. He s also a supervsor and consultant for Ph.D., master and bachelor studes. He has authored and coauthored multple research papers and partcpated n natonal research projects. Hs research nterests n Semantc Web, Ontology Reasonng and Machne Learnng. He has strong scentfc and development expertse n ontology reasonng, agent system and moble computng.