Professional Documents
Culture Documents
1 Activity Classification Using A Single Chest Mounted Tri-Axial Accelerometer
1 Activity Classification Using A Single Chest Mounted Tri-Axial Accelerometer
a r t i c l e i n f o a b s t r a c t
Article history: Accelerometer-based activity monitoring sensors have become the most suitable means for objective
Received 29 November 2010 assessment of mobility trends within patient study groups. The use of minimal, low power, IC (integrated
Received in revised form 3 May 2011 circuit) components within these sensors enable continuous (long-term) monitoring which provides
Accepted 6 May 2011
more accurate mobility trends (over days or weeks), reduced cost, longer battery life, reduced size and
weight of sensor. Using scripted activities of daily living (ADL) such as sitting, standing, walking, and
Keywords:
numerous postural transitions performed under supervised conditions by young and elderly subjects, the
Physical activity
ability to discriminate these ADL were investigated using a single tri-axial accelerometer, mounted on the
Accelerometer
Gyroscope
trunk. Data analysis was performed using Matlab® to determine the accelerations performed during eight
Discrete wavelet transform different ADL. Transitions and transition types were detected using the scalar (dot) product technique and
ADL vertical velocity estimates on a single tri-axial accelerometer was compared to a proven discrete wavelet
Postural transitions transform method that incorporated accelerometers and gyroscopes. Activities and postural transitions
Scalar product were accurately detected by this simplified low-power kinematic sensor and activity detection algorithm
Dot product with a sensitivity and specificity of 86–92% for young healthy subjects in a controlled setting and 83–89%
for elderly healthy subjects in a home environment.
Crown Copyright © 2011 Published by Elsevier Ltd on behalf of IPEM. All rights reserved.
1350-4533/$ – see front matter. Crown Copyright © 2011 Published by Elsevier Ltd on behalf of IPEM. All rights reserved.
doi:10.1016/j.medengphy.2011.05.002
1128 A. Godfrey et al. / Medical Engineering & Physics 33 (2011) 1127–1135
review of accelerometry and its application for human movement (a) provide a more straightforward, low power sensor array (one
and activity detection. The authors direct the readers to that body tri-axial accelerometer) and
of work where earlier systems and their description, accuracy, etc. (b) a reduction in the computational intensity of the activity clas-
are presented. sification algorithm.
Fig. 2. A flow chart summarizing the VESPA algorithm. This new algorithm eliminates the need for complex computations on gyroscope signals to eliminate drift introduced
by integration. Simple vertical velocity estimates and scalar products provide a complete activity monitoring system on a single accelerometer-based sensor.
A. Godfrey et al. / Medical Engineering & Physics 33 (2011) 1127–1135 1129
Fig. 3. (i) sin() after applying the DWT with decomposition into 9 scales by a fifth order Coiflet wavelet (coif5). (ii) Original vertical acceleration signal (avs ). (a, iii) DWT(avs )
between the scales of 5 and 9 with a coif5 wavelet. The circles (green) represent the detection of the type of transition, here it is a SiSt. The nearest maximum and minimum
to the time of posture transition tPT . (b, iii) Similar analysis/technique as (a, iii) but here showing a StSi. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of the article.)
2. Methods subject were recorded during each activity. The inertial sensor con-
sists of a tri-axial accelerometer sensor, consisting of two bi-axial
2.1. Sensor design Analog Devices ADXL2101 accelerometers which are sensitive to
Fig. 5. (i) This the calculation of from the scalar (dot) product (a• b) of the static vector, a, and the tri-axial accelerometer output, b, for change in body tilt – SiSt or StSi.
(ii) Original vertical acceleration signal (avs ). (a) SiSt transition for a YHS – for comparison with Fig. 3(a) – shows the potential of the new VESPA algorithm over the previous
method. (iii) Vertical estimate for the transition of SiSt – greater maximum positive peak in vertical estimate, |b| > |a| (a and b determined from t(PT) – nearest maxima). (b)
StSi transition for a YHS – for comparison with Fig. 3(b). (iii) Vertical estimate for the transition of StSi – greater maximum negative peak in vertical estimate, |a| > |b| (a and
b determined from t(PT) – nearest maxima).
1132 A. Godfrey et al. / Medical Engineering & Physics 33 (2011) 1127–1135
Fig. 6. (i) Posture duration (TD) times standing to sitting (StSi) for young healthy subjects as determined by both methods for PTs (mean with standard error). (ii) Posture
duration (TD) times for sitting to standing (SiSt) young healthy subjects as determined by both methods for PTs (mean with standard error). (iii) Standing to sitting (StSi) and
sitting to standing transition (SiSt) duration times for elderly healthy subjects as determined by the VESPA algorithm (mean with standard error).
and Scalar Product Activity (VESPA) algorithm on EHS where their performed by subjecting the accelerometer to a number of known
movement or postural transition signatures (pattern) may differ static angles versus gravity [19].
from YHS.
2.4. Signal processing
2.3. Calibration
The methods used to determine the subject’s activity were the
Calibration of the tri-axial accelerometer and gyroscope sen- use of multi-resolution analysis in the form of the discrete wavelet
sors was performed using previous methods as outlined by Bourke transform (for comparison purposes) and the new VESPA algo-
et al. [4]. The method of gyroscope calibration works by comparing rithm.
the standard deviation of three gyroscope signals to a 3-element
Gyroscope Static Threshold Vector (GSTV). The GSTV elements are 2.4.1. The discrete wavelet transform
the mean and standard deviation of each axis of the tri-axial gyro- The purpose of the discrete wavelet transform (DWT) is a
scope held static for a period of 5 s [18]. Accelerometer calibration is time-frequency representation of a signal. The DWT owes its
A. Godfrey et al. / Medical Engineering & Physics 33 (2011) 1127–1135 1133
Table 1
Sensitivity and specificity of PT and walking for the 10 YHS and 10 EHS using VESPA and the method by Najafi et al. (mean ± standard deviation).
functionality to the fast pyramid algorithm [20,21]. The pyramid 2.4.3. VESPA algorithm – scalar product
algorithm has both forward and backward (inverse) algorithms to This paper proposes a new method for trunk angle tilt estimation
compute the wavelet transform. The backward algorithm recon- using a tri-axial accelerometer mounted at the chest. This novel
structs the original signal from the component wavelets [21]. The method makes use of the scalar (dot) product. This method deals
DWT is given in Eq. (1) in terms of its recovery transform, where with vector multiplication as is defined by Eq. (3), where a, is a row
d(k,l) is a sampling of the wavelet coefficients at discrete points k vector and b is a column vector of equal length.
and l with the mother wavelet, [20]: ⎡ ⎤
b1
∞
∞ n
Here we apply the same techniques adopted by Najafi et al. to For angle tilt detection the vector a, is a 1 × 3 matrix rep-
validate the new algorithm of activity monitoring with the use of resenting the mean output value from each axis of the tri-axial
a single tri-axial accelerometer – VESPA. Use of the DWT by Najafi accelerometer during which the patient is standing and inactive.
et al. was two-fold: Similarly, b, is a 3 × N matrix, which represents the accelerometer
output for each axis, where N denotes the length of the accelerom-
(a) Firstly, DWT (with a suitable mother wavelet between the eter signal, over the entire recording period. The scalar product of
required frequency bands/scales) eliminated low frequency a and b is represented in Eq. (4). Thus, the resulting vector, or angle
drift which resulted when the gyroscope signal was integrated tilt (), is a 1 × N matrix represented in Eq. (5). The value represents
to determine change of trunk tilt (). the angle difference away from the static vector, a:
(b) Secondly, the DWT was applied to the vertical acceleration
(avs ) to determine the type of postural transition (StSi or SiSt) b11 b12 ··· b1N
transition. A combination of these techniques subsequently a · bN = [a1 , a2 , a3 ] b21 b22 ··· b2N
determined the time, t(PT), and duration of the postural transi- b31 b32 ··· b3N
tions, TD [17]. N
3
3
The wavelet toolbox of Matlab® was used to calculate the dif- = al bln · · · al blN (4)
ferent wavelet transforms used in this study. l=1 l=1 n=1
Table 2 posture transition (these duration times are presented in Fig. 6(a)
Step count for a chest mounted device determined by (a) DWT method (Najafi et al.)
and (b)).
applied to the vertical acceleration (avs ) and (b) VESPA derived step count from the
vertical estimate.
standing up (SiSt transition, Fig. 3(a)) and sitting down (StSi tran-
sition, Fig. 3(b)) from a kitchen chair for a young healthy subject 4. Discussion and conclusion
by the application of the DWT and methodology adopted by Najafi
et al. Previous activity classification studies have focused on the mul-
As discussed by Najafi et al. the nearest positive and negative tiple sensor arrangements but this often involved more complex
peaks of DWT(avs ) to the local minimum point of sin() were chosen algorithm development to determine the correct activity. More
to represent the transition of StSi and SiSt. In total there were 360 recent studies have incorporated single site sensor attachment,
PTs performed by the study subjects with a sensitivity and speci- however, those sensors incorporated more high-power devices,
ficity of 92% and 88% for StSi and 89% and 93% for SiSt respectively (such as gyroscopes) thereby limiting the monitoring duration
(these findings were similar to that achieved by Najafi et al.). capability and offline analysis of the study (due to complex algo-
Walking was detected by applying the DWT to avs between the rithms). Clinically, the application of body worn sensors without
scales of 2 and 5 using a Daubechies mother wavelet of order 4 (db4) the possibility of long term patient monitoring can be limiting due
[24]. A threshold was then applied to the negative peaks of the sig- to the added burden/cost placed upon patients and researchers to
nal, Fig. 4(a) [17]. Walking was correctly identified by this method replace sensors and/or batteries. While modern day sensors with a
with an accuracy of 100%. The total number of steps detected is single site of body attachment can record for up to seven or more
presented in Table 2. days, their activity detection algorithms can be minimal due to their
Posture transition durations (TD) were also calculated from need to save battery life. The aim of this pilot study was to show
sin() as a measure of time between the two successive peaks, P1 that a suitable low-power device with a more complete and uncom-
and P2 , around the time of posture transition (tPT ), Eq. (6) [25]: plicated activity classification algorithm was possible with a single
chest worn tri-axial accelerometer device.
TD = t(P1 ) − t(P2 ) (6) The single tri-axial accelerometer kinematic sensor and sim-
plified activity algorithm was compared directly with a proven
The posture transition duration times for both methods (Najafi
sensor configuration worn on the chest (incorporating a bi-axial
method and VESPA) are presented in Figs. 3 and 5.
accelerometer and uni-axial gyroscope) that utilised a more com-
plicated activity algorithm. Firstly, the proven activity detection
3.2. Stage 1 (YHS): VESPA technique was replicated successfully and was found to have suffi-
cient accuracy with the test group, young healthy adults (YHS) for
Lying was detected by the VESPA algorithm from the angle mea- the activities performed, Fig. 3. The activities of lying, standing, sit-
sured (trunk tilt, ) by the tri-axial accelerometer of the vertical ting, walking, and postural transitions (StSi and SiSt) were detected
accelerometer axis (avs ) with a sensitivity of 100%. Fig. 5(a) (SiSt) with success rates similar to those of Najafi et al. [17]. The new
and (b) (StSi) shows the resulting postural detection method, simi- sensor arrangement and simplified algorithm (classified as VESPA)
lar to that adopted by Najafi et al. but applied to the scalar product were then applied to the same activities as performed by the YHS.
method. However, in this study both types of transitions were dis- The data is presented in Figs. 4(b) and 5 for direct comparison to
tinguished using vve . This was achieved by examining the maximum Figs. 3 and 4(a) which show the potential for the new system con-
positive and negative peak values of the vve around the time of a figuration and activity algorithm. The activity of lying and walking
postural transition (t(PT)). A greater maximum positive peak value was detected with an accuracy of 100% under control conditions.
with a SiSt transition (Fig. 5(a, part iii)) and a greater maximum This is an improvement in walking detection over the method used
negative peak value was associated with a StSi transition (Fig. 5(b, by Najafi et el. where spurious peak detection may have resulted
part iii)). Of the 360 PTs performed by the young adults, sensitivity in misclassification (Fig. 4(a)). Importantly, the activities of walk-
and specificity were 92% and 86% for StSi and 86% and 92% for SiSt ing and step counting were more clearly identifiable using VESPA,
respectively for VESPA. Table 1 shows the overall sensitivity and with the elimination of spurious peaks (Table 1 and Fig. 4) when
specificity of transition detection for the 10 YHS. compared to the method adopted by Najafi et al. [17]. This new
Walking was detected with an accuracy of 100% by examining kinematic sensor and simplified activity algorithm technique was
the negative peaks of the vertical velocity estimate, Fig. 4(b). The then applied to data recorded from an elderly group in their own
step estimation is compared to the method adopted by Najafi et al. home environment. The same activities as performed by the EHS
in Table 2. Postural duration times were also calculated for VESPA were also identified successfully with a sensitivity and specificity
based on similar peak detection during the time of a detected of 89% and 83%, while SiSt had a sensitivity and specificity of 83%
A. Godfrey et al. / Medical Engineering & Physics 33 (2011) 1127–1135 1135
and 89% respectively. Posture duration transition times were also [6] Culhane KM, Lyons GM, Hilton D, Grace PA, Lyons D. Long-term mobility mon-
similar between the two systems for both age groups examined. itoring of older adults using accelerometers in a clinical environment. Clin
Rehabil 2004;18(3):335–43.
Limitations of the current study include a 1-s calibration before [7] Grant M, Dall P, Granat M. The feasibility of measuring activity patterns of
the performance of each activity. This required the participants women with osteoporosis. In: Procedings of the 4th world congress of the
to remain still for at least 1-s to provide the quantities required international society of physical medicine and rehabilitation. 2007.
[8] Ryan C, Gray H, Newton M, Granat MH. An investigation of the effects of psy-
by the calibration technique [4]. There is also a lack of continu- chological distress on physical activity levels in individuals with chronic lower
ous, long-term ADL for both groups and accurate recordings for back pain. In: Procedings of the 4th world congress of the international society
at home testing. Future work based upon the techniques devel- of physical medicine and rehabilitation. 2007.
[9] Leonard M, Godfrey A, Silberhorn M, Conroy M, Donnelly S, Meagher D, et al.
oped and presented here will involve the recruitment of a larger
Motion analysis in delirium: a novel method of clarifying motoric subtypes.
cohort within these groups, over an extended recording period Neurocase 2007;13(4):272–7.
(7–14 days) to test the algorithm under continuous ambulatory [10] Culhane KM, O’Connor M, Lyons D, Lyons GM. Accelerometers in rehabilitation
medicine for older adults. Age Ageing 2005;34(6):556–60.
monitoring.
[11] Coley B, Najafi B, Paraschiv-Ionescu A, Aminian K. Stair climbing detec-
The proposed low-power (<1.2 mA) sensor (single tri-axial tion during daily physical activity using a miniature gyroscope. Gait Posture
accelerometer) and new activity algorithm (VESPA) while still at 2005;22(4):287–94.
an early stage of development has a number of advantages over [12] Nyan MN, Tay FE, Tan AW, Seah KH. Distinguishing fall activities from normal
activities by angular rate characteristics and high-speed camera characteriza-
previous methods: tion. Med Eng Phys 2006;28(8):842–9.
[13] Kangas M, Konttila A, Lindgren P, Winblad I, Jamsa T. Comparison of low-
(a) Simple and easy to use activity algorithm and as a result the complexity fall detection algorithms for body attached accelerometers. Gait
Posture 2008;28(2):285–91.
new VESPA algorithm is less processor intensive. [14] Karantonis DM, Narayanan MR, Mathie M, Lovell NH, Celler BG. Implementation
(b) Use of a low-power body worn sensor (single tri-axial of a real-time human movement classifier using a triaxial accelerometer for
accelerometer on the chest) to improve long term ambulatory ambulatory monitoring. IEEE Trans Inf Technol Biomed 2006;10(1):156–67.
[15] Lyons GM, Culhane KM, Hilton D, Grace PA, Lyons D. A description of
monitoring. an accelerometer-based mobility monitoring technique. Med Eng Phys
2005;27(6):497–504.
CAALYX (Complete Ambient Assisted Living Experiment) is an [16] Godfrey A, Conway R, Meagher D, Ólaighin GM. Direct measurement of human
movement by accelerometry. Med Eng Phys 2008;30(10):1364–86.
Integrated Project supported by the European Community under
[17] Najafi B, Aminian K, Paraschiv-Ionescu A, Loew F, Bula CJ, Robert P.
the Sixth Framework Programme (IST-2006-045215). Ambulatory system for human motion analysis using a kinematic sensor:
monitoring of daily physical activity in the elderly. IEEE Trans Biomed Eng
2003;50(6):711–23.
Conflict of interest
[18] Ferraris F, Grimaldi U, Parvis M. Procedure for effortless in-field calibra-
tion of three-axis rate gyros and accelerometers. Sens Mater 1995;7(5):
The authors wish to acknowledge the assistance of the 311–30.
[19] Lötters J, Schipper J, Veltink PH, Olthuis W, Bergveld P. Procedure for in-use
volunteers, Analog Devices, BV for providing the ADXL210
calibration of triaxial accelerometers in medical applications. Sens Actuators
accelerometers and ADXRS300 gyroscopes and the CAALYX FP6 A: Phys 1998;68:221–8.
project caalyx.eu [26], for their financial support. [20] Semmlow J. Biosignal and biomedical image processing: matlab based appli-
cations. New York: Marcel Dekker; 2004.
[21] Bruce A, Donoho D, Gao H-Y. Wavelet analysis. IEEE Spectr 1996;33(10):26–35.
References [22] Degen T, Jaeckel H, Rufer M, Wyss S. SPEEDY: a fall detector in a wrist watch. In:
Proceedings of the seventh IEEE international symposium on wearable com-
[1] Coleman KJ, Saelens BE, Wiedrich-Smith MD, Finn JD, Epstein LH. Relationships puters, 2003. 2003.
between TriTrac-R3D vectors, heart rate, and self-report in obese children. Med [23] Bourke AK, O’Donovan KJ, Nelson J, Ólaighin GM. Fall-detection through vertical
Sci Sports Exerc 1997;29(11):1535–42. velocity thresholding using a tri-axial accelerometer characterized using an
[2] Clarke-Moloney M, Godfrey A, O’Connor V, Meagher H, Burke PE, Kavanagh EG, optical motion-capture system. Conf Proc IEEE Eng Med Biol Soc 2008:2832–5.
et al. Mobility in patients with venous leg ulceration. Eur J Vasc Endovasc Surg [24] Daubechies I. Ten lectures on wavelets (CBMS-NSF regional conference series in
2007;33(4):488–93. applied mathematics). Philadelphia, PA, USA: Society for Industrial and Applied
[3] Bourke AK, O’Brien JV, Lyons GM. Evaluation of a threshold-based tri-axial Mathematics; 1992.
accelerometer fall detection algorithm. Gait Posture 2007;26(2):194–9. [25] Najafi B, Aminian K, Loew F, Blanc Y, Robert PA. Measurement of stand-sit and
[4] Bourke AK, O’Donovan KJ, Ólaighin GM. The identification of vertical velocity sit-stand transitions using a miniature gyroscope and its application in fall risk
profiles using an inertial sensor to investigate pre-impact detection of falls. evaluation in the elderly. IEEE Trans Biomed Eng 2002;49(8):843–51.
Med Eng Phys 2008;30(7):937–46. [26] Boulos MNK, Rocha A, Martins A, Vicente ME, Bolz A, Feld R, et al. CAALYX:
[5] Moore ST, MacDougall HG, Gracies JM, Cohen HS, Ondo WG. Long-term moni- a new generation of location-based services in healthcare. Int J Health Geogr
toring of gait in Parkinson’s disease. Gait Posture 2007;26(2):200–7. 2007;6(9):1–6.