online gait task recognition algorithm for hip exoskeletoncga/b/0161.pdf · 2015-10-21 · online...

6
Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk Lee, Bokman Lim and Youngbo Shim Abstract—In this paper, we propose a novel online gait task recognition algorithm for hip exoskeleton. The proposed algorithm provides an automatic and prompt recognition result in just one step based on the relations between both hip joint angles at the moment of foot contact. Gait task recognition is one of the challenges that walking assist devices must address to offer adaptable and reliable assistance to users. However gait task recognition in hip exoskeleton is challenging because the sensors are very limited and fast gait task recognition is required to prevent inadequate assistance and reduce fall risk. Although in general foot contact event can be considered as crucial information during walking, it has not received attention in hip exoskeletons with no sensors corresponding foot force or pressure. In this study, we exploit foot contact event as a critical point to perform gait task recognition in hip exoskeleton. The proposed algorithm suggests a foot contact estimation method without using any foot force or pressure sensors and a rule- based inference system to recognize a new gait task in real time. Results presented from experiments will demonstrate the validity and performance of the proposed algorithm. I. INTRODUCTION Walking is an important ability for quality of daily life of elderly people. Recently many studies for walking assist activity have been performed in the form of the robotic lower limb exoskeletons [1]–[4]. These kinds of lower limb exoskeletons provide more than 2 DoFs (Degrees of Free- dom) for each leg, hip/knee or hip/knee/ankle joints. They mostly have been developed for patients with SCI and stroke for rehabilitation or for soldiers in need of power assist in combat. However, they tend to be bulky and heavy, so they may not be proper to elderly people. For elderly people, exoskeletons should be compact and light-weight. Partially assisting devices about only knee or ankle in lower limb also have been developed [5][6]. These knee or ankle type exoskeletons may not be appropriate for the elderly people because they increase effective inertial in the legs. We sug- gest that compact and light-weight hip exoskeleton could be a better choice to offer the benefits to elderly people [7]. Each gait task (e.g., level walking, stair ascent, stair descent) has different characteristics [8]. To provide efficient and suitable gait assistance and decrease fall risk, gait task recognition ability is essential. Gait task recognition is challenging for hip exoskeletons because the sensors are very limited and changes in gait task should be recognized as soon as possible to prohibit improper assistance and to decrease fall risks. In many exoskeletons, manual task selection methods have been implemented due to simplicity and reliability but it is inconvenient to users. Even though several studies for gait Authors are with the Samsung Advanced Institute of Technology, Gyeonggi-do, Korea (e-mail: {jw526.jang, kyungroc.kim, jusuk7.lee, bok- man.lim, ddalbo.shim}@samsung.com). intention recognition have been reported ([9]–[13]), to our knowledge no algorithms support the recognition ability for the gait tasks of level walking, stair ascent, and stair descent, and stand using the limited sensors in hip exoskeletons. Fig. 1. A prototype of hip exoskeleton The aim of this study is to develop an automatic and prompt gait task recognition algorithm for several gait tasks (level walking, stair ascent, stair descent, and stand) using the integrated sensors (two potentiometers for measuring left and right hip joint angles and an inertial measurement unit (IMU)) for hip exoskeleton we have developed (Fig. 1). This paper proposes a novel algorithm to recognize changing gait tasks in just one step based on the relations between both hip joint angles at the moment of foot contact. Our insight is that foot contact during gait is an inevitable event and at this moment each gait task can be identified by the amplitudes and difference between both hip joint angles. Generally, foot contact event has not been considered as a usable information in hip exoskeletons without using foot force/pressure sensors [7], [14]–[16]. In this study, however, we exploit foot contact event as a crucial point to perform gait task recognition in hip exoskeleton. To realize our insight, we propose a finite state machine (FSM) about gait tasks and inference, foot contact estimation method without using any kinds of foot force/pressure sensors, and a rule-based gait task inference system. In the literature many studies have been performed to detect foot contact by the accelerometers [17]– [21]. We present an online foot contact estimation method by the threshold integrating the freeze horizons of the vertical acceleration (given from IMU) at the lower back. For gait task inference for level walking, stair ascent and stair descent, we exploit the fuzzy inference system (FIS) since it can 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) Congress Center Hamburg Sept 28 - Oct 2, 2015. Hamburg, Germany 978-1-4799-9993-4/15/$31.00 ©2015 IEEE 5327

Upload: others

Post on 18-Apr-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Online Gait Task Recognition Algorithm for Hip Exoskeletoncga/b/0161.pdf · 2015-10-21 · Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk

Online Gait Task Recognition Algorithm for Hip Exoskeleton

Junwon Jang, Kyungrock Kim, Jusuk Lee, Bokman Lim and Youngbo Shim

Abstract— In this paper, we propose a novel online gaittask recognition algorithm for hip exoskeleton. The proposedalgorithm provides an automatic and prompt recognition resultin just one step based on the relations between both hip jointangles at the moment of foot contact. Gait task recognition isone of the challenges that walking assist devices must addressto offer adaptable and reliable assistance to users. Howevergait task recognition in hip exoskeleton is challenging becausethe sensors are very limited and fast gait task recognition isrequired to prevent inadequate assistance and reduce fall risk.Although in general foot contact event can be considered ascrucial information during walking, it has not received attentionin hip exoskeletons with no sensors corresponding foot force orpressure. In this study, we exploit foot contact event as a criticalpoint to perform gait task recognition in hip exoskeleton. Theproposed algorithm suggests a foot contact estimation methodwithout using any foot force or pressure sensors and a rule-based inference system to recognize a new gait task in realtime. Results presented from experiments will demonstrate thevalidity and performance of the proposed algorithm.

I. INTRODUCTION

Walking is an important ability for quality of daily lifeof elderly people. Recently many studies for walking assistactivity have been performed in the form of the roboticlower limb exoskeletons [1]–[4]. These kinds of lower limbexoskeletons provide more than 2 DoFs (Degrees of Free-dom) for each leg, hip/knee or hip/knee/ankle joints. Theymostly have been developed for patients with SCI and strokefor rehabilitation or for soldiers in need of power assist incombat. However, they tend to be bulky and heavy, so theymay not be proper to elderly people. For elderly people,exoskeletons should be compact and light-weight. Partiallyassisting devices about only knee or ankle in lower limbalso have been developed [5][6]. These knee or ankle typeexoskeletons may not be appropriate for the elderly peoplebecause they increase effective inertial in the legs. We sug-gest that compact and light-weight hip exoskeleton could bea better choice to offer the benefits to elderly people [7]. Eachgait task (e.g., level walking, stair ascent, stair descent) hasdifferent characteristics [8]. To provide efficient and suitablegait assistance and decrease fall risk, gait task recognitionability is essential. Gait task recognition is challenging forhip exoskeletons because the sensors are very limited andchanges in gait task should be recognized as soon as possibleto prohibit improper assistance and to decrease fall risks.In many exoskeletons, manual task selection methods havebeen implemented due to simplicity and reliability but it isinconvenient to users. Even though several studies for gait

Authors are with the Samsung Advanced Institute of Technology,Gyeonggi-do, Korea (e-mail: {jw526.jang, kyungroc.kim, jusuk7.lee, bok-man.lim, ddalbo.shim}@samsung.com).

intention recognition have been reported ([9]–[13]), to ourknowledge no algorithms support the recognition ability forthe gait tasks of level walking, stair ascent, and stair descent,and stand using the limited sensors in hip exoskeletons.

Fig. 1. A prototype of hip exoskeleton

The aim of this study is to develop an automatic andprompt gait task recognition algorithm for several gait tasks(level walking, stair ascent, stair descent, and stand) usingthe integrated sensors (two potentiometers for measuring leftand right hip joint angles and an inertial measurement unit(IMU)) for hip exoskeleton we have developed (Fig. 1). Thispaper proposes a novel algorithm to recognize changing gaittasks in just one step based on the relations between bothhip joint angles at the moment of foot contact. Our insight isthat foot contact during gait is an inevitable event and at thismoment each gait task can be identified by the amplitudesand difference between both hip joint angles. Generally, footcontact event has not been considered as a usable informationin hip exoskeletons without using foot force/pressure sensors[7], [14]–[16]. In this study, however, we exploit foot contactevent as a crucial point to perform gait task recognitionin hip exoskeleton. To realize our insight, we propose afinite state machine (FSM) about gait tasks and inference,foot contact estimation method without using any kindsof foot force/pressure sensors, and a rule-based gait taskinference system. In the literature many studies have beenperformed to detect foot contact by the accelerometers [17]–[21]. We present an online foot contact estimation method bythe threshold integrating the freeze horizons of the verticalacceleration (given from IMU) at the lower back. For gaittask inference for level walking, stair ascent and stair descent,we exploit the fuzzy inference system (FIS) since it can

2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)Congress Center HamburgSept 28 - Oct 2, 2015. Hamburg, Germany

978-1-4799-9993-4/15/$31.00 ©2015 IEEE 5327

Page 2: Online Gait Task Recognition Algorithm for Hip Exoskeletoncga/b/0161.pdf · 2015-10-21 · Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk

provide the linguistic representations and robust inferenceperformance [22]. This performs the inference for a new gaittask at the moment of foot contact detection in real time.

This paper is organized as follows. A novel gait task recog-nition algorithm will be presented in Section II. Evaluationand discussion for the proposed algorithm will be presentedin Section III. Finally, conclusions and future work will bepresented in Section IV.

II. GAIT TASK RECOGNITION ALGORITHM

Fig. 2. Finite State Machine for gait tasks and inference

TABLE IGAIT TASK ID (TID)

Gait Task tidSTAIR ASCENT (SA) 1LEVEL (LE) 2STAIR DESCENT (SD) 3STAND (ST) 4EXCEPTION (EX) 5

A. Finite State Machine (FSM) for Gait Task and Inference

We suggest a finite state machine (FSM) to manage gaittasks and perform inference for a new gait task (Fig. 2).The FSM has three states: STAND (ST), INFER-TASK (IT),and DYNAMIC-GAIT-TASK (DGT). State DGT has foursub-states, STAIR-ASCENT (SA), LEVEL (LE), STAIR-DESCENT (SD), and EXCEPTION (EX). Four transitionconditions are defined as TC1, TC2, TC3, and TC4. Eachgait task ID (tid) and all transition conditions are illustratedin Table I and II, respectively (tc , ω, α, and tms are pa-rameters of constant time, constant absolute angular velocity,constant absolute angle, and a pre-determined maximum stepduration, respectively). If a transition condition TC1 by footcontact estimation is satisfied in any state among state STand DGT, then state IT is activated and at this moment the

TABLE IITRANSITION CONDITIONS

Transition Conditions Description

TC1 foot contact detected

TC2 task inference completed

TC3 If the following (1) and (2) aresustained for a constant time (tc):(1) absolute angular velocities ofboth hip joints < ω

(2) both hip joint angles < α

TC4 NOT TC1 and NOT TC3 andelapsed time from last foot detec-tion > tms

inference for a new gait task is performed by fuzzy inferencesystem (FIS). The tid greater than 4 is reserved for theexceptional cases. At present the exceptional case for state ITis defined for the inference failure of FIS (tid 5). If gait taskinference is completed (TC2 is satisfied), then state DGT isactivated. Transition to a sub-state in DGT is determined byinvestigating tid conveyed. If transition condition TC4 in anystate among SA, LE, or SD is satisfied, then state EX canalso be activated. If a transition condition TC3 in state DGTis satisfied, then state ST is activated.

B. Foot Contact Estimation

Fig. 3. Direction of hip joint angle and coordinates of IMU integrated atthe lower back in hip exoskeleton: V vertical, AP anterior-posterior, MLmedial-lateral

Foot contact estimation is achieved by processing thevertical acceleration value (V) given from the inertial mea-surement unit (IMU) integrated at the lower back in hipexoskeleton as described in Fig. 3. At initial standing upright,we set the vertical acceleration value as zero to eliminatethe bias. Mean value of receding prediction horizon (RPH:black rectangle) of the acceleration is compared to thethreshold (TH) to determine foot contact as seen in Fig. 4.We introduce the freeze horizon (FH: blue rectangle) froma prior knowledge that some duration between foot contactsshould be required. During the freeze horizon, foot contactestimation is not performed. The freeze horizon is usefulbecause it enforces only one foot contact detection in the

5328

Page 3: Online Gait Task Recognition Algorithm for Hip Exoskeletoncga/b/0161.pdf · 2015-10-21 · Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk

Fig. 4. The vertical acceleration trajectory of lower back during levelwalking

same step. The freeze horizon is activated just after footcontact detection and lasts for some duration. We usuallyset it to 300 ms empirically for every gait task from theobservation that a step duration normally exceeds 400 mseven for fast walking and we found that it was substantiallyrobust to walking speed variations through many experi-ments. Definitely, freeze horizon can also be adjusted bywalking speed every step, e.g., half of the last step duration.

C. Gait Task Inference by Fuzzy Inference System (FIS)

Our fuzzy inference system (FIS) has one output namedGAIT that determines one of gait tasks (LE, SA, SD).An input to the FIS is a vector of the form [LeftHipAngRightHipAng AbsHipAngDiff] at the moment of foot con-tact: LeftHipAng (left hip joint angle), RightHipAng (righthip joint angle), AbsHipAngDiff (absolute difference be-tween of both hip joint angles). A variety of membershipfunctions (triangular, trapezoid, ramp, and rectangle member-ship functions) for fuzzy inputs and a triangular membershipfunction for fuzzy output are used. Mandani min-max infer-ence and centroid defuzzification method which calculates acrisp value from center of mass of the fuzzy set are used.

TABLE IIIGAIT TASK RECOGNITION FROM THE CRISP OUTPUT

Crisp Output Gait Task

0.5 < gait task < 1.5 SA

1.5 < gait task < 2.5 LE

2.5 < gait task < 3.5 SD

Membership functions of inputs (LeftHipAng,RightHipAng, and AbsHipAngDiff) and output (GaitTask) are described in Fig. 5 (flexion of hip joint: positivedirection, extension of hip joint: negative direction). Inparticular, membership functions for LOW and MID ofinput AbsHipAngDiff are overlapped significantly sincetheir right parts are extended to the right side to account

Fig. 5. Fuzzy Inference System (NEMID: negative middle, NELOW:negative low, ZERO: zero, POLOW: positive low, POMID: positive middle,POHIGH: positive high, POVHIGH: positive very high, VLOW: very low,MID: middle, VHIGH: very high, SA: stair ascent, LE: level, SD: stair de-scent):(a) Membership Function of Input LeftHipAng and RightHipAng (b)Membership Function of Input AbsHipAngDiff (c) Membership Functionof Output (Gait Task)

Fig. 6. Visualization of gait tasks (V-shape) according to the collectedinputs (LeftHipAng, RightHipAng, and AbsHipAngDiff) at the moment offoot contact in 3D space: stair ascent (solid red circles), level walking (green+), and stair descent (blue circles)

5329

Page 4: Online Gait Task Recognition Algorithm for Hip Exoskeletoncga/b/0161.pdf · 2015-10-21 · Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk

TABLE IVFUZZY RULE BASE (HALF)

Stair Ascent (SA)rule 1: if LeftHipAng is POVHIGH and RightHipAng is POLOW and AbsHipAngDiff is HIGH then GAIT is SArule 2: if LeftHipAng is POVHIGH and RightHipAng is ZERO and AbsHipAngDiff is HIGH then GAIT is SArule 3: if LeftHipAng is POVHIGH and RightHipAng is POLOW and AbsHipAngDiff is VHIGH then GAIT is SArule 4: if LeftHipAng is POVHIGH and RightHipAng is ZERO and AbsHipAngDiff is VHIGH then GAIT is SArule 5: if LeftHipAng is POVHIGH and RightHipAng is NELOW and AbsHipAngDiff is VHIGH then GAIT is SArule 6 : if LeftHipAng is POHIGH and RightHipAng is ZERO and AbsHipAngDiff is VHIGH then GAIT is SArule 7: if LeftHipAng is POHIGH and RightHipAng is ZERO and AbsHipAngDiff is HIGH then GAIT is SA

Level (LE)rule 8: if LeftHipAng is POHIGH and RightHipAng is NELOW and AbsHipAngDiff is HIGH then GAIT is LErule 9: if LeftHipAng is POHIGH and RightHipAng is NEMID and AbsHipAngDiff is VHIGH then GAIT is LErule 10: if LeftHipAng is POMID and RightHipAng is NEMID and AbsHipAngDiff is HIGH then GAIT is LErule 11: if LeftHipAng is POMID and RightHipAng is NELOW and AbsHipAngDiff is HIGH then GAIT is LErule 12: if LeftHipAng is POMID and RightHipAng is NELOW and AbsHipAngDiff is MID then GAIT is LErule 13: if LeftHipAng is POLOW and RightHipAng is NELOW and AbsHipAngDiff is LOW then GAIT is LErule 14: if LeftHipAng is POMID and RightHipAng is ZERO and AbsHipAngDiff is MID then GAIT is LErule 15: if LeftHipAng is POLOW and RightHipAng is ZERO and AbsHipAngDiff is LOW then GAIT is LE

Stair Descent (SD)rule 16: if LeftHipAng is POMID and RightHipAng is POMID and AbsHipAngDiff is VLOW then GAIT is SDrule 17: if LeftHipAng is POMID and RightHipAng is POMID and AbsHipAngDiff is LOW then GAIT is SDrule 18: if LeftHipAng is POMID and RightHipAng is POLOW and AbsHipAngDiff is VLOW then GAIT is SDrule 19: if LeftHipAng is POMID and RightHipAng is POLOW and AbsHipAngDiff is LOW then GAIT is SDrule 20: if LeftHipAng is POLOW and RightHipAng is POLOW and AbsHipAngDiff is VLOW then GAIT is SD

for SD in steep stairs with tall steps and LE in a large steplength, respectively. Table III denotes how gait task can berecognized by the crisp output. Rule base was derived fromthe following observations at the moment of foot contactfor numerous gait tasks; 1) SA - hip joint angle of the legthat last hit the ground (Leg-A) and hip joint of the otherleg (Leg-B) are large positive and small, respectively, andthe difference between them tends to be large; 2) SD - hipjoint angles of Leg-A and Leg-B are positive but not large,and the difference between them tends to be small; 3) LE- hip joint angle of Leg-A and hip joint angle of Leg-Bare positive (but not large) and negative (but not large),respectively, and the difference between them tends to besmaller than that of SA and larger than that of SD. Fig. 6shows the discriminant distribution (V-shape) for gait tasksaccording to inputs at the moment of foot contact in 3Dspace and demonstrates that our insight can be affordable.Half of the rule base for inputs is listed in Table IV. Theother half (not seen here) of rule base can be presented byexchanging LeftHipAng and RightHipAng because they aresymmetric.

III. EVALUATION

A. Experimental Protocol

To evaluate the performance of the proposed algorithm,we conducted experiments in the real environments includinglevel, stair ascent, and stair descent. The vertical accelera-tion (from off-the-shelf IMU AHRS EBIMU-9DOFV2) wasattained with the sampling frequency 100 Hz in the range

of ±8 g (sensitivity 0.24 mg). Three healthy male subjects(age 38.7 ± 5.7 years, height 173 ± 3.6 cm) were tested.They were required to walk along walking environments attheir preferred walking speed. At first they walked along levelenvironment freely, and they stopped in front of the staircasebefore ascending stairs. Then they ascended several stairsand stopped for a moment. They turned 180 degrees andstopped for a moment before descending stairs. Finally theydescended several stairs and stopped. As described above,ascending and descending stairs were performed severaltimes. Subject information and steps each subject walked aredenoted in Table V. Subject A was tested at outdoor level andstairs (approximately rise 170 mm, tread 290 mm). Subject Band C were tested at indoor level and the 8-step experimentalstaircase (rise 180 mm, tread 300 mm). Parameters used forexperiments are denoted in Table VI. For fuzzy inferenceengine, a library fuzzylite was exploited [23]. The accuracyof the algorithm was investigated by integrating the results inall the trials of all three subjects. The recognition accuracy(RA) was defined as follows:

RA =no. of steps recognized correctly

no. of total steps× 100%

B. Results

The results of gait recognition tests for three subjectsare presented in Table VII. The overall online recognitionaccuracies of SA, SD, and LE are 98%, 95%, and 99%,respectively. To demonstrate gait task recognition ability

5330

Page 5: Online Gait Task Recognition Algorithm for Hip Exoskeletoncga/b/0161.pdf · 2015-10-21 · Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk

Fig. 7. Trajectories of hip joint angles (blue line: left joint angle, red line: right hip joint angle), estimated foot contacts (green line), and the recognizedgait task IDs (black line: ID 1 - SA, 2 - LE, 3 - SD, 4 - ST) in a series of gait tasks (ST→LE→SA→LE→ST→LE→SD→LE→ST)

TABLE VSUBJECTS AND WALKING STEPS

Subject A Subject B Subject C

Sex M M M

Age (years) 34 37 45

Height (cm) 177 170 172

SA (steps) 216 112 104

SD (steps) 216 112 104

LE (steps) 440 149 155

TABLE VIPARAMETERS SETTING

TH (m/s2) 1.6 FH length (ms) 300

RPH length (ms) 30 tc (ms) 300

ω (degree/s) 5 α (degree) 20

tms (ms) 2000

in a continuous way, we performed a series of gait tasks(ST→LE→SA→LE→ST→LE→SD→LE→ST) as seen inFig. 7 which describes the trajectories of hip joint angles,estimated foot contacts, and the recognized gait task IDs.In the first level walking period, a right leg (red line)was lifted up first during stand, and foot contact (greenline) was detected when it hit the ground. At this moment,inference for gait task recognition was performed and LEwas recognized correctly in just one step (a one-step delayof recognition). This gait task was sustained until a new gait

TABLE VIIONLINE GAIT TASK RECOGNITION ACCURACY (%)

Gait Task Subject A Subject B Subject C All Subjects

SA (%) 96.8 100 98.1 97.9(209/216) (112/112) (102/104)

SD (%) 94.4 97.3 95.2 95.3(204/216) (109/112) (99/104)

LE (%) 99.8 98.0 98.1 99.1(439/440) (146/149) (152/155)

Mean (%) 97.0 98.4 97.1 97.4

task was recognized (no delay of recognition). Similarly,inferences for stair ascending (SA) and descending (SD)were also performed along walking environments. Standtasks (ST) were recognized by a transition condition TC3 asmentioned before. Interestingly, the turning tasks (althoughnot our concern in this study) were ignored and sustained asST because no foot contacts were detected.

C. Discussion

The proposed algorithm shows that each gait task can berecognized promptly in just one step with high accuracy. Forour limited subjects, the algorithm was robust for variationsin people and environment (e.g., walking speed, stair risechanges) although it requires more evaluations. Furthermore,it may not work properly if the foot contact estimationfails due to slow walking or walking with stealthy steps.Slope ascent and descent have not been considered explicitlyin this paper. In our thinking their characteristics can bechanged according to their inclinations particularly for hip

5331

Page 6: Online Gait Task Recognition Algorithm for Hip Exoskeletoncga/b/0161.pdf · 2015-10-21 · Online Gait Task Recognition Algorithm for Hip Exoskeleton Junwon Jang, Kyungrock Kim, Jusuk

exoskeleton; steep slope walking would be similar to stairwalking, and gentle slope walking would be similar to levelwalking. Without any additional sensors (e.g., foot pitchangle sensor) in hip exoskeleton, it might be challengingto distinguish slope ascent and descent to level walking,stair ascent and descent at the same time with high accuracyand reliability. As a final note, inputs to FIS we presentedshould be used as features for supervised machine learningclassification. To develop classifiers using these featuresthrough machine learning techniques (e.g., support vectormachine (SVM), multilayer perceptron, decision tree) andcompare the performance each other should be an interestingstudy for future research.

IV. CONCLUSIONS AND FUTURE WORK

To recognize gait tasks (e.g., level walking, stair ascent,stair descent) is one of the challenges that walking as-sist devices must address to support adaptable and reliableassistance to users. In this paper we presented an onlinerecognition algorithm using both hip joint angle sensors anda vertical accelerator at the lower back in hip exoskeleton.The experimental results demonstrated that the proposedalgorithm was able to discriminate gait tasks of level walk-ing, stair ascent, stair descent, and stand automatically andpromptly with high accuracy.

In future work, we will comprehensively analyze footcontact estimation method and the robustness of the proposedgait recognition algorithm according to walking speed andvarious environments (e.g, the height of steps). In addition,we will develop the assistance strategies for each gait task,which will be integrated with the gait recognition algorithmto provide adaptable and reliable walking assistance.

REFERENCES

[1] A. Esquenazi, M. Talaty, A. Packel, and M. Saulino, “The rewalkpowered exoskeleton to restore ambulatory function to individuals withthoracic-level motor-complete spinal cord injury,” American Journalof Physical Medicine & Rehabilitation, vol. 91, no. 11, pp. 911–921,2012.

[2] H. Quintero, R. J. Farris, M. Goldfarb et al., “Control and im-plementation of a powered lower limb orthosis to aid walking inparaplegic individuals,” in Rehabilitation Robotics (ICORR), 2011IEEE International Conference on. IEEE, 2011, pp. 1–6.

[3] F. Giovacchini, F. Vannetti, M. Fantozzi, M. Cempini, M. Cortese,A. Parri, T. Yan, D. Lefeber, and N. Vitiello, “A light-weight activeorthosis for hip movement assistance,” Robotics and AutonomousSystems, 2014.

[4] A. B. Zoss, H. Kazerooni, and A. Chu, “Biomechanical design ofthe berkeley lower extremity exoskeleton (BLEEX),” Mechatronics,IEEE/ASME Transactions on, vol. 11, no. 2, pp. 128–138, 2006.

[5] J. E. Pratt, B. T. Krupp, C. J. Morse, and S. H. Collins, “TheRoboKnee: An exoskeleton for enhancing strength and enduranceduring walking,” in Robotics and Automation, 2004. Proceedings.ICRA’04. 2004 IEEE International Conference on, vol. 3. IEEE,2004, pp. 2430–2435.

[6] H. Takemura, T. Onodera, D. Ming, and H. Mizoguchi, “Designand control of a wearable stewart platform-type ankle-foot assistivedevice,” International Journal of Advanced Robotic Systems, Int J AdvRobotic Sy, vol. 9, 2012.

[7] C. L. Lewis and D. P. Ferris, “Invariant hip moment pattern whilewalking with a robotic hip exoskeleton,” Journal of biomechanics,vol. 44, no. 5, pp. 789–793, 2011.

[8] R. Riener, M. Rabuffetti, and C. Frigo, “Stair ascent and descent atdifferent inclinations,” Gait & posture, vol. 15, no. 1, pp. 32–44, 2002.

[9] M. Gorsic, R. Kamnik, L. Ambrozic, N. Vitiello, D. Lefeber,G. Pasquini, and M. Munih, “Online phase detection using wearablesensors for walking with a robotic prosthesis,” Sensors, vol. 14, no. 2,pp. 2776–2794, 2014.

[10] B. Coley, B. Najafi, A. Paraschiv-Ionescu, and K. Aminian, “Stairclimbing detection during daily physical activity using a miniaturegyroscope,” Gait & posture, vol. 22, no. 4, pp. 287–294, 2005.

[11] F. Zhang, Z. Dou, M. Nunnery, and H. Huang, “Real-time implementa-tion of an intent recognition system for artificial legs,” in Engineeringin Medicine and Biology Society, EMBC, 2011 Annual InternationalConference of the IEEE. IEEE, 2011, pp. 2997–3000.

[12] K. Suzuki, Y. Kawamura, T. Hayashi, T. Sakurai, Y. Hasegawa, andY. Sankai, “Intention-based walking support for paraplegia patient,” inSystems, Man and Cybernetics, 2005 IEEE International Conferenceon, vol. 3. IEEE, 2005, pp. 2707–2713.

[13] Y. D. Li and E. T. Hsiao-Wecksler, “Gait mode recognition usingan inertial measurement unit to control an ankle-foot orthosis duringstair ascent and descent,” in ASME 2012 5th Annual Dynamic Systemsand Control Conference joint with the JSME 2012 11th Motion andVibration Conference. American Society of Mechanical Engineers,2012, pp. 743–752.

[14] “Honda walking assist device with stride management system,” http://corporate.honda.com/innovation/walk-assist/, [accessed 09-Jan-2015].

[15] G. Aguirre-Ollinger, “Learning muscle activation patterns via non-linear oscillators: application to lower-limb assistance,” in IntelligentRobots and Systems (IROS), 2013 IEEE/RSJ International Conferenceon. IEEE, 2013, pp. 1182–1189.

[16] R. Ronsse, T. Lenzi, N. Vitiello, B. Koopman, E. van Asseldonk,S. M. M. De Rossi, J. van den Kieboom, H. van der Kooij, M. C.Carrozza, and A. J. Ijspeert, “Oscillator-based assistance of cyclicalmovements: model-based and model-free approaches,” Medical &biological engineering & computing, vol. 49, no. 10, pp. 1173–1185,2011.

[17] H. B. Menz, S. R. Lord, and R. C. Fitzpatrick, “Acceleration patternsof the head and pelvis when walking on level and irregular surfaces,”Gait & posture, vol. 18, no. 1, pp. 35–46, 2003.

[18] A. Mansfield and G. M. Lyons, “The use of accelerometry to detectheel contact events for use as a sensor in fes assisted walking,” Medicalengineering & physics, vol. 25, no. 10, pp. 879–885, 2003.

[19] J. M. Jasiewicz, J. H. Allum, J. W. Middleton, A. Barriskill, P. Condie,B. Purcell, and R. C. T. Li, “Gait event detection using linearaccelerometers or angular velocity transducers in able-bodied andspinal-cord injured individuals,” Gait & Posture, vol. 24, no. 4, pp.502–509, 2006.

[20] P. C. Formento, R. Acevedo, S. Ghoussayni, and D. Ewins, “Gaitevent detection during stair walking using a rate gyroscope,” Sensors,vol. 14, no. 3, pp. 5470–5485, 2014.

[21] A. S. Anna and N. Wickstrom, “A symbol-based approach to gaitanalysis from acceleration signals: Identification and detection of gaitevents and a new measure of gait symmetry,” Information Technologyin Biomedicine, IEEE Transactions on, vol. 14, no. 5, pp. 1180–1187,2010.

[22] J.-S. R. Jang, “ANFIS: adaptive-network-based fuzzy inference sys-tem,” Systems, Man and Cybernetics, IEEE Transactions on, vol. 23,no. 3, pp. 665–685, 1993.

[23] http://www.fuzzylite.com/, [accessed 09-Jan-2015].

5332