Turn Detection and Characterization with Inertial Sensors

Sean Pearson

, Martina Mancini

, Mahmoud El-Gohary

, James McNames

and Fay Horak

APDM Inc. Suite 130, 2828 Southwest Corbett Avenue, Portland, OR, U.S.A.

Oregon Health Sciences University, Neurology, 3181 SW Sam Jackson Park Rd, Portland, OR, U.S.A.

1 OBJECTIVES

Turn detection and characterization in the home is im-

portant for continuous assessment of gait and balance

in people with movement disability. Turning often re-

sults in falling in individuals with movement disor-

ders. Researchers and clinicians would beneﬁt from

a system that identiﬁes and characterizes their daily

mobility behavior to predict their risk of falling, ben-

eﬁts or side effects of treatment, and progression of

disease. The goal of this study is to develop an algo-

rithm that is capable of reliably detecting turns dur-

ing gait with the goal of applying it over long periods

outside a lab environment. Performance of the algo-

rithm is validated against an optical marker system

and video analysis of a subset of the participants.

2 INTRODUCTION

Turning is ubiquitous during activities of daily living.

Nearly every task performed during the day requires

some amount of turning (Glaister et al., 2007). How-

ever, gait research has focused primarily on straight

ahead-walking (Studenski et al., 2003; Ganz et al.,

2007). Difﬁculty turning during gait is a major con-

tributor to mobility disability, falls and reduced qual-

ity of life in elderly and in people with movement dis-

orders. Falls during turning are particularly danger-

ous because they usually result in contact of the femur

with the ground, which results in eight times more hip

fractures compared with falls during straight-ahead

walking (Nevitt et al., 1991; Cumming and Klineberg,

1994; Feldman and Robinovitch, 2007). The ability

to modify our locomotor trajectory by turning safely

is important for functional independence but surpris-

ingly much more difﬁcult for the nervous system to

control than straight-ahead walking.

Objective measures of turning mobility are more

sensitive than gait speed or clinical measures of

mobility to detect impaired mobility. Our studies

have shown that people with movement disability

may exhibit abnormal turning characteristics even

though they have normal straight-ahead walking (Ho-

rak et al., 1992; Salarian et al., 2009; King et al.,

2011). We found that turning measures in the clinic,

such as turning duration and peak turning velocity,

can distinguish between people with movement dis-

orders and healthy age-matched controls, even when

clinical measures of balance or gait did not (Spain

et al., 2010; Zampieri et al., 2010). Other stud-

ies found that measurements of turning discrimi-

nate elderly fallers from non-fallers (Dite and Tem-

ple, 2002). Researchers have suggested that turning-

related neural systems may be more vulnerable to im-

pairments than straight-ahead, linear gait. This is due

to the fact that turning involves more interlimb coordi-

nation, more coupling between posture and gait, and

modiﬁcations of locomotor patterns requiring frontal

lobe cognitive and executive function that plays a role

in postural transitions (King et al., 2012; Herman

et al., 2010).

In this study, we develop a new algorithms for ini-

tial sensors to detect and analyze a wide variety of

types of turning activities performed at different gait

speeds in the laboratory. We validate our algorithm

using a Motional Analysis system as a gold-standard.

As a second method of validating our inertial sensor

algorithms for turning, two raters review the videos

and annotate the beginning and end times of each turn.

In the following sections, we present the algorithm

and discuss the results and future work.

3 METHODS

We examined 17 healthy control subjects wearing a

set of four Motion Analysis reﬂective markers and

an Opal inertial sensor (APDM Inc.) on the lumbar

spine. Opal includes triaxial accelerometers, gyro-

scopes, and magnetometers and record signal data at

128 Hz. Subjects were instructed to walk on a path

composed of a mixed route with short straight seg-

ments and turns ranging from 30 to 180 degrees. Each

subject performed 12 repetitions: 4 at a slow speed, 4

at a preferred speed, and 4 at a fast speed.

Pearson S., Mancini M., El-Gohary M., McNames J. and Horak F..

Turn Detection and Characterization with Inertial Sensors.

DOI: 10.5220/0004647000190022

In Proceedings of the International Congress on Sports Science Research and Technology Support (icSPORTS-2013), pages 19-22

ISBN: 978-989-8565-79-2

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: An example of the Opal (APDM Inc.) inertial

sensor placed on the lumbar.

For the purpose of turn detection and characteriza-

tion, the angular rate about the vertical axis (gravity

aligned) is an ideal measurement. A three-axis gyro-

scope is capable of measuring this vertical axis angu-

lar velocity, but aligning the sensor and maintaining

accurate alignment is difﬁcult. An accelerometer can

be used to measure the direction of gravity during a

stationary period and then, the gyroscope measure-

ments could be projected on to that axis throughout

the trial.

Orientation angles are commonly estimated using

sensor fusing, taking advantage of the accelerome-

ter measurement of gravity to correct drift from inte-

gration of angular velocity measurements (El-Gohary

and McNames, 2012). We take advantage of the ori-

entation estimates to obtain angular velocity about the

vertical axis using appropriate transformation. Opal

sensor provide orientation estimates, q, in quaternion

form and can be used directly to transform body frame

sensor measurements into the inertial frame. We

transform the angular velocity measurements in the

body frame, ω

, to that in the inertial frame (with x-

y-z axes corresponding to magnetic North, East, and

Up), ω

using the following quaternion multiplication.

= q ∗ ω

∗ q

−1

(1)

From the inertial frame angular velocity we extract

the z component, w

about the vertical axis. This an-

gular velocity about the vertical axis can be integrated

to obtain an estimate of the relative turn angle.

(2)

This turn angle drifts over time due to integration er-

ror, but over short time periods is very accurate. The

Opal magnetometer could be used to help compen-

sate for this error. However, since only relative angles

during short period while turning are of importance

for detecting and characterizing turns, the potential

drawbacks of being susceptible to magnetic interfer-

ence from nearby objects outweigh the beneﬁts of us-

ing the magnetometer (Bachmann et al., 2004).

The vertical axis angular velocity (ω

) is lowpass

ﬁltered with a 1.5 Hz cutoff frequency Butterworth

ﬁlter to remove high frequency components. Candi-

date turns are then detected from segments where this

ﬁltered angular velocity is greater than 15 degrees per

second. Start and end of each turn are set to the point

where the ﬁltered angular velocity drops below 5 de-

grees per second. The precise cutoff value has little

effect on the total turn duration or angle, and is in-

tended to account for a slight bias which could prevent

the angular velocity from dropping exactly to zero. It

is difﬁcult for humans to make more than a very slight

turn with a duration < 0.5 seconds or to complete an

extremely slow turn with a duration > 10 seconds,

during gait. Therefore, only turns with duration be-

tween 0.5 and 10 seconds, with turn angles over 45

degrees are considered. We combine any turns in the

same direction separated by a brief pause < 50 mil-

liseconds.

4 RESULTS

Turn detection was performed on the optical Motion

Analysis data. In a subset of trials, a standard video

was also reviewed by two raters who annotated begin-

ning and end times of turns. Timing accuracy of the

turn detection is important for characterizing turn du-

ration, therefore, we use one based on time intervals

rather than a turn by turn accuracy comparison. All

turn detection metrics were resampled at 128 Hz, and

comparison was calculated on a sample by sample ba-

sis, rather than by turns,

Table 1 shows sensitivity of the inertial algo-

rithm detecting turns compared to Motion Analysis

and video raters. Sensitivity of the inertial algorithm

is 0.86 compared with the optical marker algorithm.

Similarly, sensitivity for the inertial algorithm is 0.77

compared to both video rater 1 and 2. The optical al-

gorithm has sensitivity of 0.62 and 0.65 with video

rater 1 and2 respectively. Video rater 1 has a sensitiv-

ity of 0.89 compared with video rater 2.

Table 1: Sensitivity.

Optical Video 1 Video 2

Inertial 0.86 0.77 0.77

Optical 0.62 0.65

Video 1 0.89

Table 2 show speciﬁcity of the inertial algorithm

detecting turns compared to Motion Analysis and

icSPORTS2013-InternationalCongressonSportsScienceResearchandTechnologySupport

−1500 −1000 −500 0 500 1000 1500 2000

−2000

−1500

−1000

−500

500

1000

1500

X position (mm)

Y position (mm)

Figure 2: The blue trace is the X-Y position of the center of

mass from optical markers. Overlaid in red are the segments

detected as turns by the inertial algorithm, and in green the

segments detected as turns by the optical algorithm.

0 10 20 30 40 50

−100

100

200

300

400

500

Time (s)

Yaw Angle (deg)

Figure 3: The blue trace is the yaw angle from the optical

system. The dashed blue line is relative yaw angle from the

inertial algorithm. The dots represent onset (green) and end

(red) of turns detected by the optical algorithm. The vertical

dashed lines represent onset (green) and end (red) of turns

detected by the inertial algorithm.

video raters. Speciﬁcity of the inertial algorithm is

0.70, 0.66 and 0.54 with the optical marker algorithm,

video rater 1 and 2 respectively. The optical marker

algorithm has speciﬁcity of 0.77 and 0.68 with video

rater 1 and 2. Rater 1 has a speciﬁcity of 0.64 com-

pared with video rater 2.

Table 2: Speciﬁcity.

Optical Video 1 Video 2

Inertial 0.70 0.66 0.54

Optical 0.77 0.69

Video 1 0.64

The number of detected turns for the inertial al-

gorithm averaged 9.6 (sd = 1.4) across all subjects

and trials. The algorithm based on optical markers

detected an average of 10.4 turns (sd = 3.4). The pre-

scribed pattern had 10 turns larger than 45 degrees,

including one pair of 90 degree turns to pick up a bas-

ket. Table 3 shows turn metrics detected by the in-

ertial algorithm, including average number of terms,

turn peak and mean velocity in degrees per seconds,

and turn duration in seconds.

Table 3: Turn metrics.

Metric Slow (sd) Normal (sd) Fast (sd)

# of turns 9.6 (1.1) 9.6 (0.9) 8.3 (1.3)

Peak Vel. 112.2 (16.8) 131.2 (19.4) 169.7 (27.8)

Mean Vel. 56.8 (6.6) 64.0 (8.2) 76.8 (10.7)

Duration (s) 2.2 (0.3) 2.1 (0.3) 2.1 (0.4)

5 DISCUSSION

Turning is ubiquitous during activities of daily living.

Nearly every task performed during the day requires

some amount of turning (Glaister et al., 2007). How-

ever, gait research has focused primarily on straight

ahead walking. In this study, we use inertial sensors

including gyroscopes and accelerometers to charac-

terize turns during gait. We validated our inertial al-

gorithm using Motional Analysis system and turn data

from 2 raters analyzing video recordings of healthy

subjects.

Inertial sensors can be used to detect turns during

walking at least as well as an optical marker system.

Subjects were very inconsistent with the basket turns,

with some avoiding it altogether by picking up the

basket without any turning at the waist. In this close

pattern with the rapid series of turns it is sometimes

the case that multiple turns become blended together,

where a subject will transition ﬂuidly from one to the

next rather than following the line precisely. Both of

these factors contribute to variability in the number of

detected turns, especially in the fast trials.

Turn detection and characterization in the home

is important for continuous assessment of gait and

balance in people with movement disability. Asking

an individual to execute a turn in a clinical environ-

ment often does not reveal their impairments. We

hypothesis that we can use the algorithm developed

in this study to measure locomotor activities and to

distinguish turning characteristics in healthy subjects

and individuals with movement disorders, in the home

throughout the day, using wearable inertial sensors.

ACKNOWLEDGEMENTS

Research reported in this study was supported by

the National Institute of Neurological Disorders and

Stroke (NINDS) of the National Institutes of Health

under award number R41NS076088-02.

TurnDetectionandCharacterizationwithInertialSensors

REFERENCES

Bachmann, E. R., Yun, X., and Peterson, C. (2004). An in-

vestigation of the effects of magnetic variations on in-

ertialklagnetic orientation sensors. In Proceedings of

the 2004 IEEE International Conference on Robotics

and Automation, pages 1115–1122.

Cumming, R. G. and Klineberg, R. J. (1994). Fall frequency

and characteristics and the risk of hip fractures. Jour-

nal of the American Geriatrics Society, 42(7):774.

Dite, W. and Temple, V. A. (2002). Development of a

clinical measure of turning for older adults. Amer-

ican journal of physical medicine & rehabilitation,

81(11):857–866.

El-Gohary, M. and McNames, J. (2012). Shoulder And El-

bow Joint Angle Tracking with Inertial Sensors. IEEE

Transactions on Biomedical Engineering, 59(9):577–

585.

Feldman, F. and Robinovitch, S. N. (2007). Reducing hip

fracture risk during sideways falls: evidence in young

adults of the protective effects of impact to the hands

and stepping. Journal of biomechanics, 40(12):2612–

2618.

Ganz, D., Bao, Y., Shekelle, P., and Rubenstein, L. (2007).

Will my patient fall. the journal of the American Med-

ical Association, 297(1):77–86.

Glaister, B., Bernatz, G., Klute, G., and Orendurff, M.

(2007). Video task analysis of turning during activ-

ities of daily living. Gait & Posture, 2(25):289–294.

Herman, T., Giladi, N., and Hausdorff, J. M. (2010). Prop-

erties of the ?timed up and go?test: more than meets

the eye. Gerontology, 57(3):203–210.

Horak, F., Jones-Rycewicz, C., Black, F., and Shumway-

Cook, A. (1992). Effects of vestibular rehabilitation

on dizziness and imbalance. Otolaryngol Head Neck

Surg, 106(2):175–180.

King, L., Mancini, M., Priest, K., and Salarian, A. (2011).

Instrumented and clinical measures of turning in

Parkinson’s disease. Journal Neurologic Physical

Therapy, page In Press.

King, L., Mancini, M., Priest, K., Salarian, A., Rodrigues-

de Paula, F., and Horak, F. (2012). Do clinical scales

of balance reﬂect turning abnormalities in people with

Parkinson’s disease? Journal of Neurologic Physical

Therapy, 36(1):25–31.

Nevitt, M. C., Cummings, S. R., and Hudes, E. S. (1991).

Risk factors for injurious falls: a prospective study.

Journal of gerontology, 46(5):M164–M170.

Salarian, A., Zampieri, C., Horak, F., Carlson-Kuhta, P.,

Nutt, J., and Aminian, K. (2009). Analyzing 180 de-

grees turns using an inertial system reveals early signs

of progression of Parkinson’s disease. In Proceedings

of the 2009 IEEE Eng Med Biol Soc, pages 224–227.

Spain, R., St George, R., Salarian, A., Bourdette, D., and

Horak, F. (2010). Validation of an instrumented test

of mobility, iMOBILITY, in multiple sclerosis. In

American Academy of Neurology 62

Annual Meet-

ing, Toronto, Canada. American Academy of Neurol-

ogy.

Studenski, S., Perera, S., Wallace, D., Chandler, J., Dun-

can, P., Rooney, E., Fox, M., and Guralnik, J. (2003).

Physical performance measures in the clinical set-

ting. Journal of the American Geriatrics Society,

51(3):314–322.

Zampieri, C., Salarian, A., Carlson-Kuhta, P., Aminian, K.,

Nutt, J., and Horak, F. (2010). The instrumented timed

up and go test: potential outcome measure for disease

modifying therapies in Parkinson’s disease. J Neurol

Neurosurg Psychiatry, 81(12):171–176.

icSPORTS2013-InternationalCongressonSportsScienceResearchandTechnologySupport