Automated Recognition of Human Movement States

using Body Acceleration Signals

Md. Rafiul Hassan

, Rezaul K. Begg

, Ahsan H. Khandoker

and Robert Stokes

Department of Computer Science and Software Engineering

The Unversity of Melbourne, VIC 3010, Australia.

CARES, Victoria University, Melbourne 8001, Australia.

Department of Electrical Engineering

The University of Melbourne, VIC 3010, Australia.

Abstract. Automated recognition of human activity states has many advan-

tages, e.g., applications in the smart home environment for the monitoring of

physical activity levels, detection of accidental falls in the older adults in the

home environment or assessment of the recovery phase of patients living inde-

pendently at home. In this paper, we describe an accelerometer-based system to

recognize three activity states, e.g., steady state gait or walking, sitting and

simulated sudden accidental falls. The recorded 3D movement accelerations of

the trunk were processed using wavelets, and the features were extracted for

recognition of movement states through the use of a fuzzy inference system.

The system was trained and tested using 58 different data segments represent-

ing the three states. Cross-validation test results indicated an overall recognition

accuracy by the machine classifier to be 89.7% with an ROC area of 0.83. The

results suggest good potential for the system to be applied for various situations

involving activity monitoring as well as gait and posture recognition. Further

tests are required using various population groups.

1 Introduction

Falls in older people during locomotion is a major public health issue, as it leads to

injuries, hospitalization and a significant cost to the community. In Australia, falls

related injury costs have been estimated to be A$2.4 billions per annum [1]. Re-

searchers around the globe are investigating this subject using various techniques.

One methodology involves gait analysis in a laboratory environment to identify the

significant changes in gait that occur due to ageing. Some of these changes may be

important in the understanding of why people fall and this information may also help

us to devise techniques for the prevention of falls. Automated monitoring of gait and

posture of an individual within the home environment has recently received signifi-

cant attention. The main advantage of such monitoring is to identify people at risk of

falls and also to assess their overall well-being. For example, such technique could be

used for checking the healthy life style of patients recovering at home or to monitor

the health risk of an older person living independently at home. In order for a system

to be effective in such applications, the system has to be robust enough to differenti-

Raﬁul Hassan M., K. Begg R., H. Khandoker A. and Stokes R. (2006).

Automated Recognition of Human Movement States using Body Acceleration Signals.

In Proceedings of the 2nd International Workshop on Biosignal Processing and Classiﬁcation, pages 135-143

DOI: 10.5220/0001225601350143

 SciTePress

ate between the various events and activities such as walking, sitting, resting, or acci-

dental falls. In this paper, we attempt to develop an automated system that could be

applied to distinguish such sequences of events.

In the past, a number of studies have used accelerometer technology to automate the

monitoring of daily activities of humans such as walking, sitting, standing, climbing

stairs, walking velocity, etc, classified and recognized by analyzing the various fre-

quency components of the acceleration signals [2-3]. To acquire the useful knowl-

edge from the raw dataset, wavelet analysis was extensively used in some of studies

[4-6].

To automate recognition of activities from either pre-processed raw data, various

techniques and classifiers have been proposed in several studies, such as Fractal esti-

mation [4], Neural Networks [3,7]. However, these techniques have their own merits

and limitations, which depend on the dataset. In our previous work [8], we have suc-

cessfully applied fuzzy rules to recognize gait changes due to trip-related falls in a

laboratory environment. In this study, we apply fuzzy classifier to recognize the daily

activities in a home environment.

2 Methods

In this study, data were collected from a simulated home environment. A triaxial

accelerometer unit (giving acceleration information in horizontal, vertical and lateral

positions) was placed on a subject’s body to collect movement data. Acceleration data

were pre-processed using wavelet technique and classified using a set of fuzzy rules.

2.1 Description of the System

A schematic diagram of the experimental setup is shown in Figure 1. This measure-

ment system consisted of a Crossbow ±4G triaxial accelerometer unit, a National

Instruments 6013 16-bit PCI analog to digital converter interface card, a Labview 7.0

data acquisition program [http://www.ni.com/labview/] and a personal computer. The

accelerometer unit was attached to the upper body (trunk) of the subject.

Fig. 1. Activity recognition of individuals in a smart home.

Tri-axial Accelerometer

Breakout

Box

Labview

Data

Acquisition

Feature

Extraction

Fuzzy

Classifier

Frequency

Threshold

Activity

Reco

nition

NI 6013

Data

Acquisition Card

136

A breakout board was used to connect the accelerometer with the analog to digital

converter card such that data could be acquired by the computer. The physical data

was digitized using the 16-bit analogue-to-digital (A/D) converter and the LabView

program captured and recorded the data signal onto the PC harddrive. In this study,

data were collected and digitized at 100 Hz (100 bits/sec) sampling rate.

The experiments were set up in a room of size 3.5m by 3.5m Various items of furni-

ture (chairs, tables, etc) were arranged to mimic an actual home environment. For the

purpose of data collection, each subject entered into the room through the only door

of the room and walked to the chair, sat on the chair, typed something using the PC

keyboard, sometimes suddenly fell (simulated falling) onto the ground and conse-

quently stood up. The experiments were performed on healthy subjects with no

known gait problems.

2.2 Feature Extraction

2.2.1 Wavelets and Transformations

Wavelet functions as defined by [15], are constructed from translations and dilations

of a single function called the “mother wavelet”

)(t

)(

⎟

⎠

⎞

⎜

⎝

⎛

−

0,,

≠

ℜ

∈

aba

(1)

where, the scaling parameter ‘a’ represents the degree of compression (its value is

linked to the frequency of the wavelet), and the translation parameter ‘b’ represents

time location of the wavelet. Higher wavelet frequencies are indicated using ‘a’ <1,

whereas lower frequencies are represented when ‘a’>1.

A discrete wavelet transformation of a signal

)(xf can be represented using the

following equations

∫

= dxxxfka

)()()(

(2)

∫

= dxxxfkd

)()()(

(3)

∑∑∑

kjkJ

xkdxkaxf

)()()()()(

ψφ

(4)

In equations (2),(3) and (4), ‘j’ stands for the dilation index, ‘k’ represents the index

in time, ‘J’ is the depth of the decomposition level and ‘*’ symbolizes complex con-

jugation [4]. The scaling function coefficient

gives the average value of the signal

over the given interval, which is used to calculate the approximation coefficients. The

details coefficients obtained using equation 3 keeps the high frequency information

while the approximation coefficients in each level contain the low frequency informa-

tion.

137

The details coefficients were used in this study to obtain the frequency of a signal in

a specific time period. To detect the frequency of a function representing a fixed

length signal (length of signal was selected to be 1 second in this research), a wavelet

decomposition followed by a translation of scale to frequency was computed. First,

the signal was decomposed using wavelets of up to ‘n’ labels and then the energy of

the details coefficients for each of the levels was obtained. The scale value (

a )

related to the maximum energy from the set of energy values was then taken into

consideration to dig out the original frequency of the signal. Next step was to find out

the frequency of a wavelet for the corresponding scale.

2.2.2 Translation from Scale to Frequency

As the frequencies in a signal predominate the scales chosen in wavelet decomposi-

tion, by analyzing the relationships between the scale and frequency the pseudo-

frequency of the signal can be obtained. Following the methodology outlined in [16],

the pseudo-frequency of a wavelet corresponding to a scale was computed. In doing

so, first the center frequency F

of the wavelet was computed by the following equa-

tion [10]:

where,

a = scale, Δ = sampling period,

F = center frequency of the wavelet in Hz.

F = pseudo-frequency corresponding to the scale a in Hz.

(5)

Using this method, the frequency calculated at scale ‘

a ’ became the pseudo-

frequency of the signal. The next step in the recognition system was the inclusion of

the fuzzy classifier for the identification of different activities (e.g., walking, sitting or

simulated falls) of the subjects using the calculated frequency information of the

acquired signals.

2.3 Fuzzy Classifier

Fuzzy logic provides a framework for modeling uncertainty, human way of thinking,

reasoning and the perception process, and has been found to be very useful in ex-

pressing the nonlinear relationships among the inputs and outputs [11-13]. Condi-

tional statements such as ‘if-then’ are the linguistic representation of fuzzy rules.

Recently, we have applied fuzzy rules to classify gait patterns of healthy adults and

adults with balance problems and the gait classification accuracy was found to be

over 89% [8]. Here, we propose to apply fuzzy rules to classify movement states

using features translated into the various wavelet coefficients.

With our present study, we don’t have any prior knowledge about the rules relating

to the subject’s movement data. So defining rules for these dataset is a problem. How-

ever, there are techniques to identify a reasonable number of rules from a given data-

138

set in order to establish a fuzzy system. For example, Chiu [14] has developed a tech-

nique based on subtractive clustering to extract rules. This technique has been tested

in gait pattern classification for rule generation [8], and was also applied in this re-

search .

2.4 Simulated falls event detection

In order to detect the occurrence of sudden/accidental falls, a signal of a fixed time

interval (1 second) was considered. During such events, the movement frequency is

expected to be momentarily high. To identify whether there was any instantaneous

increase in frequency, the details coefficient

coeff

values for the signal were ana-

lyzed. As the wavelet transformation provides both the frequency and time informa-

tion, any sudden increase in frequency can be monitored by studying the details coef-

ficients.

Fig. 2. The details coefficient values at label 3 to detect simulated falls.

Figure 2 shows a plot of the details coefficient and illustrates an abrupt increase in its

magnitude during the falls event (~200ms). To automate the system’s falls identifica-

tion capability, a threshold

was selected and the following threshold logic was

applied:

for i=1 to length(signal)

coeff

D >

then

identify sudden fall

end if

end for

In this study, details coefficient at level 3 (D3) was found to be the most suitable for

the problem, especially in the vertical direction.

Falls

139

3 Results

We tested the proposed wavelet-fuzzy method using data related to the 60 state

changes (change from one state to another state). As mentioned above, three types of

body states were considered: sitting, walking and a sudden simulated fall on the

ground). A change in state might take place according to one of the following se-

quences:

Walking, Sitting, Walking….

Walking, Sudden fall, Standing up and Walking again….

Sitting, Walking…

Figure 3(a-c) shows typical raw acceleration signal recorded during movement repre-

senting the various states in the horizontal anterior-posterior, medio-lateral and verti-

cal axes respectively. Figure 4 plots the computed detailed coefficients using equation

(3) for a mother coifflet wavelet (at levels 1 to 9) of the acceleration signal displayed

in figure 3(a).

For this data, ‘coifflet’ wavelet was found to be suitable for calculating the pseudo-

frequency of the acceleration signals as well as to identify the instantaneous increase

in the frequency. During walking task, the major frequency component was close to

4.0 Hz while that for the sitting posture on a chair or on the ground was found to be

~0.14 Hz. There were slight variations in the frequencies along the three axes as can

be seen in Figure 5. Therefore, information from more than one source is necessary

for correct recognition, and in this case, all three acceleration signals from the tri-

axial accelerometer were used for training and testing the fuzzy classifier.

Fig. 3. The collected signal using tri-axial accelerometer a) Horizontal anterior-posterior (a

b) Medio-lateral (a

) and c) vertical(a

) signals.

Table 1. Classification accuracy of the model (sitting and walking).

Total Activity

(Training & Test-

ing)

Accurate classifi-

cation

Classification ac-

curacy

ROC area

58 52 89.7% 0.83

Walk Falls Walk Sitting Walk Sitting Walk Sitting

140

Table 2. Sensitivity and specificity of the model.

Activity Name Total activity Sensitivity

Specificity

Walking 26

Sitting 22

0.917 0.833

Simulated Falls 10 All the fall events were recognized correctly with 100% accu-

racy

Fig. 4. The details coefficients (d1 to d9) at levels 1-9 of the signal in figure 3(a).

Fig. 5. The frequencies of segmented signal (into 1 second each) of the raw signals presented

in figure 3.

Classifier’s performance was tested using a four-fold cross validation test. Table 1

illustrates the overall result suggesting an overall accuracy of 89.7% in differentiating

141

walking and sitting. This accuracy is also reflected in the ROC curves that recorded a

high area (0.83) (see Figure 6). However, the recognition accuracy for all the simu-

lated falls was found to be 100%.

ROC Curve for movement recognition

Specificity

1.00.75.50.250.00

Sensitivity

1.00

.75

.50

.25

0.00

Fig. 6. ROC curve for identifying the two states: sitting and walking.

4 Discussions and Conclusion

Human movement behavior and the associated body positions change continually in

reaction to the demands of the environment in which we move. Accordingly, the

signal collected for a specific action of the subject was not the same all the time. It

might be a difficult task to recognize the subject’s current activity with only the ap-

plication of simple thresholds. Classifying different activities with the application of

sensor data on the same ground level (where the value of ’g’ is constant) makes the

problem more challenging. Most of the previous tasks related to activity identification

and classification dealt with movements in different levels (e.g., walking downstairs

i.e., ‘moving with g’ or walking upstairs i.e., ‘moving against g’) [4- 6]. In this study,

as the value of ‘g’ remained fixed throughout the experiment, movement task differ-

ences were reflected only in the acceleration frequencies (see figure 5).

It was observed that signals collected from only one axis could identify some of the

activities at times but not always. With the use of 3D accelerations the system was

found to identify and recognize the subject’s activities more robustly. Moreover,

being random in characteristics, fuzzy systems were found particularly suitable for

analyzing the in-depth information in the recorded signals. However, for detecting

accidental falls event, a threshold level detector might be suitable due to its sudden

increase in frequency level.

142

The benefits of the proposed model are: (1) the accelerometer can be attached to the

garments without the complicated setup procedures, (2) activities can be identified on

the same floor/level, (3) any accidental falls can be accurately identified.

This specific model could be applied in a smart home to monitor and discriminate

any discrepancy in subject’s behavior. In particular, this model could potentially help

older adults who are living independently in their homes. The usage of garments for

recording accelerometer signals, allows a non-intrusive monitoring technique of the

subject’s gait activities.

References

1. Fildes B. 1994. Injuries among older people. Melbourne: Collins Dove.

2. Fahrenberg, J., Foerster, F., Smeja, M., Muller, W. (1997), ”Assessment of posture and

motion by multichannel piezoresistive accelerometer recordings” , Psychophysiology. Vol :

34(5), pp. 607-612.

3. Mantyjarvi, J., Himberg, J., Seppanen, T. (2001), “Recognizing human motion with multi-

ple acceleration sensors”, Proceedings of IEEE international conference on Systems, Man

and Cybernetics, pp. 747-752.

4. Sekine, M., Tamura, T., Akay, M., Fujimoto, T., Togawa, T., and Fukui, Y.

(2002),”Discrimination of Walking Patterns Using Wavelet-Based Fractal Analysis”, IEEE

Transactions on Neural Systems and Rehabilitation Engineering, Vol : 10 (3), pp. 188-196.

5. Cho, S. Y., Park, C.G., and Jee, G.I., (2002),”Measurement system of walking distance

using low-cost accelerometers”, Proceedings of the 4

Asian Control Conference, pp.

1799-1803.

6. M. N. Nyan, Tay, F.E.H., Seah, K.H.W., Sitoh, Y.Y. (2005),”Classification of gait patterns

in the time-frequency domain”, Journal of Biomechanics, (In press).

7. Aminian, K., Robert, P., Jequier, E., and Schutz, Y. (1995), ”Estimation of Speed and

incline of walking using neural network”, IEEE Transaction on Instrumentation and Meas-

urement, Vol : 44(3), pp 743-746 .

8. Hassan, M.R., Begg, R., and Taylor, S. (2005), “Fuzzy Logic-based recognition of gait

changes due to trip-related falls”, Proceedings of the 2005 IEEE Annual Conference on

Engineering in Medicine and Biology(EMBS’05).

9. Debnath, L. (2001), Wavelet Transforms & Their Applications, Birkhäuser , Boston ,USA.

10. Misiti, M., Misiti, Y., Oppenheim, G., Poggi, J-M. (2005), Wavelet Toolbox User’s Guide-

For use with Matlab, The Mathworks Inc.

11. Meunier, B. B, Yager, R. R., and Zadeh, L. A. (Eds)(2000), Uncertainty in Intelligent and

Information Systems, World Scientific Publishing Company, Singapore.

12. Pedrycz, W. (1995) , Fuzzy Sets Engineering, CRC Press.

13. Zadeh, L. A. (1965), Fuzzy Sets, Information and Control, Vol : 8, pp. 338-353.

14. Chiu. S.L.(1997), Extracting Fuzzy Rules from Data for Function Approximation and

Pattern Classification, Chapter 9 in Fuzzy Information Engineering: A guided Tour of Ap-

plications (eds) D. Dubois, H. Prade, and R. Yager, John Wiley & Sons.

15. Morlet, J. , Aerens, G., Fourgeau, E., ans Giard, D. (1982) ,”Wave propagation and sam-

pling theory, Part I: Complex signal land scattering in multilayer media”, Journal of Geo-

physics, Vol : 47, pp. 203-221.

16. Abry, P. (1997), Ondelettes et turbulence. Multirésolutions,algorithmes de décomposition,

invariance d’échelles, Diderot Editeur, Paris.

143