Book Title Book Editos IOS Pess, 005 Weaable inetial sensos fo am motion tacking in home-based ehabilitation Huiyu Zhou a,, Huosheng Hu a and Nigel Hais b a Univesity of Essex, Colcheste, CO4 3SQ, UK b Univesity of Bath, Bath, BA RL, UK Abstact. We in this pape intoduce a eal-time human am movement tacking system that can be used to aid the ehabilitation of stoke patients. A 3-axis inetial senso is used to captue am movements in 3-D space and in eal time. The tacking algoithm is based on a kinematical model that just consides the human foeams at this stage. To impove accuacy and consistency, a Lagangian based filteing stategy is adopted. The expeimental esults demonstate that the poposed famewok can be used to tack foeam motion. Keywods. Rehabilitation, motion tacking, inetial senso, filteing. Intoduction Stoke is the biggest cause of sevee disability in the UK. Ten thousand people each yea expeience a fist stoke, and a futhe 30,000 have a futhe stoke [5]. Moe than 75% of these people equie multi-disciplinay assessments and appopiate ehabilitative teatments afte they ae dischaged fom hospital []. This places a lage demand on community healthcae sevices, which often have quite limited theapy esouces. As a esult, thee is consideable inteest in taining aids o intelligent systems that conduct ehabilitation in patient s home envionment athe than in hospital. The goal of ehabilitation is to enable a peson who has expeienced a stoke to egain the highest possible level of function. Although some functional abilities may be spontaneously estoed soon afte a stoke, ecovey is an ongoing pocess and the patient must pefom epeatative movements, coecting any undesied motion behavio in ode to egain fine contol of the uppe o lowe limbs. Duing these ehabilitation execises, if the movements of stoke patients can be tacked incoect motion pattens can be eadily identified and coected. Theefoe, devices that can accuately tack the position of limbs in space is an essential component of such a ehabilitation system. Hee, we will biefly summeize the existing human tacking systems, which can be classified as non-vision based, vision-based with makes, vision-based without makes, and obot-guided systems. Non-vision based: Systems can deploy sensos, e.g. inetial, mechanical and magnetic ones, to continuously collect movement infomation. Fo example, inetial and Coespondence to: H. Zhou, Depatment of Compute Science, Univesity of Essex, Colcheste, UK. Tel.: +44 06 87409 Fax: +44 06 87788 E-mail: zhou@essex.ac.uk.
H. Zhou et al. / IAS-9 Figue. Illustation of the poposed ehabilitation system. magnetic sensos based devices (i.e. [6]) exploit mico-electomechanical systems (MEMS). These devices can be used in most cicumstances without any limitations (i.e. illumination, tempeatue, o space, etc.). They have bette pefomance in accuacy against mechanical sensos based devices. The main dawback of using inetial sensos is that accumulating eos (o dift) will become significant afte a shot peiod of time. Vision-based with makes: Numeous make-based tacking systems ae nowadays available in makets and academics. CODA (Chanwood Dynamics, UK) and Qualisys (Qualysis AB, Sweden) ae two examples: the fome uses active" makes, and the latte exploits passive" makes that ae gasped by the allocated cameas. Howeve, these systems inevitably suffe fom the occlusion poblem due to the existence of the line of sight. Vision-based without makes: As a less estictive motion captue technique, makeless based sensing is capable of patially ovecoming the occlusion poblem as it concens boundaies o featues on human bodies. This appoach has a obust pefomance but is inefficient in computation and less effective in igoous cicumstances. To solve the inefficiency poblem, fo example, Mihailidis et al. [3] designed a sensing agent fo an intelligent envionment that assists olde people with dementia duing thei life. It has shown pomising esults of -D hand motion tacking. Howeve, this system failed to povide 3-D motion estimation. Robot-guided: To find out whethe execise theapy influences plasticity and ecovey of the bain following a stoke, an automatic system, named MIT-MANUS, was designed to move, guide, o petub the movement of a patient s uppe limb, whilst ecoding motion-elated quantities, e.g. position, velocity, o foces applied []. This system was successfully implemented. Unfotunately, duing execise then ams has to be attached to the obot am so the patient is unable to cay out fee motion. In this pape, a kinematic model of human foeam motion is developed, which can povide consistent estimates of foeam movements such as position and oientation based on a commecially available inetial senso MT9 (Xsens Motion Technology, Holland). At this peliminay stage, the tacking system equies the elbow joint to be fixed,
H. Zhou et al. / IAS-9 3 Z X Y S (0,0,0) senso Human am S (x,y,z) wz Foeam S 3 (,0,0) Fixed point Figue. Kinematics of a human foeam (the fixed point is the elbow joint). howeve late wok will emove this constaint. The designed motion tacking famewok has been integated within a home-based ehabilitation system illustated in Figue. The est of the pape is oganized as follows. Section pesents a novel appoach to conduct the 3-D tacking based on the collected angula ate, which exploits the kinematics of the human foeam movements. A weighted least squaes filteing method is poposed in Section 3 that educes the eos whose Euclidean distance is lage than a theshold. Section 4 intoduces expeimental esults. Conclusions and futue wok ae finally povided in Section 5.. Kinematic modelling of human am motion A kinematic model of the foeam is poposed in this section. Conside a igid human foeam moving in the 3-D inetial space. Figue shows the kinematics of a human foeam, whee the elbow pesumably is fixed and the inetial senso is attached neaby the wist. is the distance between the the cente of the senso and the fixed point, which can be known a pioi. Let the coodinates of an abitay point be denoted by u(x, y, z), then one can have pojected coodinates on thee othogonal planes, i.e. x-y plane { x = z cos ω z y = z sin ω z () y-z plane { y = x cos ω x z = x sin ω x () x-z plane { x = y sin ω y z = y cos ω y (3) whee ω x, ω y and ω z ae Eule angles aound x-, y- and z-axis, espectively.
4 H. Zhou et al. / IAS-9 Assume that a = (cos ω x ), b = (cos ω y ) and c = (cos ω z ). x then is x = ± ( + abc)( c + bc abc) + abc (4) Substituting Equation (4) to the emaindes of Equation (), () and (3), the solutions fo y and z will be explicitly available as follows { y = z sin ω z z = x sin ω x (5) Solutions fo x, y and z ely on the estimated Eule angles, which ae the integation of the collected tuning-ates [7]. The angles usually accompany noise o difts due to the inetial popeties. These eos might be up to 5 o moe, which can significantly bias the estimated 3-D positions. One example is shown in Figue 3, whee one can see significant discepancy between the estimates of x-axis position by a standad motion tacke and ou method, espectively. So, a eal-time filte is needed fo estoing the tue data. 3. Real time filteing We intend to econstuct a tue data point u fom its obsevation ũ (e.g. x, y and z positions) ũ = u + ɛ (6) whee ɛ is noise o an unknown eo. One of the common denoising techniques is to minimise a function of gadient given as min F ε,p (u), and, F ε,p (u) = ε u p dx + λ ũ u, (7) u Ω whee λ 0 is a Lagange multiplie, and ε is a egulaization coefficient: ε u = (u + ε ). (8) To solve the minimisation poblem, Equation (7), we use the Eule-Lagange equation as follows: ( u ) u = ε u p + β(ũ u), (9) whee β (= λ p ) is the constaint paamete (indicating the descent diection), and λ can be available if we take the deivative fo Equation (7) with espect to u and then set it to zeo. The equied deivatives ae yielded as follows:
H. Zhou et al. / IAS-9 5 8 6 motion tacke ou method 4 X POSITION (CM) 0 4 6 8 0 0 00 400 600 800 000 SAMPLES Figue 3. Discepancy between the estimates by a standad motion tacke and ou method, espectively. x ω x = ab c sin ω x (+a b c ) (+a b c )d x ω y = a bc sin ω y (+a b c ) a bc d sin ω y (+a b c ) + a bc sin ω y( c +b c a b c ) (+a b c )d + ( bc sin ω y+a bc sin ω y) d x ω z = a b c sin ω z (+a b c ) a b cd sin ω z (+a b c ) + a b c sin ω z( c +b c a b c ) (+a b c)d + (c sin ω z b c sin ω z+a b c sin ω z) d z ω x = x x sin ωx ωx + x a z x ω y = x x sin ωx ωy x z ω z = x x sin ωx ωz y x ω x = z z sin ωz ωx z y ω y = z z sin wz ωy y z ω z = z z sin ωz ωz + z c. z ab c d sin ω x (+a b c ) + ab c sin ω x( c +b c a b c ) (0) whee d = ( + a b c )( c + b c a b c). When p = and ε = 0, Equation (7) will become a poblem of total vaiation (TV) [4]. Howeve, noise o outlies often exist in the estimation because of iegula motion o the soft-skin effect, etc. These points possibly mislead the solution by Equation (7) to an incoect position. To emove these eos befoe computing, we exploe a smoothing scheme: in a small neighbohood, the point with the vaiance of the Euclidean distance between its position and the weight cente of the egion lage than twice of the aveaging vaiance will be consideed as eos and be emoved fom the computation list. Fom the efficient point of view, the size of the neighbohood is 5-7. This filteing stategy is illustated in Figue 4, whee squae 7 is the filteed output of cicle 7 given the pevious seven points, and the dotted cicle consists of the inlies. To conduct a fast minimization fo Equation (8), we use the iteative Levenbeg- Maquadt (L-M) algoithm. Although L-M can only seek a locally optimal solution, we wish that it conveges to the coect solution by setting the stating point as the weight cente of the investigated point aea. 4. Expeimental wok To evaluate the pefomance of the filte-on kinematic model against the filte-off technique, we make use of a commecial human motion tacking system ( Qualysis"). This
6 H. Zhou et al. / IAS-9 4 5 6 7 8 0 3 9 7 Figue 4. Illustation of the denoising stategy. system povides absolute positions fo the human am movements duing tajectoies. Qualysis uses eto-eflective ball makes that can be identified by the cameas suounding the object peson. It diectly econstucts 3-D positions of the moving human limbs afte pope calibation is achieved. To set up the expeimental envionment, a Qualysis make is attached to one side of the wist joint while the MT9 inetial senso is mounted on the othe side. This method has bette accuacy than the configuation whee the make and the inetial senso both ae placed on the same side. The distance between the elbow joint (fixed point) and the cente of the senso is 4 cm. Figue 5 illustates the expeiemental set-up in which the Qualysis" has 3 video cameas to obseve uppe limb movements via the ball make. Befoe the expeiments stat, we need to un a calibation fo aligning the coodinate system of the inetial senso with that of the Qualysis system. Inetial measuements can be obtained by going though the following pocedue. The MT9 inetial senso collects data and then pe-filtes it to emove high-fequency noise. This is followed by computing the 3-D position and oientation based on the poposed kinematical modelling. Due to the pesence of noise and eos a dynamic filte is applied fo suppessing the noise o eos. This pocess will un continuously until it is manually teminated. In the expeiments, these two systems wok independently and asynchonously, so we cannot diectly compae thei 3-D motion estimations. A simple way to get aound this poblem is to develop a least-squaed fitting" method, which is used to fit individual estimates to one-cycle cuves (see Figues 6 and 7). By doing this, one can easily judge whose pefomance is bette. A numbe of epeated motion pattens have been captued in ode to obtain compehensive compaisons. In this pape, only two cases ae investigated: () up-down motion, and () cyclic otation in 3-D space. It should be noticed that the elbow joint is kept stationay on the testing bench duing the foeam movements. Case : up-down movements In this expeiment, the up-down foeam motion is captued by both the MT9 inetial senso and the Qualysis" tacking system. Figue 6 shows 3-D motion tajectoies of the foeam movements that ae captued. Note, the solid line is the Qualysis" data, the dashed line is the inetial data afte filteing, and the dotted line is the inetial data without filteing. These esults ae the mean values poduced by the weighted least-squae calculation. The mean eo between the Qualysis" data and the inetial data afte filte-
H. Zhou et al. / IAS-9 7 video camea X Z MT9 Y 3 Figue 5. Illustation of the expeimental set-up. ing is.09 cm (SD: 0.47 cm), while the mean eo between the Qualysis" data and the inetial data without filteing is.5 cm (SD:.05 cm). Case : cyclic otation Figue 7 shows 3-D motion tajectoies of this motion style, and the symbols can be efeed to the ealie desciption. The mean eo between the Qualysis" data and the inetial data afte filteing is.73 cm (SD: 0.93 cm), while the mean eo between the Qualysis" data and the inetial data without filteing is.7 cm (SD:.56 cm). 5. Conclusions and futue wok We have pesented a weaable inetial sensing based tacking system that integates kinematics of human am movements and a dynamic filteing stategy. Compaed to the commecial tacking system Qualysis" that uses makes, ou system is able to delive eal time human foeam motion estimation with a simple set-up. Also, ou system has achieved easonably accuate esults. The futue wok will be addessed to extend the ideas pesented hee in ode to implement a potable device that consides eal theejoints am movements with highe degees of feedom. The kinematic model of the uppe am is the same as that of the foeam, although the fixed point is now located at the shoulde. Anothe MT9 senso will be used to epesent this motion. Acknowledgements This is pat of SMART Rehabilitation Poject funded by the UK EPSRC unde Gant GR/S9089/0. We ae gateful to Ms Yaqin Tao fo helping us to set up the expeiments. Refeences [] J.H. Cauaugh, S. Kim, Two coupled moto ecovey potocols ae bette than one electomyogam-tiggeed neuomuscula stimulation and bilateal movements, Stoke 33 (00), 589 594.
8 H. Zhou et al. / IAS-9 3 5 0 Z (CM) 5 0 5 5 0 5 0 5 6.5 6 5.5 5 X (CM) 4.5 Y (CM) Figue 6. Case - Up-down motion tajectoies: the Qualysis" data (solid line - ), the inetial data afte filteing (dashed line - ) and the inetial data without filteing (dotted line - 3). 5 3 0 Z (CM) 5 0 0 Y (CM) 0 0 5 0 5 X (CM) Figue 7. Case - Cyclic motion tajectoies: the Qualysis" data (solid line - ), the inetial data afte filteing (dashed line - ) and the inetial data without filteing (dotted line - 3). [] H.I. Kebs, B.T. Volpe, M.L. Aisen, N. Hogan, Inceasing poductivity and quality of cae: obot-aided neo-ehabilitation, Jounal of ehabilitation eseach and development 37 (000), 639 65. [3] A. Mihailidis, B. Camichael, J. Boge, The use of compute vision in an intelligent envionment to suppot aging-in-place, safety, and independence in the home, IEEE Tansaction on infomation technology in medicine 8 (004), 38 47. [4] L. Rudin, S. Oshe, E. Fatemi, Nonlinea total vaitation based noise emoval algoithms, Physica 60 (99), 59 68. [5] The Stoke Association, Speaking out about stoke sevices, London: the Stoke Association, 00. [6] H. Zheng, N.D. Black, and N.D. Hais, Position-sensing technologies fo movement analysis in stoke ehabilitation, Medical & Biological Engineeing & Computing 43 (005), 43 40. [7] H. Zhou, H. Hu, Inetial motion tacking of human am movements in home-based ehabilitation, Poc. of IEEE Int. Conf. on Mechatonics and Automation, Niagaa Falls, Canada, 005.