Sensors 2012, 12, 6102-6116; doi:10.3390/s120506102 OPEN ACCESS sensors ISSN 1424-8220 www.mdpi.com/journl/sensors Review Inertil Sensor-Bsed Methods in Wlking Speed Estimtion: A Systemtic Review Shuozhi Yng nd Qingguo Li Deprtment of Mechnicl nd Mterils Engineering, Queen s University, Kingston, ON K7L 3N6, Cnd; E-Mil: yngs@me.queensu.c Author to whom correspondence should be ddressed; E-Mil: qli@me.queensu.c; Tel.: +1-613-533-3191; Fx: +1-613-533-6489. Received: 28 Februry 2012; in revised form: 26 April 2012 / Accepted: 29 April 2012 / Published: 10 My 2012 Abstrct: Self-selected wlking speed is n importnt mesure of mbultion bility used in vrious clinicl git experiments. Inertil sensors, i.e., ccelerometers nd gyroscopes, hve been grdully introduced to estimte wlking speed. This reserch re hs ttrcted lot of ttention for the pst two decdes, nd the trend is continuing due to the improvement of performnce nd decrese in cost of the miniture inertil sensors. With the intention of understnding the stte of the rt of current development in this re, systemtic review on the exiting methods ws done in the following electronic engines/dtbses: PubMed, ISI Web of Knowledge, SportDiscus nd IEEE Xplore. Sixteen journl rticles nd ppers in proceedings focusing on inertil sensor bsed wlking speed estimtion were fully reviewed. The existing methods were ctegorized by sensor specifiction, sensor ttchment loction, experimentl design, nd wlking speed estimtion lgorithm. Keywords: wlking speed; mbultory; sptio-temporl prmeters; biomechnics; inertil sensors; git segmenttion; review 1. Introduction Wlking is one of the most importnt dily ctivities for people to get from one plce to nother. Not surprisingly, sptil nd temporl git prmeters during wlking (i.e., wlking speed, stride length, stride frequency, nd git symmetry) hve been extensively studied for helthy nd pthologicl popultions. Besides these prmeters, self-selected wlking speed hs long been recognized s
Sensors 2012, 12 6103 proxy mesure of mbultion qulity nd is used to quntify the progress of git rehbilittion [1 4]. Trditionlly, stopwtch hs been used to determine the verge speed during wlking through known distnce, which is commonly prcticed in clinicl settings. To obtin more ccurte results nd instntneous wlking speed, lbortory-bsed cmer bsed motion trcking systems (i.e., OptoTrck, NDI) nd instrumented wlkwys (i.e., the GAITRite system) hve been introduced to determine the individul stride length nd stride frequency s well s the wlking speed [5]. Despite their ccurcy, these systems re restricted in lbortory settings due to their size nd cost. With the dvncement of microelectromechnicl systems (MEMS), inertil sensors (ccelerometers nd/or gyroscopes) becme more nd more populr in physicl ctivity monitoring becuse of their portbility, improved performnce, nd low cost [6]. Since the 1990s, mny inertil sensor-bsed methods hve been developed for estimting sptio-temporl prmeters during wlking. The sptio-temporl prmeters reported in these studies include wlking speed, stride length, nd totl distnce wlked. The estimtion of the lst two prmeters is equivlent to estimting wlking speed, but with different emphsis. The stride length defines the trvel distnce in one stride, nd the wlking speed cn be clculted by multiplying the ssocited stride frequency. The stride length tends to be used for evluting git vribility between wlking trils. On the other hnd, the totl distnce wlked quntifies the distnce covered, nd the verge wlking speed cn be esily clculted by dividing the distnce trveled with time spent. Although the selection of prmeters re different between studies, the essence is the sme in estimting the distnce covered in fixed time intervl. Therefore, this review does not distinguish stride length estimtion nd wlking distnce estimtion from wlking speed estimtion. As the nme suggests, n ccelerometer is device for mesuring ccelertions, including those induced by grvity. A gyroscope mesures ngulr velocity. The combintions of these sensors re referred to s inertil mesurement units (IMU). During wlking, body segments undergo cyclic motion nd the movement pttern of ech segment repets every stride cycle. The cyclic motion of body segment induces periodic ccelertion nd ngulr velocity chnges, which cn be sensed by the ttched inertil sensors. The embedded fetures in the mesurements enble the possibility in estimting wlking speed. A generic structure of wlking speed estimtion method is illustrted in Figure 1. Inertil sensors cn be ttched to different prts of the body (Figure 1()) nd mesure different signls (Figure 1(b)). With these mesurements, different models nd methods cn be used to estimte wlking speed (Figure 1(c)). Under this generic structure, wide rnge of inertil sensors nd configurtions hve been dopted in estimting wlking speed, nd it is certinly of interest to exmine how they were utilized nd how well they performed. There re comprehensive reviews in the re of werble sensors for git nlysis [7], reviews on inertil sensors in monitoring lower limb biomechnics [8], nd in the git event detection [9]. However, there is no review on current sttus of inertil sensors in wlking speed estimtion, which is the primry purpose of this pper.
Sensors 2012, 12 6104 Figure 1. Generic wlking speed estimtion method. (). Inertil sensors (ccelerometer nd/or gyroscope) re ttched to different prts of the user. (b). Inertil sensors mesure the ccelertions nd/or ngulr velocities which contin informtion relted to the wlking speed. (c). A wlking speed estimtion lgorithm extrcts the wlking speed informtion from these sensor mesurements. (d). Different wlking speed re distinguished s outputs of the lgorithm. 400 200 0 () -200 (b) 2.0 S T estimted speed (m/s) 1.5 1.0 0.5 0.8 m/s 1.0 m/s 1.2 m/s 1.4 m/s 1.6 m/s 1.8 m/s (d) (c) 2. Methods 2.1. Review Questions We systemticlly reviewed the literture regrding the inertil sensor-bsed wlking speed estimtion methods, nd ttempted to nswer the following questions: (1) Wht re the existing inertil sensor bsed wlking speed estimtion methods (including stride length estimtion nd wlking distnce estimtion)? (2) Wht types of inertil sensors were used in relted studies? (3) Where were the inertil sensors ttched? (4) How were the experiments conducted? nd (5) How ws the performnce of these studies? In order to nswer ll these questions, we reviewed the literture on inertil sensor bsed wlking speed, stride length nd wlking distnce estimtion methods. 2.2. Article Selection The reserch method is grphiclly depicted in Figure 2 for better understnding of the procedure. We systemticlly serched for published journl rticles nd ppers in proceedings in PubMed (from 1950), ISI Web of Knowledge (Science Cittion Index Expnded, from 1899; Socil Sciences Cittion Index, from 1956; Art & Humnities Cittion Index, from 1975), SportDiscus (from 1950) nd IEEE Xplore (from 1950) t the first week of July in the yer 2011. These four electronic engines/dtbses were chosen becuse of their populrity nd their coverge of literture in engineering, medicine nd biomechnics. The serched keyword string ws (ssessment OR estimtion OR clcultion OR computtion OR mesurement) AND (inertil sensor OR ccelerometer OR gyroscope OR inertil mesurement unit) AND (speed OR velocity OR step length OR stride velocity OR stride length) AND
Sensors 2012, 12 6105 (wlking) for ppernce in the title, bstrct nd keyword fields of the rticles. The initil totl number of identified rticles ws 344. The title nd the bstrct of ech rticle ws red crefully for the first selection stge, nd unrelted nd duplicted rticles were excluded, which reduced the number of rticles to 47. In the second selection stge, these 47 full rticles were retrieved from the Queen s University librry system nd completely reviewed. A totl of 16 full rticles were ultimtely included in this review. The inclusion criteri were s follows: (1) the study involved inertil sensors, nd (2) the study reported wlking speed, stride length or wlking distnce estimtion results. However, s this review focuses on the method development, rticles only contining the following contents were excluded: (1) performnce evlution of commercil vilble product without reveling the wlking speed estimtion method, (2) performnce comprison between existing methods, nd (3) pplictions bsed on previous reported methods. Figure 2. Article review procedures. After the initil serch, the title nd bstrct were reviewed first to exclude unrelted rticles. The full rticles were then retrieved nd reviewed with the detiled inclusion/exclusion criteri. 16 rticles were finlly included in this review. Initil Serch PubMed 73 ISI Web of Knowledge 89 SportDiscus 72 IEEE Xplore 110 1st Selection stge: title nd bstrct Totl Articles 47 2nd Selection stge: full rticle Totl Articles 16 3. Results 3.1. Sensor Specifiction Lrge vrieties of inertil sensors re currently vilble on the mrket, rnging from unixil ccelerometers/gyroscopes to IMUs with 6 degrees of freedom (6DOF). The mesurement rnge of the inertil sensors vried with the specifictions, from ±2 g to ±50 g for ccelerometers nd from ±150 /s to ±1,000 /s for gyroscopes. The inertil sensor mesurements were smpled nd filtered with different frequencies. Depending on the purpose of the study nd the system design, different inertil sensors nd sensor configurtions were dopted. Tble 1 in Appendix shows the detiled specifictions of sensors used in these studies. No cler rtionle in sensor selection hs been reported in the reviewed rticles. As generl rule, the section of sensor type (ccelerometer nd/or gyroscope) is dependent on the wlking speed estimtion lgorithm. For exmple, the direct integrtion bsed pproch requires mesurements of both ccelertion nd ngulr velocity. On the other hnd, the selection of sensor mesurement rnge
Sensors 2012, 12 6106 is determined by the loction of the sensor, nd in most cses, more distlly ttched sensor should hve bigger rnge of mesurement. 3.2. Sensor Attchment Loction Most body motion during humn git occurs in the lower limbs; therefore, most of the reviewed studies chose to ttch the inertil sensors on the thigh nd the shnk or the foot of the subjects. One study [10] ttempted to cpture the motion with n ccelerometer ttched to the chest. Four studies [11 13] utilized sensors ttched t the lower spine to estimte the wlking speed. One study dditionlly used force/moment sensor [14] s n iding component in the system. Tble 2 in Appendix shows the sensor types nd the ttchment loctions used in the reviewed rticles. We followed similr illustrtion method s proposed in [9]. The effect of sensor ttchment loction on the qulity of the mesurements ws not discussed in the reviewed studies. The following fctors need to be considered when selecting sensor loction. (1) The linkge between the sensor mesurement nd the wlking speed. Even wlking t the sme speed, the chrcteristics of ccelertion nd ngulr velocity re different from loction to loction. A wlking speed estimtion lgorithm extrcts wlking speed informtion from these chrcteristics nd therefore is sensor loction dependent. The fesibility of deriving wlking speed informtion from sensor mesurements needs to be considered first in selecting specific sensor loction. (2) The reltive movement between the sensor nd the body segment. As the sensors re typiclly ttched to the body segment with strp or tpe, reltive motion is unvoidble nd will cuse discrepncy between the mesured ccelertion nd ngulr velocity nd those of the body segment. The differences ultimtely ffect the wlking speed estimtion ccurcy nd robustness of the lgorithm. (3) Robustness to the disturbnce cused by the bnorml motion. In ech wlking speed estimtion method, certin ssumptions hve been mde under which the lgorithm will function properly. Any bnorml motion tht devites from the ssumption will generte errors in the estimted wlking speed. As n exmple, sensor ttched to the foot my be ffected by the bnorml foot motion, such s the equinus git often observed in children with cerebrl plsy [15]. A concurrent comprison study demonstrted tht the foot ttched sensor is prone to wlking speed estimtion error during toe-out wlking s compred with the shnk-mounted sensor when using 2D inertil sensor [16]. The robustness needs to be considered in selecting sensor loction, especilly when pplying the wlking speed estimtion method to popultion with git bnormlities. 3.3. Experimentl Design Most of the reviewed studies focused on experimentl vlidtion with helthy subjects. Some studies iming to pply their methods in ge-relted or pthologicl git chose to include elderly subjects [17,18], spinl cord injured subjects [19] or ptients with prostheses or hemiplegic git [20]. In some studies, the proposed methods were vlidted with only one subject [19,21 24]. Although single subject verifiction might not be sufficient to demonstrte the robustness of the proposed methods, the ide of the proposed methods ws clerly explined. Two mjor forms of experimenttion were tredmill wlking nd overground wlking t either preset speed or preferred speed. For those studies tht involved elderly or impired subjects [17 20], the experiments were designed with cre while providing
Sensors 2012, 12 6107 resonble comprison with young/helthy subjects. When compred with helthy young subjects, the wlking speed estimtion method produced less ccurte results for subjects with pthologicl git in most cses. Four out of 16 studies [14,18,22,23] concentrted on inertil sensor bsed personl nvigtion systems tht were cpble of monitoring the subjects wlking in 3D environments; 12 out of 16 studies focused on the motion of the subjects in the sgittl plne only. The detiled experimentl design informtion is shown in Tble 3 in Appendix. 3.4. Wlking Speed Estimtion Algorithm The lgorithms of using inertil sensors to estimte wlking speed cn be grouped into three ctegories: (1) bstrction model (three studies: [10,11,13]), (2) humn git model (five studies: [12,17,19,20,25]) nd (3) direct integrtion (eight studies: [14,18,21 24,26,27]). 3.4.1. Abstrction Model Insted of clculting wlking speed bsed on certin physicl model, some studies decided not to look into the detils of the humn git biomechnics, but to bstrct the system nd construct blck-box model for the complex reltionship between the sensor mesurements nd the wlking speed from n informtion processing point of view. In 1995, Aminin et l. [11] proposed wlking speed estimtion lgorithm with the id of rtificil neurl networks (ANNs) with four ccelertion mesurements s inputs. In this study, the system consisted of two two-lyer ANNs, in which the input ccelertion signls were collected from the bck trunk (trixil) nd the heel (unixil), nd the first ANN generted the incline estimtes while the second ANN estimted the wlking speed. Before the clssifiction phse (wlking speed estimtion), the trining phse (lerning process) of ech ANN used set of 360 ccelertion signl ptterns obtined from tredmill wlking experiments, nd ssocited the ccelertion signls with the ctul wlking speed by djusting the weights nd bises of the ANN to minimize the sum squred errors. At the end of the trining phse, blck-box ANN model with fixed weights nd bises ws generted nd lter used in the clssifiction phse to mp the ccelertion signls to the wlking speed, which chieved mximum reltive error of 16% from the overground wlking experiments. A similr method ws dopted by Song et l. [10]. Song employed two-stge structure consisting of three ANNs to process the ccelertion signls collected from n ccelerometer (trixil) ttched to the chest. At the first stge, the wlking/running clssifiction network clssified the type of git, either wlking or running; t the second stge, the wlking neurl network (NNW) or the running neurl network (NNR) processed the ccelertion signls ccordingly to estimte wlking or running speed. The overll root men squred error (RMSE) ws 0.54 km/h bsed on wlking/running experiments t vrious speed between 4.7 km/h nd 17.14 km/h. Different from the ANNs used in these two studies, Yeoh et l. [13] defined the verge net ccelertion (ANA) of the left nd right thighs nd estimted the wlking speed using third-order polynomil model. Before the wlking experiment, sufficient mount of ccelertion signl dt t vrious wlking speed were collected (trining phse), nd polynomil model ws determined by fitting the men vlue of ANA with respect to the wlking speed using lest squres pproch. This method ws derived bsed on the fct tht the force exerted by n object is directly proportionl to
Sensors 2012, 12 6108 the ccelertion nd the physicl ctivity intensity (or wlking speed) cn be expressed s function of ccelertion. The overll men squred error ws 1.76 km/h bsed on the wlking trils t speed rnging from 1 km/h to 13 km/h. The implementtion nd the experimentl results of these three studies demonstrted the fesibility of the bstrction model bsed method in estimting wlking speed. As n inherent property of the bstrction model bsed method, the reltionship between the wlking speed nd the inertil sensor mesurements is modeled s blck-box model with set of prmeters. Although the prmeters do not hve ny explicit mening, mp cn be estblished through trining. Although the off-line trining could be time consuming, the wlking speed estimtion method itself is fst nd highly simplified, which is suitble for rel time implementtion. Additionlly, since no physicl model is required in this method, lrge vriety of signls cn potentilly be used s inputs to the bstrction model, which implies tht the sensor loction nd type re highly flexible (e.g., ttched to bck of the trunk nd heel in [11] nd to chest in [10]). Due to the fct tht the bstrction model is n pproximtion of the ctul physicl system, the ccurcy of the wlking speed estimtion depends on the completeness of trining dt set. The ccurcy of the estimtion is generlly low nd it is difficult to control the performnce consistency cross multiple subjects. 3.4.2. Humn Git Model Some reserchers chose to estimte wlking speed with stride length, bsed on some predefined humn git models. This clss of methods ws motivted by the fct tht some spects of the lower limb kinemtics, other thn stride length, cn be determined from the mesurements of the inertil sensor ttched to the leg. With the ssistnce of humn git model, the stride length cn be indirectly estimted bsed on the mesured lower limb kinemtics. Most erlier studies tried to employ simplified git model to void complicted sensor configurtion nd to reduce the computtion complexity. In 1997, Miyzki [20] proposed stride length nd wlking speed estimtion method using gyroscope (unixil) ttched to the thigh nd symmetric git model. In this method, ech leg ws modeled s one single segment, nd the two legs were ssumed symmetricl; thus, t heel-strike, two legs nd the distnce between the feet formed isosceles tringle. The ngle between the leg segments ws clculted by integrting the ngulr velocity mesurement from the gyroscope ttched to the thigh, nd then the distnce between two feet (step length, one hlf of stride length) ws clculted using the properties of isosceles tringle. The overground wlking experiment showed tht this simple git model method chieved reltive error of 15%. Similr to [20], Tong et l. [19] lso modeled ech leg s one segment nd ttched gyroscope (unixil) on the shnk. When clculting the stride length, Tong implicitly used pendulum model in which the one-segment leg swung bck nd forth bout the hip joint during wlking. The stride length ws simply clculted s the product of the leg segment inclintion rnge (rd) nd the leg segment length (m). Another method of using the single segment git model ws proposed by Tnk et l. [25]. One dditionl ccelerometer (bixil) ws ttched to the shnk long with gyroscope (unixil). The lgorithm used the sme model s Miyzki s method, but introduced the ccelertion mesurement to determine initil thigh ngle just before wlking initition. In 2002, Aminin et l. [17] utilized foot switch (to monitor temporl prmeter) nd two gyroscopes (unixil) to estimte wlking speed. Discrding the
Sensors 2012, 12 6109 simplified git model reported in [19,20], Aminin et l. chose to solve the complete git model with seprte shnk nd thigh segments with the sme ssumption of symmetry between two legs. In this method, the rottion ngles of the shnk nd the thigh were trcked by two gyroscopes, nd ech stride cycle ws divided into stnce phse nd swing phse using the foot switch. The trvel distnce in ech phse ws solved geometriclly using the rottion ngles nd the lengths of shnk nd thigh. Evluted from both tredmill nd overground wlking experiments, the overll RMSE ws 0.06 m/s (6.7%) for wlking speed estimtion nd 0.07 m (7.2%) for stride length estimtion. A different model of humn git ws first introduced by Zijlstr et l. [12] in wlking speed estimtion. The proposed method, rther thn ttching sensors on the lower limb, used the verticl displcement of the center of mss (CoM) to estimte wlking speed. Since the CoM movements in the sgittl plne follow circulr trjectory bout the foot during ech single support phse, upon the determintion of the CoM verticl displcement, the stride length cn be derived geometriclly. The experimentl results showed tht the mximum reltive error is bout 16%. Using predefined git model long with inertil sensor mesurements to estimte wlking speed is beneficil in severl spects: (1) simple sensor setup, (2) ese of use. First, with the support of git model, the inertil sensors were usully used to provide only one or two kinemtic prmeters, e.g., the shnk/thigh ngle [17,19,20,25] nd the verticl displcement of the CoM [12]. Since only one or two sensor mesurements were processed in this method, less effort ws required to del with the inevitble sensor errors, i.e., noise nd bis. Second, unlike the bstrction model method, the git model bsed method followed physicl principles to construct the humn git model; thus, no subject-specific trining phse ws required before the ctul wlking speed estimtion ppliction. This method lso hs some shortcomings. The ccurcy highly depends on the vlidity of the model, nd the git model directly ffects the complexity of the lgorithm. Compring the performnce of [17] nd [20], the wlking speed estimtion error of the simplified git model ws bout two times bigger thn tht of more complete git model. However, the improved ccurcy required much more complicted clcultion procedure [17]. In ddition, subject-specific nthropometric mesurements, i.e., lower limb segment length, must be tken in order to construct the humn git model. 3.4.3. Direct Integrtion In recent yers, more nd more studies strted to use direct integrtion method to estimte wlking speed. A generic direct integrtion lgorithm includes the following steps: (1) define strting nd ending point of ech stride cycle; (2) determine the orienttion of the inertil sensor with respect to the globl coordinte system; (3) project the ccelertion mesurement into the globl coordinte system bsed on the instntneous orienttion of the inertil sensor nd remove the ccelertion due to grvity; nd (4) integrte the ccelertion in the globl coordinte system from the strting point to obtin instntneous sensor velocity nd the ssocited stride length. The direct integrtion methods hve been developed seprtely for humn git nlysis nd personl nvigtion. Erly studies using body-fixed inertil sensors mostly focused on humn git nlysis. In the reviewed rticles, the first direct integrtion wlking speed estimtion method ws proposed by Sbtini et l. [26] in 2005, which used n IMU (bixil ccelerometer nd bixil of gyroscope) fixed on the instep of the foot. With resonble ssumption tht the foot (with the shoe) ws rigid
Sensors 2012, 12 6110 enough, the estimted sensor velocity could be viewed s the velocity of the foot. The foot flt (FF) event ws defined s the strting point of ech stride cycle nd the ngulr velocity dt ws used to detect the FF event in the stnce phse. One importnt procedure in this lgorithm ws the zero velocity updte (ZUPT), which determined the initil sensor orienttion nd estimted the sensor mesurement offsets during the period of the stride cycle. This method chieved n overll RMSE of 0.18 km/h bsed on the tredmill wlking experiments t vrious speed rnging from 3 km/h to 6 km/h. Alvrez et l. [21] used similr method to estimte the foot displcement over one stride cycle; however, Alvrez utilized one IMU (trixil ccelerometer nd unixil gyroscope) on ech foot nd dt fusion lgorithm to reduce the estimtion error. Although the experimentl results (reltive error 10.1 ± 6.2%) showed limited improvement from the results of [26], Alvrez extended Sbtini s study nd introduced method to combine the informtion obtined from multiple sensors tht could potentilly increse the wlking speed estimtion ccurcy. Attching the sensor to the foot provided lot of benefits, such s the possibility to implement ZUPT t foot flt, but the flexibility of the nkle joint brought concerns the influence of the bnorml git on the inertil sensor dt, such s out of plne motion [28]. To void such issue, Li et l. [27] ttched n IMU (bixil ccelerometer nd unixil gyroscope) to the lterl side of mid-shnk nd estimted the wlking speed with direct integrtion method. Different from ttching sensor on the foot, ZUPT technique could not be used for the sensor ttched to the shnk. Insted, they defined the strt nd end points of ech stride cycle s when the shnk ws verticl (shnk ngle zero), nd mde use of the inverted pendulum model to ssume tht the initil sensor velocity ws zero. This ssumption is bsed on the fct tht the CoM ws t its highest point nd the kinetic energy ws ll trnsformed into potentil energy t the shnk verticl event. In their study, percentge RMSEs of 7% nd 4% were obtined from the tredmill nd the overground wlking experiments, respectively. These methods [21,26,27] focused on wlking speed estimtion in the sgittl plne (or direction of progression) only, since most biomechnicl studies used wlking speed evluted long stright line, e.g., 10-meter wlking test (10MWT) [29]. In 2010, Mrini et l. [18] ttempted to use n IMU (trixil ccelerometer nd trixil gyroscope) ttched to the bck of the heel to estimte the stride length, stride velocity nd turning ngle in three-dimensionl spce. Quternion representtion of the sensor orienttion in three-dimensionl spce ws first obtined. The ccelertion mesurement ws then projected to the globl coordinte system bsed on the sensor orienttion, nd the ccelertion due to grvity ws removed from the projected sensor ccelertion mesurement. After the double integrtion of the projected sensor ccelertions, the foot position in ech stride ws expressed in 3-dimensionl spce. The stride length ws determined s the distnce between the positions of the foot t two djcent foot flt events. The stride length nd stride velocity estimtion reltive error were chieved s 1.3 ± 6.8% nd 1.5 ± 5.8%, respectively. In prllel to humn git nlysis ppliction, miniture inertil sensors hve been used in personl nvigtion s potentil lterntive to the GPS system. One ttempt ws conducted by Ojed nd Borenstein [22]. They developed nvigtion system using n IMU (trixil ccelerometer nd trixil gyroscope) ttched to the lterl side of the foot. They lso used the quternion representtion of the sensor orienttion to determine the instntneous sensor orienttion. The overll movement estimtion ws through process clled ded reckoning, with which the current position ws determined by using previously determined position. The trvel distnce estimtion error ws less thn 2% s reported
Sensors 2012, 12 6111 in [22]. In 2010, Mrtin et l. [14] used n IMU s (trixil ccelerometer nd trixil gyroscope) nd two force sensors ttched beneth the heel nd the forefoot. The force sensors were used to detect the time instnt of heel down (HD) tht ws defined s the strting point of the stride cycle. The overll stride length estimtion error obtined from the 10 MWT ws 34.1 ± 2.7 mm. Although the bsic estimtion procedure ws the sme, Hung et l. [24] used the direction cosines representtion to trck the orienttion of the IMU (trixil ccelerometer nd trixil gyroscope) ttched to the rch of the foot, nd chieved wlking distnce estimtion error of bout 2%. Moreover, concerning the effect of the sensor noise in the estimtion ccurcy, Bebek et l. employed extended Klmn filter (EKF) to reduce the sensor noise nd bis through the stnce phse. A pressure sensor rry ws plce between the heel of the shoe nd the shoe insole, nd n IMU (trixil ccelerometer nd trixil gyroscope) ws ttched to the lterl side of the foot. The pressure sensor rry ws used to detect the zero velocity period of the stride cycle, in which the ZUPT with EKF ws pplied. The reltive error of the system ws 0.40%, evluted in the outdoor wlking experiments with n verge distnce of 1,215 m. The direct integrtion methods benefited from the incresing ccurcy of the miniture IMUs nd the sophisticted new lgorithms. Compred to the bstrction model bsed method nd the humn git model bsed method, this clss of methods is esier to use without the troubles from the trining process or subject-specific model building/clibrtion. Currently, the IMU is ttched to the foot or the shnk in order to tke dvntge of using the ground s reference in the lgorithm. As discussed in mny studies, two importnt components of the direct integrtion method were the determintion of the sensor orienttion nd the sensor error correction. Since the direct integrtion method solely used the IMU mesurement in the estimtion nd very little externl informtion ws vilble on-the-fly, the sensor orienttion determintion hevily relied the ngulr velocity mesurement or the ngle representtion such s Euler ngle, quternion nd direction cosines. On the other hnd, the sensor noise nd bis must be resonbly corrected in the estimtion process to ensure the ccurte estimtion results. One common sensor error correction technique ws ZUPT. The sensor noise nd bis were evluted during the zero velocity period of the stride cycle, nd then compensted from the clcultion. With the direct integrtion method, three-dimensionl motion monitoring ws lso mde possible [14,18,22 24]. In contrst to humn git nlysis ppliction, personl nvigtion requires estimting 3D position bsed on only inertil sensor mesurements nd the ccurcy of the estimtion is criticl becuse of long durtion opertion. The reviewed rticles indicte tht the results from personl nvigtion pplictions tend to hve better ccurcies. Although most ge-relted or pthologicl git reserch still considered wlking speed during the stright line wlking, the three-dimensionl motion trcking cpbility will definitely extend the ppliction of inertil sensors in humn git nlysis. 4. Discussion For the pst two decdes, lrge mount of work hs ttempted to use inertil sensors in estimting wlking speed. The existing methods cn be ctegorized into three groups: bstrction model, humn git model nd direction integrtion. Ech method hs its merits nd limittions. There is cler trend tht more nd more methods use the direct integrtion method with 2D or 3D IMUs. In the direct integrtion method, the double integrtion process mplifies mesurement errors, leding to the requirement of inertil sensors with higher ccurcy. Although the sensor performnce hs been rmped
Sensors 2012, 12 6112 up drmticlly, the inertil sensor mesurement error is unvoidble, especilly for miniture MEMS sensors. As one of the future reserch directions, development should focus on sensor error modeling nd ccommodtion to further improve prmeter estimtion ccurcy [30]. A method tht combines humn git model nd direct integrtion my be potentil cndidte for this purpose. Since this systemtic review is developed with focus on lgorithm development, performnce evlution or system vlidtion rticles were not included. As found from the reviewed studies, the experimentl designs nd the presenttion of results re quite different between studies, which mde comprison between methods difficult. A stndrd experimentl benchmrk or performnce evlution protocol should be developed to verify the performnce of new lgorithm. Another re of lgorithm development should focus on the unlevel ground wlking nd popultion with pthologicl git. There hs been recent lgorithm development for inclined wlking [31]. Similrly, Using inertil sensors in estimting git other thn wlking (e.g., running [32]) is lso very interesting. As limittion nd the inherent nture of systemtic review, it is unvoidble to omit relevnt literture due to the choice of keywords in the serch. However, this review did cover the mjor methods in the re of inertil sensor bsed wlking speed estimtion methods nd provide useful informtion for future lgorithm development. Acknowledgements We would like to grtefully cknowledge the support from NSERC discovery grnt nd Queen s SARC grnt. We lso thnk Emily Gogrty for her help in editing the mnuscript. References nd Notes 1. Cohen, J.; Sveen, J.; Wlker, J.; Brummel-Smith, K. Estblishing criteri for community mbultion. Topics Geritric Rehb. 1987, 3, 71 78. 2. Richrds, C.; Mlouin, F.; Dums, F.; Trdif, D. Git velocity s n outcome mesure of locomotor recovery fter stroke. In Git Anlysis: Theory nd Appliction; Mosby: St Louis, MO, USA, 1995; pp. 355 364. 3. Schmid, A.; Duncn, P.; Studenski, S.; Li, S.; Richrds, L.; Perer, S.; Wu, S. Improvements in speed-bsed git clssifictions re meningful. Stroke 2007, 38, 2096 2100. 4. Mudge, S.; Stott, N. Outcome mesures to ssess wlking bility following stroke: A systemtic review of the literture. Physiotherpy 2007, 93, 189 200. 5. Blsubrmnin, C.; Bowden, M.; Neptune, R.; Kutz, S. Reltionship between step length symmetry nd wlking performnce in subjects with chronic hemipresis. Arch. Phys. Med. Rehb. 2007, 88, 43 49. 6. Sbtini, A. Computtionl Intelligence for Movement Sciences: Neurl Networks, Support Vecotr Mchines nd Other Emerging Techniques; Ide Group Inc.: Hershey, PA, USA, 2006; pp. 70 100. 7. To, W.; Liu, T.; Zheng, R.; Feng, H. Git nlysis using werble sensors. Sensors 2012, 12, 2255 2283. 8. Fong, D.; Chn, Y. The use of werble inertil motion sensors in humn lower limb biomechnics studies: A systemtic review. Sensors 2010, 10, 11556 11565.
Sensors 2012, 12 6113 9. Rueterbories, J.; Spich, E.; Lrsen, B.; Andersen, O. Methods for git event detection nd nlysis in mbultory systems. Med. Eng. Phys. 2010, 32, 545 552. 10. Song, Y.; Shin, S.; Kim, S.; Lee, D.; Lee, K. Speed estimtion from tri-xil ccelerometer using neurl networks. In Proceedings of 29th Annul Interntionl Conference of the IEEE Engineering in Medicine nd Biology Society, Lyon, Frnce, 23 36 August 2007; pp. 3224 3227. 11. Aminin, K.; Robert, P.; Jéquier, E.; Schutz, Y. Estimtion of speed nd incline of wlking using neurl network. IEEE Trns. Instrum. Mes. 1995, 44, 743 746. 12. Zijlstr, W.; Hof, A. Assessment of sptio-temporl git prmeters from trunk ccelertions during humn wlking. Git Posture 2003, 18, 1 10. 13. Yeoh, W.; Pek, I.; Yong, Y.; Chen, X.; Wluyo, A. Ambultory monitoring of humn posture nd wlking speed using werble ccelerometer sensors. In Proceedings of 30th Annul Interntionl Conference of the IEEE Engineering in Medicine nd Biology Society, Vncouver, BC, Cnd, 20 24 August 2008; pp. 5184 5187. 14. Mrtin, S.; vn Asseldonk, E.; Bten, C.; Veltink, P. Ambultory estimtion of foot plcement during wlking using inertil sensors. J. Biomech. 2010, 43, 3138 3143. 15. Goldstein, M.; Hrper, D. Mngement of cerebrl plsy: Equinus git. Develop. Med. Child Neurol. 2001, 43, 563 569. 16. Ludnski, A.; Yng, S.; Li, Q. A concurrent comprison of inerti sensor-bsed wlking speed estimtion methods. In Proceedings of 33th Annul Interntionl Conference of the IEEE Engineering in Medicine nd Biology Society, Boston, MA, USA, 30 August 3 September 2011; pp. 3484-3487. 17. Aminin, K.; Njfi, B.; Bul, C.; Leyvrz, P.; Robert, P. Sptio-temporl prmeters of git mesured by n mbultory system using miniture gyroscopes. J. Biomech. 2002, 35, 689 699. 18. Mrini, B.; Hoskovec, C.; Rocht, S.; Bül, C.; Penders, J.; Aminin, K. 3D git ssessment in young nd elderly subjects using foot-worn inertil sensors. J. Biomech. 2010. 19. Tong, K.; Grnt, M. A prcticl git nlysis system using gyroscopes. Med. Eng. Phys. 1999, 21, 87 94. 20. Miyzki, S. Long-term unrestrined mesurement of stride length nd wlking velocity utilizing piezoelectric gyroscope. IEEE Trns. Biomed. Eng. 1997, 44, 753 759. 21. Alvrez, J.; González, R.; Alvrez, D.; Lóprez, A.; Rodríguez-Urí, J. Multisensor pproch to wlking distnce estimtion with foot inertil sensing. In Proceedings of 29th Annul Interntionl Conference of the IEEE Engineering in Medicine nd Biology Society, Lyon, Frnce, 23 26 August 2007; pp. 5719 5722. 22. Ojed, L.; Borenstein, J. Non-GPS nvigtion for security personnel nd first responders. J. Nvig. 2007, 60, 391 407. 23. Bebek, O.; Suster, M.; Rjgopl, S.; Fu, M.; Hung, X.; Cvusoglu, M.; Young, D.; Mehregny, M.; vn den Bogery, A.; Mstrngelo, C. Personl nvigtion vi high-resolution git-corrected inertil mesurement units. IEEE Trns. Instrume. Mes. 2010, 59, 3018 3027. 24. Hung, C.; Lio, Z.; Zho, L. Synergism of INS nd PDR in self-contined pedestrin trcking with miniture sensor module. IEEE Sensors J. 2010, 10, 1349 1359.
Sensors 2012, 12 6114 25. Tnk, S.; Motoi, K.; Nogw, M.; Ymkoshi, K. A new portble device for mbultory monitoring of humn posture nd wlking velocity using miniture ccelerometers nd gyroscope. In Proceedings of 26th Annul Interntionl Conference of the IEEE Engineering in Medicine nd Biology Society, Sn Frncisco, CA, USA, 1 5 September 2004; Volume 1, pp. 2283 2286. 26. Sbtini, A.M.; Mrtelloni, C.; Scpellto, S.; Cvllo, F. Assessment of wlking fetures from foot inertil sensing. IEEE Trns. Biomed. Eng. 2005, 52, 486 94. 27. Li, Q.; Young, M.; Ning, V.; Doneln, J.M. Wlking speed estimtion using shnk-mounted inertil mesurement unit. J. Biomech. 2010, 43, 1640 1643. 28. Jsiewicz, J.; Allum, J.; Middleton, J.; Brriskill, A.; Condie, P.; Purcell, B.; Li, R. Git event detection using liner ccelerometers or ngulr velocity trnsducers in ble-bodied nd spinl-cord injured individuls. Git Posture 2006, 24, 502 509. 29. Slbch, N.; Myo, N.; Higgins, J.; Ahmed, S.; Finch, L.; Richrds, C. Responsiveness nd predictbility of git speed nd other disbility mesures in cute stroke. Arch. Phys. Med. Rehb. 2001, 82, 1204 1212. 30. Yng, S.; Ludnski, A.; Li, Q. Inertil sensors in estimting wlking speed nd inclintion: An evlution of sensor error models. Med. Biol. Eng. Comput. 2012, 50, 383 393. 31. Prk, S.; Suh, Y. Height compenstion using ground inclintion estimtion in inertil sensor-bsed pedestrin nvigtion. Sensors 2011, 11, 8045 8059. 32. Yng, S.; Mohr, C.; Li, Q. Ambultory running speed estimtion using n inertil sensor. Git Posture 2011, 34, 462 466. Appendix Source Yer Sensor Model Tble 1. Inertil Sensor Specifictions in Reviewed Studies. Number of Sensors Type nd Specifiction Accelerometer Gyroscope Mesuring Axes Mesuring Rnge Mesuring Axes Mesuring Rnge Smpling Frequency (Hz) Aminin et l. [11] 1995 IC Sensors 3021 4 3 40 16 Miyzki [20] 1997 ENC-05S, Murt 1 1 ±150 /s Tong nd Grnt [19] 1999 ENC-05EA, Murt 1 1 50 0.3 4 Aminin et l. [17] 2002 ENC-03J, Murt 3 1 200 Zijlstr nd Hof [12] 2003 Kistler 1 3 ±2g 100 20 Tnk et l. [25] 2004 Sbtini et l. [26] 2005 cc: MC301, Wcoh gyro: ENC-03J, Murt cc: ADXL210E, Anlog Device gyro: ENC-03J, Murt 3 1 25 1 2 ±10g 2 ±300 /s 200 Dt Filtering (Hz) cc: 3 gyro: 0.3 20 cc: 17 gyro: 15 Alvrez et l. [21] 2007 MTx, Xsens 2 3 ±2g 1 100 Ojed nd Borenstein [22] 2007 SiIMU01, BAE 1 3 ±50g 3 ±1000 /s Song et l. [10] 2007 ADXL330 1 3 ±3g 200 Yeoh et l. [13] 2007 Crossbow 3 2 ±2g 25 Mrtin et l. [14] 2010 MTx, Xsens 2 3 3 50 15 Bebek et l. [23] 2010 InertiCube3, InterSense 1 2 ±6g 1 Hung et l. [24] 2010 1 3 3 Li et l. [27] 2010 cc: ADXL320 gyro: ADXRS300 1 2 1 1000 Mrini et l. [18] 2010 S-Sense 1 3 ±3g 3 roll, yw: ±300 /s pitch: ±800 /s 200 17 Empty entries indicte vribles unspecified by the uthor.
Sensors 2012, 12 6115 Tble 2. Sensor Type nd Attchment Position. In totl, nine different types of inertil sensors, including unixil, bixil nd trixil ccelerometer nd gyroscope, were ttched to 12 positions, including the chest, bck of the trunk, thigh, shnk nd foot. 3 2 1 1 2 3 C1 Cx Chest Sx Shnk B1 Bx Bck Thx Thigh Fx Hx Foot Heel One-dimensionl Two-dimensionl Three-dimensionl Th2 Th1 Accelertion Angulr Velocity Sensor Unit S2 S1 H1 F4 H2 F1 F2 F3 Position References 1 2 3 C1 [10] B1 [13] [11,12] Th1 [20] Th2 [17] [13] S1 [19] S2 [17] [27] F1 [26] [24] F2 [21] F3 [14] F4 [22] [23] H1 [11] [18] H2 [14]
Sensors 2012, 12 6116 Tble 3. Experimentl Design in Reviewed Studies. Source Subject Experimentl Design Aminin et l. [11] 5 helthy subjects Tredmill wlking t preferred speed. Miyzki [20] Tong nd Grnt [19] Aminin et l. [17] Zijlstr nd Hof [12] 18 helthy subjects; 7 ptients with bove knee prostheses; 10 hemiplegic subjects 1 incomplete spinl cord injured subject; 1 unimpired subject 9 young subjects; 11 elderly subjects 25 helthy subjects Helthy subjects: Ptients with bove knee prostheses: Hemiplegic subjects: Overground 4.5 m wlking t preferred speed. Tnk et l. [25] 10 helthy subjects Overground wlking t vrious speed. Overground 25 m wlking t preferred speed; Overground 25 m wlking t low, medium nd high speed; Overground 15 m wlking t low, medium nd high speed. Young subjects: Tredmill wlking t preferred speed, under nd over preferred speed; Overground 30m wlking t preferred speed Elderly subjects: Overground 30m wlking t preferred speed. Tredmill wlking t 6 speed (0.5m/s, 0.75m/s, 1.0m/s, 1.25m/s, 1.5m/s nd 1.75m/s); Overground wlking t preferred speed, slow nd fst speed. Sbtini et l. [26] 5 helthy subjects Tredmill wlking t combintions of 7 speed (3km/s to 6km/s in steps of 0.5km/s). Alvrez et l. [21] 1 helthy subject Overground 10 m wlking t preferred speed. Ojed nd Borenstein [22] 1 helthy subject Overground wlking t norml nd brisk pce; Overground wlking long squre-shped loop pth on 1, 2 nd 4 floors including stirs; Overground 14-minute nd 12-minute wlking long city streets. Song et l. [10] 17 helthy subjects Tredmill wlking nd running t vrious speed from 4.8km/h to 15.4km/s. Yeoh et l. [13] 5 helthy subjects Tredmill wlking nd running t vrious speed from 1km/h to 13km/s. Mrtin et l. [14] 10 helthy subjects Overground 10 m wlking t preferred speed. Bebek et l. [23] 1 helthy subject Overground hlf-hour wlking long loop. Hung et l. [24] 1 helthy subject Overground wlking long the squre, rectngle, J-shped pths nd n thletic trck. Li et l. [27] 8 helthy subjects Tredmill wlking t 6 speed (0.8m/s, 1.0m/s, 1.2m/s, 1.4m/s, 1.6m/s nd 1.8m/s); Overground 100 m wlking t preferred speed. Mrini et l. [18] 10 young subjects; 10 elderly subjects Overground 5 m U-turn, 3 m 8-turn nd 25 m 6-minute wlking. c 2012 by the uthors; licensee MDPI, Bsel, Switzerlnd. This rticle is n open ccess rticle distributed under the terms nd conditions of the Cretive Commons Attribution license (http://cretivecommons.org/licenses/by/3.0/.)