Skip to main content

Table 3 Comparison of acceleration based fall detectors using external accelerometers

From: Challenges, issues and trends in fall detection systems

Article

Year

Basis

Detection technique

Fall types

Subjects

Declared perform

Position

Elders Yes/No

Comments

Lindeman et al.[33]

2005

A fall detector placed at head level

TBM considering the spatial direction of the head, the velocity right before the initial contact with the ground and the impact

Falls to the front, side with a 90° turn, back, back with hip flexion.

A young volunteer and an elderly woman (83 years)

High sensitivity and specificity

Ear

Yes

Accelerometers were integrated into a hearing-aid housing, which was fixed behind the ear

Falls backwards against a wall, while picking up an object and collapse.

Chen et al.[34]

2005

Detect the occurrence of a fall and the location of the victim

TBM considering the impact and the change in orientation

Backward and sideways falls

2 subjects

The acceleration for ADL is much less than the observed from falling

Waist

No

The final orientation of the wearer is considered

Zhang et al. [45]

2006

Fall detection using machine learning strategies

MLM. 1) Extraction of temporal and magnitude features from the acceleration signal, 2) One-class Support Vector Machine classifier

Soft fall

12 subjects (8 males, 4 females, ages 10–70)

Accuracy 96,7%

Waist

Yes

To the best of our knowledge, this study is the first in using machine learning techniques

Hard fall in the ground, stairs and slopes (using a mannequin)

Bourke et al.[35]

2007

Investigation into the ability to discriminate between falls and ADL

TBM using information from the impact

Forward falls, backward falls and lateral falls left and right, performed with legs straight and flexed

10 subjects (ages 21–29)

Trunk

Trunk, thigh

Yes

The trunk appears to be the optimum location for a fall sensor

SP:100%

10 community-dwelling elderly subjects (3 females, 7 males, ages 70–83)

Thigh

SP: 83.3%

Doukas et al.[46]

2007

Accelerometers transmit patient movement data wirelessly to the monitoring unit

MLM. The acceleration in the three axis is classified using Support Vector Machine

Not specified

1 subject

SE: 98.2%

Foot

No

If a fall is suspected it also transmits video images to remote monitoring units

SP: 96.7%

Kangas et al.[36]

2008

Comparison of 3 low-complexity algorithms

TBM considering the start of the fall, the velocity, the impact and the lying posture

Forward, backward, and lateral falls

3 volunteers (1 female, 2 males; ages 38, 42, 48)

Waist

Wrist, head, waist

No

Waist worn accelerometer might be optimal for fall detection considering the fall associated impact and the posture after the fall

SP: 100%

SE: 98%

Kangas et al.[37]

2009

To validate the data collection of a new fall detector prototype

TBM considering two or more of the following phases of a fall event: start of the fall, falling velocity, fall impact, and posture after the fall

Syncope, tripping, sitting on empty air, slipping, lateral fall, rolling out of bed

20 subjects (40–65 years old), 21 voluntary older people (58–98 years old)

SP: 100%

Waist

Yes

Middle-aged persons could be considered to mimic the fall events of older people more adequately than young subjects would

SE: 97.5%

Li et al.[38]

2009

Fall detection system using both accelerometers and gyroscopes

TBM analyzing the intensity of the activity, the posture and whether the transition to a lying posture was unintentional or not

Forward, backward, sideways and vertical falls. Falling on stairs and fall against walls ending with a sitting position

3 male subjects (age 20)

SP: 92%

Chest, thigh

No

Human activities are divided into static postures and dynamic transitions

SE: 91%

Shan et al.[42]

2010

Investigation of a pre-impact fall detector

MLM 1) A discriminant analysis is applied to time-domain statistical characteristics to select the features, 2) Support Vector Machine is used for fall recognition

Forward falls, backward falls, lateral falls left and right (subjects were instructed to keep their postures for about 2 seconds after the fall)

5 male subjects (ages 21 – 28)

SP: 100%

Waist

No

Impending falls are detected in their descending phase before the body hits the ground

SE: 100%

Bianchi et al.[39]

2010

Augmentation of accelerometer-based systems with a barometric pressure sensor

TBM considering the impact, the postural orientation, and the change in altitude associated with a fall

Forward, backward and lateral falls (ending lying, with recovery, with attempt to break the fall)

20 subjects (12 male, 8 female; mean age: 23.7)

SP: 96.5%

Waist

No

A system based on a barometric pressure sensor is compared with an accelerometry-based technique.

The acceleration and air pressure data are recorded using a wearable device

SE: 97.5%

5 subjects (2 male, 3 female; mean age: 24)

Resting against a wall, then sliding vertically down to the end in the sitting position

5 subjects (5 male, mean age: 26.4)

Bourke et al.[40]

2010

It compares novel fall-detection algorithms of varying complexity

TBM considering the fall impact, the velocity and the posture

Forward falls, backward falls, lateral falls left and right all performed with both legs straight and with knees relaxed

10 male subjects (age 24–35)

SP: 100%

Waist

Yes

The algorithms were tested against ADL performed by elderly subjects

10 older subjects (6 male, 4 female, age 73–90)

SE: 94.6%

Lai et al.[41]

2011

Several acceleration sensors for joint sensing fall events

TBM to differentiate dynamic/static states using the acceleration of the three axis

Forward, backward, rightward or leftward falls

16 subjects

Accuracy 92.92%

Neck, hand, waist, foot

No

After a fall accident occurs, the system determines the level of injury

Bagala et al.[9]

2012

Benchmark the performance of published fall-detection methods when they are applied to real-world falls

TBM including, among others, the algorithms published in [35, 36]

Real-world falls: indoor/outdoor, forward /backward /sideward, impact against the floor /wall or locker before hitting the floor / sofa or bed/ desk

9 subjects (7 women, 2 men, age: 66.4±6.2)

Average 13 algorithms

Lower back

Yes

Algorithms that were successful at detecting simulated falls did not perform well when attempting to detect real-world falls

SP: 83.0% ±30.3%

15 subjects

29 subjects

SE: 57.0%

1 subject

±27.3%

Yuwono et al.[43]

2012

Use of a sophisticated fall detection method

MLM. 1) Discrete wavelet transform, 2) Associate a cluster to the input feature vector; fuse cluster information with input, 3) Combined classification (vote majority): Multilayer Perceptron and Augmented Radial Basis Function

Not specified

8 individuals (age 19–28)

SP: 99.6%

Waist

No

Training and clustering are done off-line. Clustering is done using Regrouping particle swarm optimization

SE: 98.6%

Kerdegari et al.[44]

2012

Investigation of the performance of different classification algorithms

MLM. Input is pre-processed using windowing techniques. Features include acceleration, angular velocity, velocity, position and time domain features: maximum, minimum, mean, range, variance and standard deviation. Several methods are compared.

With flexed knees: forward, backward, sideways falls

50 volunteers (18 male, 32 female, average age 32)

SE: 90.15%

Waist

No

Multilayer Perceptron, Naive Bayes, Decision tree, Support Vector Machine, ZeroR and OneR algorithms are compared.

Base on wall: backward, sideways falls

Backward falls sitting on empty, turning left and right

Results show that the Multilayer Perceptron algorithm is the best option

Cheng et al.[47]

2013

Daily activity monitoring and fall detection

TBM using a decision tree: 1) A decision tree is applied to the angles of all the body postures to recognize posture transitions, 2) the impact magnitude is thresholded to detect the falls

Four types of falls: from standing to face-up lying, face-down lying, left-side lying, and right-side lying

10 subjects (6 males, 4 females, age 22–26)

SE: 95.33%

Chest

No

Dynamic gait activities are also identified using Hidden Markov Models. Surface electromyography signals are combined with the acceleration signals.

SP: 97.66%

Thigh

  1. TBM: Threshold Based Method, MLM: Machine Learning Method.