Fall Detection Algorithm Compiled

32
Introduction According to World Health Organization (2012), fall is defined as an event that a person coming to rest inadvertently to a lower ground level. Fall is considered as a second leading cause of accidental injury death worldwide. The yearly estimated fatal fall is 424,000. The group of people who have the greatest falling risk is adults whom are older than 65 years old. Besides aging, there are several other risk factors for falling: occupation at elevated heights, side effects of medication that will result in a loss of balance, poor mobility, cognition and vision and underlying medical condition for example Alzheimer's disease that results in three times of falling risk. ( Abbate, S., Avvenuti, M., Corsini, P., Light, J., & Vecchio, A., 2010)

description

Full report of a Capstone, "Fall Detection" come with MATLab coding.

Transcript of Fall Detection Algorithm Compiled

Page 1: Fall Detection Algorithm Compiled

Introduction

According to World Health Organization (2012), fall is defined as an event that a

person coming to rest inadvertently to a lower ground level. Fall is considered as a

second leading cause of accidental injury death worldwide. The yearly estimated fatal

fall is 424,000. The group of people who have the greatest falling risk is adults whom

are older than 65 years old. Besides aging, there are several other risk factors for

falling: occupation at elevated heights, side effects of medication that will result in a

loss of balance, poor mobility, cognition and vision and underlying medical condition

for example Alzheimer's disease that results in three times of falling risk. (Abbate, S.,

Avvenuti, M., Corsini, P., Light, J., & Vecchio, A., 2010)

Fall can occur along one of the two anatomical planes, sagittal plane shown in

Figure 1 (a) and coronal plane in Figure 1 (b). The fall that occurs along the sagittal

plane is either forward or backward whereas the fall that occurs along the coronal

plane is either left or right. Besides, fall can occur in any circumstances. It is not only

occurring when the person is standing but it happens when the person is sitting on

chair or is lying on the bed.

Figure 1: Falling direction

Falling events can be monitored by using a network of sensors including pressure

sensor on chairs, camera, Radio Frequency Identification (RFID) tags embedded

throughout the home of elderly people or elderly care unit and the clothes, which

Page 2: Fall Detection Algorithm Compiled

communicate with the tag readers in walls, floor mats and shelves. The main

advantage of having this kind of monitoring system is that the person needs not to

wear the sensors on his body. However, it is only useful if the person at risk is

confined to a place but not for outdoor activities. The effect of falls can be bone

fractures or sometimes leading to fatal due to 'long lie'. 'Long lie' refers to

involuntarily remaining on ground for an hour or more than that. (Wild D, Nayak,

1981) According to research done by Wild D Nayak (1981), more than half of the

elderly who experienced 'long lie' died within 6 months even though there is no direct

injury after the fall. Therefore a real time monitoring of fall is needed to reduce the

occurrence of 'long lie'. Monitoring the activities of daily living (ADL) requires a

wireless sensor to analyze and capture body movements, triggers the alarm when a

fall is detected. The wireless sensor is light, comfortable and easy to carry. Besides, it

does not intrude the privacy of the person and is not confined to a certain environment

only.

Although commercial devices on fall detection exist, they do not fulfilled

satisfaction due to high false alarm rate, high initial and maintenance cost, and their

non-ergonomic nature. (Noury, 2008) In order to improve the robustness of fall

detection sensor, a reliable fall detection algorithm is needed to reduce the false alarm

rate. To achieve this, we need sufficient data for algorithm development. Since it takes

a longer time to get the data from involuntary fall, we limit our study to voluntary

(simulated) falls.

Many different approaches have been investigated to detect fall. It can be using

accelerometer only or inertial based approach which is the combination of

accelerometer and gyroscopes. The advantage of the combination of accelerometer

and gyroscope is that the specific event, either voluntary (ADLs) or involuntary (fall)

can be determined clearly based on statistical or threshold based algorithm.

Objectives

1. To design an experiment to simulate a fall event.

Page 3: Fall Detection Algorithm Compiled

2. To develop a fall detection algorithm.

Literature Review

There are two types of fall detection algorithms. There is bi-axial and tri-axial.

TRI-AXIAL ACELERAOMETER ALGORITHMS

According to NingJia on hispaper Detection of human fall, he used the tri-axial

accelerometer (ADLX 345) to measure the fall. He said that, combination of the

parameter on the inertial sensor form the entire fall detection algorithms will detect

the fall event. Healso said that, when the fall is occurred, it can cause the sensor

system to raise an appropriate alert that a fall has occurred. In this paper, NingJia

main research on the principles of fall detection focuses on the changes in

acceleration that occur when a human is falling.

Page 4: Fall Detection Algorithm Compiled

Figure 2: Block diagram for the flow chart of the fall detection algorithm based on tri-axial sensor (Jia, 2009).

According to him, the data is continuously analyzed algorithmically to determine whether the individual’s body is falling or not. If an individual falls, the device can employ GPS and a wireless transmitter to determine the location and issue an alert in order to get assistance. The core element of fall detection is an effective, reliable detection principle and algorithm to judge the existence of an emergency fall situation.

BI-AXIAL GYROSCOPE ALGORITHMS

Another type of fall detection algorithms is from paper A.K Bourke and G.M

Page 5: Fall Detection Algorithm Compiled

Lyons. They use the bi-axial gyroscope sensor (ADXRS 300) to detect the fall detection. They said that, to distinguish the Activity of Daily living (ADL) and fall, the bi-axial sensoris used that place on the trunk to measuring the pitch and roll angular velocities and a threshold-based algorithm.

Figure 3: block diagram of flow chart of fall detection on the bi-axial sensor (Bourke, Lyons 2008).

As from the figure 3, the sensor is measuring two parameters, the angular velocity of pitch (ωp) and the angular velocity of roll (ωr). From this two data, the resultant force for this velocity signals (ωres) was derived by taking a root mean square of these two parameters. Thus, this would provide a combination measurement of angular velocity in sagittal and frontal planes.

By setting a threshold which is referred to the Fall Threshold 1 (FT1) Fall Threshold 2 (FT2) and Fall Threshold 3 (FT3), the falls recorded in this study will be correctly identified as falls, by cascading these thresholds, and that no ADL events would be misdetected as falls.

Page 6: Fall Detection Algorithm Compiled

Figure 4: The analysis data of the ADL [2]

(a) An ADL that exceeded FT1, FT3 but not FT2 is an ADL(b) An ADL that exceeded FT1 and FT2 but did not exceeded FT3 is a Fall(c) The signals from a typical fall, where all three thresholds are exceeded.

Methodology

A. Pre-experiment

1. The whole procedure of the experiment are explains to the subjects.

2. Subjects are required to sign on concern form if agree with the experiment

Page 7: Fall Detection Algorithm Compiled

progress.

3. Subjects must be healthy where there will be no implants in the body and do

not perform any surgery for the past 6 months at least.

B. During-experiment

1. Subjects are required to remove all the accessories such as watch, necklace,

rings, bracelet, spectacles and others.

2. Subjects will have the rights in stopping the experiment anytime when he/she

not feeling well.

3. Mattresses will be placing at the falling point of the subjects from the chair.

4. The falling device will be switched on and calibrated and connected to the

laptop using WiFi.

5. The falling device will be put onto the waist of the subject using the belt of

the device.

6. Subject will be request to walk around for 1 minute and then sit on the chair.

7. Subject will sit on chair for few seconds for self preparation before falling

down from the chair to the mattresses.

8. Subject will be falling to their right hand side only throughout the whole

experiment.

9. The chair will be hold in preventing the chair falling along with subjects and

cause injury.

10. Subject will have to stay on the mattresses for few seconds.

11. Subject will then repeat step 6 to 10 for two more times.

12. If subject encounter any injury, the experiment will be halted. Subject’s

condition will be checked and will be rush to hospital if needed.

13. The whole process will be repeated until all 10 subject’s data are collected.

14. The recorded data will be analyzed and shown in the next chapter of this

report.

C. Post-experiment

Page 8: Fall Detection Algorithm Compiled

1. Simple body check up will be done on subjects and if there is any injury of

feeling not well, subjects will be rush to nearby hospital for a proper check

up and get medication if needed.

Results

Table 1: BMI for each subject

Subject Gender Height (m)

Weight (kg)

BMI Status

1 M 1.52 52 22.51 Normal2 M 1.67 65 23.31 Normal3 M 1.71 63 21.55 Normal4 F 1.56 54 22.19 Normal5 F 1.42 48 23.80 Normal6 F 1.52 45 19.48 Normal7 F 1.52 56 24.24 Normal8 M 1.67 70 25.10 Overweight9 M 1.68 60 21.25 Normal10 F 1.50 38 16.89 Underweight

Table 2: Peak value of acceleration and time during falling event for each subject

Subject Peak value of acceleration (g) Time of falling (s)1 1.4420 63.922 1.0586 64.483 1.4299 68.184 0.7439 68.255 0.9078 59.946 0.6067 65.647 1.4190 83.828 1.2348 63.229 1.6165 63.4910 1.3412 65.47

Graphs of the falling event for all subjects are as follow:

Subject 1

Page 9: Fall Detection Algorithm Compiled

Subject 2

Subject 3

Page 10: Fall Detection Algorithm Compiled

Subject 4

Page 11: Fall Detection Algorithm Compiled

Subject 5

Subject 6

Page 12: Fall Detection Algorithm Compiled

Subject 7

Subject 8

Page 13: Fall Detection Algorithm Compiled

Subject 9

Subject 10

Page 14: Fall Detection Algorithm Compiled

Discussion

This experiment involves 5 male and 5 female subjects whereby subject 8 is

overweight, subject 10 is underweight and the other subjects are within the normal

range. The subjects are asked to carry out activity of daily living (ADL) which is

walking for one minute before intentional falling from the chair. A threshold-based

tri-axial accelerometer fall detection algorithm is used to differentiate between ADL

and falling. It is an automatic approach of detecting fall event from the acceleration

value when exceeds a threshold of 80%of the peak value as shown on Figure 5. A

fixed acceleration value cannot be set as threshold because each person has different

threshold value due to different BMI.

Page 15: Fall Detection Algorithm Compiled

Figure 5:Acceleration pattern of subject 3

Tri-axial accelerometer records the acceleration change due to force in x, y and z

axes of direction. The magnitude of vector sum can be computed by taking the root-

sum-of-square of acceleration from these three different directions. The peak value

indicates the peak impact force experienced by the body, which is different from the

value produced during the performance of normal ADL.Data analysis is performed

using MATLAB to determine the peak acceleration and falling time.

SV=√ ( Ax )2+ ( A y )2+( A z )2-------------------------------------------------- Equation 1

In order to record the acceleration pattern resultant from these activities, the

accelerometer is mounted onto the waist. The waist-attached accelerometer is able to

provide reliable information on the movement and posture of subjects because it is at

the nearest location to the body’s center of location, with the exception of movements

of the legs and arms (Kangas, 2011). Most of the fall detection applications place

accelerometer at waist since the fall detection sensitivity ranges from 70% to 100%

and the specificity from 95% to 100% (Kangas, 2011).

Page 16: Fall Detection Algorithm Compiled

Based on the results, falling event is detected for each subject by the elevated

acceleration level at around 70s ± 10s. The falling subject initially accelerates as

being pulled by the downward force of gravity. According to Newton’s law, the

amount of acceleration is directly proportional to the net force acted upon the subject.

It is supposed that when the subject has larger BMI (larger mass), he will experience

greater downward force of gravity with acceleration. However, based on the results,

subject 1, 3, 7 and 9 who are within the normal range of BMI depicts greater amount

of acceleration than subject 8 who is overweight. While subject 10 who is

underweight supposed to show the lowest peak acceleration, but it does not show in

the experiment. So, this condition may be affected by the varying sensitivity and

specificity of the device while recording the acceleration pattern.

The falls occurred during the experiment come intentionally and being

constrained. The subjects are instructed to fall onto the mattresses. It is opposed to

real life fall onto the hard surfaces and with fully attempt to break a fall. Therefore,

the peak acceleration from the simulated falls will be expected to be much lower than

that of real conditions. For further research, the fall detection experiment can be

conducted on crash test dummy to determine peak acceleration differences between

soft mattresses and typical domestic surfaces. Another limitation of this experiment is

not determining the falling posture of the subjects using gyroscope. It is crucial to

know as different fall postures will injure specific body part. Moreover, more subjects

of different ages should be involved in order to confirm the workability of the fall

detection algorithm.

Conclusion

As conclusion, accelerometer is able to detect falling event although working

alone. But the specificity and sensitivity will be limited. If combining with gyroscope,

the sensitivity and specificity will be much higher. In this experiment, subjects are

intentionally fall and there will be mattresses at the falling area. Thus, the acceleration

Page 17: Fall Detection Algorithm Compiled

during the falling event will be lower than the “full falling” event where it will be not

intentionally and direct fall onto the ground. Besides, the relationship between BMI

and acceleration during fall event is not clear yet. Suggestions to this unsolved matter

will be collect and analyze larger data pool and improve the falling algorithm.

Nevertheless, objectives of this experiment are fulfilled where the fall events are

simulated successfully and fall detection algorithm had been developed.

References

Abbate, S., Avvenuti, M., Corsini, P., Light, J., & Vecchio, A. (2010). Monitoring of

human movements for fall detection and activities recognition in elderly care

using Wireless sensor network: a survey.

Bagalà, F., Becker, C., Cappello, A., Chiari, L., Aminian, K., Hausdorff, J. M., ... &

Klenk, J. (2012). Evaluation of accelerometer-based fall detection algorithms

on real-world falls. PloS one, 7(5), e37062.

Bourke, A. K., & Lyons, G. M. (2008). A threshold-based fall-detection algorithm

using a bi-axial gyroscope sensor. Medical engineering & physics,30(1), 84-

90.

Jia, N. (2009). Detecting Human Fall With a 3-Axis Digital Accelerometer. Analog

Dialogue, 43, 1-7.

Kangas, M. (2011).Development of accelerometry-based fall detection.

Özdemir, A. T., & Barshan, B. (2014). Detecting falls with wearable sensors using

machine learning techniques. Sensors, 14(6), 10691-10708.

World Health Organization. Ageing, & Life Course Unit. (2008). WHO global report

on falls prevention in older age. World Health Organization.

Appendices

Fall detection algorithm

Page 18: Fall Detection Algorithm Compiled

% Coding for detecting fall event and the time it occurred% read data from deviceraw_data=xlsread('dataset_105_export.csv'); %put the name of the data fileac1=raw_data(:,2); % "2" is refer to second column, x-axisac2=raw_data(:,3); % "3" is refer to second column, y-axisac3=raw_data(:,4); % "4" is refer to second column, z-axiss=(((ac1.^2)+(ac2.^2)+(ac3.^2)).^(1/2))/10000;

len=length(raw_data(:,1)); % finding total number of datadata_plot=[1:len]'; time_plot=data_plot/100; % make X-axis in terms of second, assuming sampling rate is 100Hz% detecting thresholdthreshold=0.8*max(s'); % setting threshold as 80% of the peak value[peak,tpc_dataplot]=findpeaks(s,'MinPeakHeight',threshold); % finding the peak valuetime_fall_occured=tpc_dataplot/100; % finding the time (sec) for peak value% graph plottingplot(time_plot,s); % graph of time vs amplitudeholdplot(time_fall_occured,s(tpc_dataplot),'ro'); % put red circle on threshold valueplot(xlim,[threshold threshold], 'r') % out a reference line with threshold valuexlabel('Time(sec)'); ylabel('Amplitude');title('Accelerometer value');%display the datapeaktime_fall_occured