Scholarly article on topic 'Smartphone Based Data Mining for Fall Detection: Analysis and Design'

Smartphone Based Data Mining for Fall Detection: Analysis and Design Academic research paper on "Materials engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"fall detection" / "machine learning" / "supervised learning" / "cellular phone" / smartphone}

Abstract of research paper on Materials engineering, author of scientific article — Abdul Hakim, M. Saiful Huq, Shahnoor Shanta, B.S.K.K. Ibrahim

Abstract Falls can be devastating to the affected individual, yet a common event and hence one of the major causes of injury or disability within the aged population in Malaysia and worldwide. This paper aims to detect human fall utilizing the built inertial measurement unit (IMU) sensors of a smartphone attached to the subject's body with the signals wirelessly transmitted to remote PC for processing. Matlab's mobile and the Smartphone Sensor Support is used to acquire the data from the smartphone which is then analysed to design an algorithm for the detection of fall. Falls in human are usually characterized by large acceleration. However, focusing only on a large value of the acceleration can result in many false positives from fall-like activities such as sitting down quickly and jumping. Thus, in this work, a threshold based fall detection algorithm is implemented while a supervised machine learning algorithm is used to classify activity daily living (ADL). This combination has been found effective in increasing the accuracy of the fall detection. The aim is to develop and verify the high precision detection algorithm using Matlab Simulink, followed by a few real time testing.

Academic research paper on topic "Smartphone Based Data Mining for Fall Detection: Analysis and Design"

(8)

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedia Computer Science 105 (2017) 46-51

2016 IEEE International Symposium on Robotics and Intelligent Sensors, IRIS 2016, 17-20 December 2016,

Tokyo, Japan

Smartphone Based Data Mining for Fall Detection: Analysis and Design

Abdul Hakim, M. Saiful Huq, Shahnoor Shanta, B.S.K.K. Ibrahim

Faculty of Electrical and Electronic Engineering, University Tun Hussein Onn Malaysia 86400 Parit Raja, Johor, Malaysia

Abstract

Falls can be devastating to the affected individual, yet a common event and hence one of the major causes of injury or disability within the aged population in Malaysia and worldwide. This paper aims to detect human fall utilizing the built inertial measurement unit (IMU) sensors of a smartphone attached to the subject's body with the signals wirelessly transmitted to remote PC for processing. Matlab's mobile and the Smartphone Sensor Support is used to acquire the data from the smartphone which is then analysed to design an algorithm for the detection of fall. Falls in human are usually characterized by large acceleration. However, focusing only on a large value of the acceleration can result in many false positives from fall-like activities such as sitting down quickly and jumping. Thus, in this work, a threshold based fall detection algorithm is implemented while a supervised machine learning algorithm is used to classify activity daily living (ADL). This combination has been found effective in increasing the accuracy of the fall detection. The aim is to develop and verify the high precision detection algorithm using Matlab Simulink, followed by a few real time testing.

© 2017 The Authors.Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors(IRIS 2016).

Keywords: fall detection, machine learning, supervised learning, cellular phone, smartphone

1. Introduction

In Malaysia, people with disabilities can be considered as a part of the most vulnerable minority group within the Malaysian population [1]. World Health Organization (WHO) and World Bank (2011) estimated that there are approximately 15% of the world population having some form of disability. According to the Malaysian Department of Social Welfare in December 2015, there were 305,640 disable people in Malaysia. Among them, 27,363 are visual 39,303 hearing, 180 speech, 106,252 physical, 117,699 learning, 2,130 mental and 12,713 multiple disable people [2]. Thus, standing balance for elderly and in post-acute stroke is a part of physical disabilities faced by physiotherapists. Usually stroke patients showed excessive postural sway and instability [3]. Therefore, this problem can potentially cost individuals more money and eventually burden the national economy and destabilize the society. In this context, assistive devices that could ease this significant wellbeing are a dire social need.

Existing fall detection approaches can be explained and categorized into three different classes depending on the different detection methods used. They are: (i) device based, (ii) ambience sensor based and (iii) vision based. The wearable devices can be further classified into (a) posture based and (b) motion based devices. The ambience based devices can be further classified into (a) presence- and (b) posture- based sensors. The vision based systems, on the other hand, can be further categorised into three classes depending on (a) shape change, (b) inactivity and (c) 3D head motion.

Wearable devices are based entirely on sensing the acceleration since fall is primarily characterized by large accelerations. Various distinctive methodologies for automatic detection of falls have been reported in recent years [4]. Some rely on the

1877-0509 © 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of the 2016 IEEE International Symposium on Robotics and Intelligent Sensors(IRIS 2016). doi:10.1016/j.procs.2017.01.188

orientation of the body relative to the ground or near horizontal orientation of the subject, following the fall [4]. Most fall-detection frameworks rely on detecting the stun experienced by the body upon fall using accelerometers [4-6]. Some recent studies implementing fall detection using tri-axial accelerometers have reported better results in detection of fall than the previous systems. These wearable devices have their advantages as well as disadvantages. The largest advantage is the lower cost of wearable device. Installation and setup of the design is also not very complicated. Therefore, the devices are relatively easy to operate. The disadvantages include intrusion and fixed relative relation with the subject, which could cause the device to be easily disconnected. Such disadvantages make wearable devices an unfavourable choice for the elderly [7].

Ambience based devices are based on sensing the vibrational data. The detection of the events and changes using vibrational data can be useful in many ways such as monitoring, tracking and localisation. A completely passive and unobtrusive system was introduced in [8] where the authors have developed the working principle and the design of floor vibration-based fall detector. In this approach, the detection of human fall is estimated by monitoring the floor vibration patterns, since floor's vibration signature generated by human fall is different from normal activities, such as walking. With these approaches pressure sensors are used to sense high pressure of the object due to the object's weight for detection and tracking. It is very cost effective and less intrusive for the implementation of surveillance systems. Still, it has huge disadvantages of sensing pressure of everything in and around the object and generating false positives, which leads to low detection accuracy [7].

Vision based approaches are based on using cameras. Nowadays, the application of such approaches are on the rise, especially in home assistive systems, as they convey multiple advantages over other systems. Cameras can be used to detect multiple events simultaneously with less interruption. Image analysis requires efficient and accurate shape modelling methods. Mihailidis [9] placed a video camera on the ceiling and developed scene algorithms to detect a fall. The system which was tested on 21 volunteers carrying out simulated falls, was reported successful in detecting 77% of falls. This approach, learned through extended observation such as the interpretation of human activities in a scene, provides contextual representation of activities. Vision based systems tend to deal with intrusion better than the other approaches, yet it also introduces some ethical issues concerning the confidentiality and privacy of the user.

Recently, there is a growing interest for identifying and detecting fall events using an inertial measurement unit (IMU) which is a combination of accelerometers, gyroscopes and magnetometers [10]. Various studies have shown that the IMU based wearable systems can be effectively used to recognize fall events by examining the impact of the body with the ground and the body orientation prior to, during, and following the fall. However, the techniques fail to bolster pervasive fall detection since they require particular hardware and software, which causes increase in cost and limits commercial sustainability. Smartphone is still not being used widely in comparison with the previous categories according to the several studies conducted on fall detection recently. In fact, this is a relatively novel technology considering the fact that the first study using smartphone only appeared in 2009 [11]. Although these studies still need to incorporate a more exhaustive evaluation, there are signs that this area is in fact an emerging field [12].

Universal fall detection can be achieved by taking advantage of a widely available technology, such as smartphones. Smartphone based fall detection techniques possess immense potential as they are ultra-portable, widely available while packed with high quality sensing and processing ability, featuring embedded motion sensors, increasingly powerful microprocessors, considerable memory capacity, open source operating systems, and telecommunication services, making them ideal for easily programmable and customizable for fall detection [11, 13].

Fall is usually characterized by larger acceleration and hence accelerometer can be easily used for its detection. However, such approach is prone to false positives as tasks like sitting down quickly or jumping may result in larger values of accelerations. Existing acceleration based fall detection solutions can be divided into two categories. The first category only analyses acceleration to detect fall while the second category utilizes both acceleration and body information to detect fall. The second category possesses increased accuracy and sensitivity in detection of human fall [14]. In this work, we propose to implement the latter category and compile the statistics of the output from the combination of various motion and set threshold to detect fall.

The ultimate objective of the overall design is to develop a wearable device that can be donned and doffed very easily, rendering it as a feasible safety device for day to day regular user activity. Since smart phones are widely used and are packed with high quality sensing elements, it holds promising potential to make use of such highly available, ultra-portable, hi-tech piece of equipment in detecting falls in vulnerable group humans, e.g. in elderly or geriatrics.

2. Methods

2.1. Hardware Platform

The design of the proposed fall detection system involves utilizing smartphone which incorporates a powerful central processing unit (CPU), adequate memory capacity, flexible software environment, and built in sensors. To acquire acceleration,

jf" ) IÜ

Fig. 1. System Design

orientation and angular velocity data for analysis, the Sony C6002 Xperia Z was selected. It features an embedded IMU (i.e. triaxial accelerometer, tri-axial gyroscope, proximity and compass), an adequately powerful CPU (Quad-core 1.5Ghz krait) and a substantial memory capacity (16GB internal storage, 2GB RAM). To validate the developed system, Apple iPhone 4s was selected, featuring an embedded IMU (i.e., tri-axial accelerometer, tri-axial gyroscope, digital compass, proximity and ambient light), an adequately powerful CPU (Dual-Core ARM Cortex A-9), a substantial memory capacity of 16GB internal storage and 512MB RAM.

The smartphone application used in this study was MATLAB Mobile. Recently, MathWorks developed this software which is a lightweight desktop for iPhone, iPad or Android™ devices that connect to a MATLAB session running on MathWorks Cloud or computer. This software can evaluate MATLAB commands, run scripts, create and manipulate plots and figures and view the results. MATLAB mobile can acquire data from built in sensors which includes motion sensors like accelerometer and position sensors like the GPS. Measurements such as acceleration, magnetic field, latitude, longitude and altitude can be viewed on the smartphone. Since we are interested with acceleration, orientation and angular velocity only these measured parameters will be sent to a computer for further analysis. Fig. 1 shows the conceptual schematic of the real-time smartphone based fall detection system. Acceleration and gyroscope is referred respectively asAj, ,4V, As and cj^, ii)v, (i)z. In addition, orientations measures are henceforth referred to SD(pitch), 0v(yaw), and 0r(roll) representing Euler rotation angles. All these parameters are used for analysis and classification of the fall events.

2.2. Experimental Setup

Eight healthy subject s took part in an experiment, six of them for training the classifier and two of them for validation. The activities consist of four types of fall (fall forward, fall backward, and fall toward the left and right directions) and activity daily living (ADL) which are sitting, standing, walking, laying, walking upstairs and walking downstairs. The smartphones are located in the right, left and front-pockets. Three axis data of accelerometer, gyroscope and orientation in each sensor node is acquired during the experiment. Data are sampled at 10 Hz and are transmitted via wireless transmission to a remote PC for further processing. Falls and ADLs were annotated by custom written application in Matlab on the remote PC. Thus, ADLs are classified using machine learning approach, where the data need to be combined and classified using the training session and the validation. Fall detection using design threshold methods, has been shown to accurately identify fall events in post-analysis [13, 15, 16]. The program timed each activity for 10 seconds by recording the time for start and end of each activity for one person.

2.3. Fall Events

Six subjects were told to intentionally fall onto a 15cm thick cushion. Thus, the sample data were collected within 10 seconds and sent remotely to a PC via the Wi-Fi. Within these 10 seconds, due to an impact with the cushion, even though not unintentional fall, there is difference value of resultant fall and ADL. Equation 1 defines the resultant vector AT for both the events. This process was performed for fall forward, fall backward, and falls in the right left directions. Thus, from the ADL and fall resultant vector, the level of threshold which is above the ADL resultant vector was computed. For event detection, the undefined events that exceed the level threshold value are classified as potential fall events.

At = J A* +4 +AI (1]

2.4. ADL events

One of the overall aims of the desired system in this work is to be able to distinguish between falls and ADLs. During the experiment, postures and activities such as sitting, standing, walking upstairs, walking downstairs, walking and laying were maintained to 10 seconds for each subject. Six subjects with ADLs sample data was collected to be trained for machine learning and the data from two other subjects were used for validation. In this section, the signals are detected from embedded accelerometer, gyroscope and orientation sensors in the smart phone. Machine learning consists of supervised learning and unsupervised learning. Since ADL is predicted activity being done by human beings, it can be classified as supervised machine learning. Firstly, the collected data need to be trained and iterated until the optimal model is obtained as shown in Fig. 2.

Fig. 3. Validation Workflow

Angular Velocity

In the pre-process phase, the loaded data was labelled manually as events since the data were collected according to ADLs. Next, the data was filtered as mean, standard deviation and principle component analysis. The three pre-processed data gives distinctive informative information of the collected data since the goal is to find discriminatory values of data to make the predictor work as well as possible.

Four different classification algorithms were used for detection and classification: support vector machines (SVM), Decision trees, Nearest Neighbour Classifiers and Discriminant Analysis. Matlab has the machine learning toolbox where all the listed algorithm can be trained. V-fold cross validation were used for real practice. The level of accuracy was chosen to be used for validating the fall detection algorithm.

New raw data was used to validate and test the optimal model in order to integrate in the fall detection algorithm as shown in Fig. 3.

2.5. Fall Detection Algorithm

A fall detection algorithm is still under development and the control algorithm will be developed using SIMULINK for real time testing. Human activities can be divided into two categories which are static posture and dynamic transition. Therefore, those categories can produce false positive fall activity. In this phase, the algorithm will detect false positives from fall-like activities such as sitting down quickly and jumping since there is model of machine learning that recognize the ADL. Since falls represent a form of impact of a body to the ground, it means the resultant acceleration will be

above a threshold value and the model will detect as laying posture. Thus, the algorithm will detect fall as shown in Fig. 4. The

ADL is processed using the optimal model of machine TABLE I: Accuracy of Algorithm learning, the fall detection is set up conferring to the threshold

, value. Since falls can be detected either false positive or false negative, these two data can classify if the positive fall is detected. The condition is when threshold value is above the set value and the body is categorize as laying position.

Fig. 4. Workflow of fall algorithm

Algorithm

Acceleration, Angular Velocity, Orientation (%)

Acceleration, Angular Velocity

Acceleration, Orientation (%)

Acceleration

Decision Tree

Complex Tree Medium Tree Simple Tree Discriminant Analysis

Discriminant Quadratic Discriminant Support Vector Machine

Linear SVM Quadratic SVM Cubic SVM Fine Gaussian SVM Medium Gaussian SVM Coarse

Gaussian SVM Nearest Neighbour Classifier

Fine KNN Medium KNN Coarse KNN Cosine KNN Cubic KNN Weighted KNN

96.0 82.5

94.5 92.0 80.3 92.8 91.0 94.0

The results are shown

81.3 81.5 73.8

86.8 91.3 89.7 79.2

86.2 84.8 74.3 85.5 84.3 87.0

93.7 93.5 78.3

94.8 98.3

97.7 86.0

95.5 94.5 84.3 94.3 93.2 95.2

77.2 79.0 75.2

81.8 85.3 81.5

80.7 82.2

74.7 79.5 81.2

3. Results and Analysis

Various machine learning algorithms were used to train the classifiers for classifying the ADL. To do so, 3600 samples data of six subject described in section II were first manually labelled with the activities, which were then assigned to separate attribute vectors. Finally these vectors were used as training data for four machine learning algorithms: Support Vector Machines (SVM), Decision Trees, Nearest Neighbor Classifiers and Discriminant Analysis. Machine learning experiments were conducted in two steps. In the first step of the machine learning experiment, the classification accuracy of four main algorithm for all three parameters which are acceleration, angular velocity and orientation were compared.

i in Table I. The accuracy was computed with the V-fold cross validation. The accuracy of best attribute set for each algorithm is given in bold type; the accuracy of the best algorithm for each attribute set is shown on grey background.

The following step of the machine learning experiment involved using the same algorithm for computing the accuracy performance for the parameter set: acceleration and angular velocity, acceleration and orientation and acceleration itself. The

purpose of this was to compare the performance of the parameter set so that the best one can be chosen for the ADL classification.

Accordingly, Table I shows that the best algorithm to use for the model is Support Vector Machine, since it shows the highest accuracy for all the attribute set. Within four attribute sets accuracy above 90% were obtained for some of them and the optimal attribute set is proposed to be the combination of acceleration, angular velocity and orientation.

According to the confusion matrix as shown in Fig. 5, the most accurately classified positions are the laying, sitting and standing positions where the positive predictive values (PPV) of 100% and false discovery rates (FDR) of 0% are obtained. For walking classification there were 1% FDR for the identification of the walking downstairs and 99% PPV for the walking case. For walking downstairs, SVM algorithm predicts 97% PPV, 3% is predicted as FDR out of which 2% as walking and 1% as walking upstairs. 98% PPD prediction

was found for walking

upstairs FDR was

and only found

2% for

Fig. 5. Confusion Matrix

Forward Backward Left Right

Average Value

26.11 24.67 22.67 23.54 24.25

walking downstairs. However, there is a small confusion between walking upstairs, walking downstairs and walking but the accuracy is above 97%.

Fig. 6 demonstrates the resultant vector of ADL with 3600 sample data where it shows the noisy data during walking upstairs, walking downstairs and walking. Thus, this activity can lead to false positive for threshold value system. The maximum value for ADL indicate 24.2 m/s*.

Fig. 7 determine the resultant acceleration of fall waveform of six subjects. There are four events, viz. fall forward,

fall backward and fall right and left. The collected data is considered as intentional fall, thus it can be concluded that subject were trying to fall and the collected data does not in fact represent unintentional fall. As fall is usually characterised by larger acceleration, the threshold value is set higher with respect to ADL. Table II shows the average peak value of fall events, hence the selected value for threshold value is 24.25

In the proposed approach, a fall-like event is defined as an acceleration peak of magnitude greater than 2.5g of threshold value. In general, the threshold value ranging from 2.5g to 3.5g have been widely used in other fall detection systems reported in the literatures [17, 18]. The 2.4g value is small enough to avoid false negative, since ADL peak value are 2.4g as shown in table II in the collected data. Thus, there is possibility of having more false negatives if only the acceleration is being utilized. Meanwhile, to overcome this problem, machine learning algorithm with high accuracy should be utilized to make sure fall is positively detect.

4. CONCLUSIONS

Smartphones can be considered as self-contained devices for such application, since it presents a mature hardware and software environment for pervasive fall detection. We wanted to use smartphone and standard state of the art machine learning algorithms in order to perform a lusty fall detection and ADL classification. By applying supervised machine learning such as SVM the results with near flawless accuracies were obtained in all cases. This was possible with the standard IMU sensors found in almost all modern smartphones and could thus be implemented in phone apps.

Smartphones are comprised of advantageous platforms to analyze movements and hence can potentially detect falls. They have built-in communication protocols that allow simple data logging to the device and wireless transmission possible. This allows real-time response or, in an experimental setting, compliance verification. Because mobile phones are widely adopted, compliance without verification is already high, as people are used to carrying them [19]. Previous work showed how the accelerometer in mobile phones can be used to identify activities such as sitting, standing, walking and running, fall also being mentioned in their works [20-24]. Our approach using SVM is also supported by the results which show that activity recognition

can be increased with the accuracy level as high as 99%, when the combination of acceleration, angular velocity and orientation parameters are utilized compared to using them separately. The least accuracy is obtained using acceleration itself. This could increase the accuracy of fall detection as well as the false negatives since fall detection using just the threshold value is prone to false positives.

Although, the analyses showed high accuracy, there are several issues that will still need to be addressed before such technique could be implement in real life situations. For instance, the smartphone used in this study were located in an identical position with all the subjects, which is very unlikely to be the case in real life. This allowed highly stereotypical measurement that might have boosted the accuracy ratings.

5. Discussions

Fall detection promises to be important in the context of healthcare [19], thus accuracy and practicality is vital in this sort of application. Many of the fall detection approaches try to implement wearable devices on body parts such as chest, waist and head. But, it seems not very practical because of the way the users need to wear a device which might not be very comfortable in performing their daily activity. In this paper, we propose utilizing smartphone as a tool for fall detection development. We analyse the fall events and ADLs to design algorithms that can increase the accuracy of fall detection with pleasant position of the mobile phone. It is shown that using acceleration, angular velocity and orientation the accuracy of ADL recognition and fall detection is increased substantially.

ACKNOLEDGEMENT

This research was supported by the Office for Research, Innovation, Commercialization and Consultancy Management (ORICC), University Tun Hussein Onn Malaysia.

References

1. Kamarulzaman, K. (2007) Adult learning for people with disabilities in Malaysia: Provisions and services, The Journal of Human Resource and Adult Learning. 3.

2. Islam, M. R. & Cojocaru, S. (2015) Advocacy for Empowerment: A Case of the Learning Disabled People in Malaysia, Revista de cercetare [i interven] ie social. 50, 38-52.

3. Chern, J.-S., Chang, H.-S., Lung, C.-W., Wu, C.-Y. & Tang, S. F. (2013). Static ankle-foot orthosis improves static balance and gait functions in hemiplegic patients after stroke. Paper presented at the Engineering in Medicine and Biology Society (EMBC), 2013 35th Annual International Conference oof the IEEE.

4. Doughty, K., Lewis, R. & Mcintosh, A. (2000) The design of a practical and reliable fall detector for community and institutional telecare, Journal of Telemedicine and Telecare. 6, 150-154.

5. Prado, M., Roa, L., Reina-Tosina, J. & Sánchez, G. (2004). Preliminary evaluation of a full-time falling monitor for the elderly. Paper presented at the Engineering in Medicine and Biology Society, 2004 IEMBS'04 26th Annual International Conference of the IEEE.

6. Noury, N., Barralon, P., Virone, G., Boissy, P., Hamel, M. & Rumeau, P. (2003). A smart sensor based on rules and its evaluation in daily routines. Paper presented at the Engineering in medicine and biology society, 2003 Proceedings of the 25 th annual international conference of the IEEE.

7. Mubashir, M., Shao, L. & Seed, L. (2013) A survey on fall detection: Principles and approaches, Neurocomputing. 100, 144-152.

8. Alwan, M., Rajendran, P. J., Kell, S., Mack, D., Dalal, S., Wolfe, M. & Felder, R. (2006). A smart and passive floor-vibration based fall detector for elderly. Paper presented at the Information and Communication Technologies, 2006ICTTA'06 2nd.

9. Lee, T. & Mihailidis, A. (2005) An intelligent emergency response system: preliminary development and testing of automated fall detection, Journal of telemedicine and telecare. 11, 194-198.

10. Maenaka, K. (2008). MEMS inertial sensors and their applications. Paper presented at the Networked Sensing Systems, 2008INSS 2008 5th International Conference on.

11. Sposaro, F. & Tyson, G. (2009). iFall: an Android application for fall monitoring and response. Paper presented at the Engineering in Medicine and Biology Society, 2009 EMBC 2009 Annual International Conference of the IEEE.

12. Igual, R., Medrano, C. & Plaza, I. (2013) Challenges, issues and trends in fall detection systems, BiomedEng Online. 12, 1-66.

13. Dai, J., Bai, X., Yang, Z., Shen, Z. & Xuan, D. (2010). PerFallD: A pervasive fall detection system using mobile phones. Paper presented at the Pervasive Computing and Communications Workshops (PERCOM Workshops), 2010 8th IEEE International Conference on.

14. Kim, J., Kwak, H., Lee, H., Seo, K., Lim, B., Lee, M., Lee, J. & Roh, K. (2012). Balancing control of a biped robot. Paper presented at the Systems, Man, and Cybernetics (SMC), 2012 IEEE International Conference on.

15. Bourke, A. K. & Lyons, G. M. (2008) A threshold-based fall-detection algorithm using a bi-axial gyroscope sensor, Medical engineering & physics. 30, 8490.

16. Li, Q., Stankovic, J. A., Hanson, M. A., Barth, A. T., Lach, J. & Zhou, G. (2009). Accurate, fast fall detection using gyroscopes and accelerometer-derived posture information. Paper presented at the Wearable and Implantable Body Sensor Networks, 2009 BSN 2009 Sixth International Workshop on.

17. Bourke, A., O'brien, J. & Lyons, G. (2007) Evaluation of a threshold-based tri-axial accelerometer fall detection algorithm, Gait & posture. 26, 194-199.

18. Li, Q., Stankovic, J., Hanson, M., Barth, A. T., Lach, J. & Zhou, G. (2009). Accurate, fast fall detection using gyroscopes and accelerometer-derived posture information. Paper presented at the Wearable and Implantable Body Sensor Networks, 2009 BSN 2009 Sixth International Workshop on.

19. Albert, M. V., Kording, K., Herrmann, M. & Jayaraman, A. (2012) Fall classification by machine learning using mobile phones, PloS one. 7, e36556.

20. Brezmes, T., Gorricho, J.-L. & Cotrina, J. (2009) Activity recognition from accelerometer data on a mobile phone in Distributed computing, artificial intelligence, bioinformatics, soft computing, and ambient assisted living pp. 796-799, Springer.

21. Gyorbíró, N., Fábián, Á. & Hományi, G. (2009) An activity recognition system for mobile phones, Mobile Networks and Applications. 14, 82-91.

22. Bieber, G., Voskamp, J. & Urban, B. (2009) Activity recognition for everyday life on mobile phones in Universal Access in Human-Computer Interaction Intelligent and Ubiquitous Interaction Environments pp. 289-296, Springer.

23. Kwapisz, J., Weiss, G. & Moore, S. (2013) Activity recognition using cell phone accelerometers. SIGKDD Explor. Newsl. 12 (2), 74-82 (2011) in

24. Wang, Y., Lin, J., Annavaram, M., Jacobson, Q. A., Hong, J., Krishnamachari, B. & Sadeh, N. (2009). A framework of energy efficient mobile sensing for automatic user state recognition. Paper presented at the Proceedings of the 7 th international conference on Mobile systems, applications, and services.