sensors Article

Highly Portable, Sensor-Based System for Human Fall Monitoring Aihua Mao 1, *, Xuedong Ma 2 , Yinan He 1 and Jie Luo 3, * 1 2 3

*

School of Computer Science & Engineering, South China University of Technology, Guangzhou 510006, China; [email protected] School of Electronic and Information Engineering, South China University of Technology, Guangzhou 510006, China; [email protected] School of Fine Art and Artistic Design, Guangzhou University, Guangzhou 510006, China Correspondence: [email protected] (A.M.); [email protected] (J.L.); Tel.: +86-20-39380285 (A.M.)

Received: 12 August 2017; Accepted: 11 September 2017; Published: 13 September 2017

Abstract: Falls are a very dangerous situation especially among elderly people, because they may lead to fractures, concussion, and other injuries. Without timely rescue, falls may even endanger their lives. The existing optical sensor-based fall monitoring systems have some disadvantages, such as limited monitoring range and inconvenience to carry for users. Furthermore, the fall detection system based only on an accelerometer often mistakenly determines some activities of daily living (ADL) as falls, leading to low accuracy in fall detection. We propose a human fall monitoring system consisting of a highly portable sensor unit including a triaxis accelerometer, a triaxis gyroscope, and a triaxis magnetometer, and a mobile phone. With the data from these sensors, we obtain the acceleration and Euler angle (yaw, pitch, and roll), which represents the orientation of the user’s body. Then, a proposed fall detection algorithm was used to detect falls based on the acceleration and Euler angle. With this monitoring system, we design a series of simulated falls and ADL and conduct the experiment by placing the sensors on the shoulder, waist, and foot of the subjects. Through the experiment, we re-identify the threshold of acceleration for accurate fall detection and verify the best body location to place the sensors by comparing the detection performance on different body segments. We also compared this monitoring system with other similar works and found that better fall detection accuracy and portability can be achieved by our system. Keywords: fall detection; sensors; easy portability; body segment

1. Introduction The aging society is becoming a serious public health issue in many areas, especially in developed countries. According to the World Health Organization, falls are the second leading cause of accidental or unintentional injury deaths worldwide [1], and more than one third of elderly people fall once or more each year [2]. For this group of people, once they fall, they may suffer serious health problems. Damage may be greatly reduced if they have access to timely rescue. Thus, a reliable fall monitoring system has great application value and development prospects. In the last decade, many methods have been developed for fall detection; most of which are based on optical sensors [3–6], accelerometers [7–11], and accelerometers combined with gyroscopes [12–14]. The principle of the optical sensor-based method is to capture the image by visual sensors, such as digital cameras and Kinect, and then distinguishes the human body with other items using digital image processing algorithms to detect falls. For example, Yang et al. [15] used a Kinect sensor to capture three-dimensional (3D) depth images and the dense spatio-temporal context (STC). The dense STC algorithm was applied to track the head position. They calculated the distance between the head and floor plane. If the distance is lower than an adaptive threshold, then the centroid height of Sensors 2017, 17, 2096; doi:10.3390/s17092096

www.mdpi.com/journal/sensors

Sensors 2017, 17, 2096

2 of 15

the human was used to determine whether a fall accident occurred. Gasparrini et al. [16] proposed a fall detection method for indoor environments based on a Kinect sensor. In this method, an Ad hoc segmentation algorithm was used to analyze the raw depth data from the Kinect and recognize all the elements captured in the depth scene. The system extracts the elements and classifies blobs in the scene to recognize human subjects among the blobs, followed by a tracking algorithm between different frames. Lee et al. [17] used shape feature variation extracted from the captured image to detect abnormal events, such as fall, and utilized a shadow removal algorithm to improve the accuracy of object detection and tracking. Anderson et al. [18] recognized the silhouette of individuals from the image captured by cameras and then extracted the features and trained hidden Markov models to recognize future performances of these known activities, such as fall. Rougher et al. [19] used a single camera to obtain images and obtained the 3D trajectory of the human head and then distinguished fall from activities of daily living (ADL) by the trajectory. Methods based on optical sensors have high reliability and accuracy. However, their monitoring range is limited. Once the users leave the captured range of the picture or are blocked by other objects, falls cannot be detected. Furthermore, intensive computation is required to finish the detection, leading to difficulty in real-time monitoring. Accelerometer-based methods measure acceleration of the human body by an accelerometer, and falls are detected based on a threshold value of acceleration. For example, Mario et al. [20] proposed an algorithm to detect human basic movements, such as fall events from wearable measured acceleration data. The algorithm was designed to minimize computational requirements while achieving acceptable accuracy levels based on characterizing some particular points in the temporal series obtained from a single sensor. Hsieh et al. [21] reported a novel hierarchical fall detection algorithm based on the measured acceleration data involving threshold-based and knowledge-based approaches to detect a fall event. Kurniawan et al. [22] used a triaxis accelerometer to obtain the acceleration of the human body and then determined whether the fall occurs depending on the acceleration. Albert et al. [23] used the accelerometer of the mobile phone to obtain the acceleration data, which is applied in the machine learning classifier to detect falls and classify the type of falls. These methods are not limited by the monitoring space and are not vulnerable to external interference. However, they used only acceleration data, which often mistakenly determine ADL, such as jumping and running, as a fall event and then trigger a false fall alarm. Thus, the accuracy of this kind of method may be affected if it is only based on an accelerometer. Thus, to improve the accuracy, an alternative approach is to combine a gyroscope and even a magnetometer with the accelerometer to work together. Pierleoni et al. [24] fixed an inertial unit on the waist of a human body by a belt to obtain the acceleration and the attitude angle and then detect the falls. However, they can only send the detection result to a nearby device via Bluetooth. If nobody is nearby, the fall alarm would not be noticed by other people. Wibisono et al. [25] used an accelerometer and a gyroscope embedded in a mobile phone to obtain the acceleration and angular velocity of the human body. A threshold-based algorithm was applied to detect falls based on the acceleration and angular velocity. However, this kind of method based on the mobile phone required the mobile phone provided with acceleration and gyroscope, which is not available in some mobile phones. Besides that, the weight of most mobile phones is far greater than that of the sensor devices and may lead to discomfort when carried for prolonged periods. Furthermore, though the mobile phone now is a necessity in daily life, it is not practical to fix it on a body location all the time. He et al. [26] introduced a fall detection and alerting system consisting of a smart phone and a customized vest integrated with a triaxial accelerometer and a gyroscope. The raw triaxial gyroscope and accelerometer data are preprocessed by a program running on the smart phone by the Kalman filter to reduce the noise measurement from the sensors. Then, a Bayes network classifier was used to determine whether the individual falls or not. The system only considers four typical subcategories of ADLs and two kinds of falls and only tried the sensors on the vest near to the neck. Ntalampiras et al. [27] proposed a learning mechanism based on hidden Markov models for automatic recognition of physical activities of humans and assessed the performance based on previous datasets from an accelerometer and a gyroscope. However, the performance of this kind of method based on machine learning is dependent on the training dataset.

Sensors 2017, 17, 2096 Sensors 2017, 17, 2096

3 of 15 3 of 15

In this datasets study, we propose a highly portable, sensor-based system for monitoring human previous from an accelerometer and a gyroscope. However, the performance of this kind falls. of A small self-packaged sensorlearning unit andisadependent mobile phone composed system. The sensor unit includes method based on machine on the training this dataset. this study, weapropose a highly portable, sensor-based system for monitoring falls. Aby a triaxisInaccelerometer, triaxis gyroscope, and a triaxis magnetometer, which can be human easily ported p 2 ), A , A , sensor unitsquare and a mobile composed(RMS this system. The2 + sensor unit thesmall user self-packaged to obtain the root mean (RMS)phone of acceleration = ( Ax Ay2 + Azincludes x y triaxis accelerometer, a triaxis gyroscope, and a triaxis magnetometer, which can be easily ported anda A are the x-axis, y-axis, and z-axis acceleration) and orientation of the user. All the sensor data are z byto the user to obtain the via rootBluetooth mean square (RMS) of acceleration ( RMS  processed Ax, detection Ay, and ( Ax 2  Ay 2 by Az 2a) ,fall sent the mobile phone communication and is further algorithm. If falls arey-axis, detected recognized, the mobile phone will issue an All alarm automatically x-axis, andand z-axis acceleration) and orientation of the user. theand sensor data are Az are the sent to the mobile phone via Bluetooth communication and is further processed by a fall detection call the emergency contact for timely rescue. To determine the threshold of acceleration for accurate If falls arethe detected andlocation recognized, thethe mobile phone will issue an alarm and fallalgorithm. detection and verify best body to place sensors, we designed seven fall activities automatically callconduct the emergency contact for timely rescue. To the determine the on threshold of and nine ADLs and the fall detection experiment by placing sensor unit the subjects’ acceleration for accurate fall detection and verify the best body location to place the sensors, we shoulder, waist, and foot, respectively. We also compare this system with other similar works in terms designed portability, seven fall activities and nine and conduct the fall detection experiment bydue placing of accuracy, and comfort. OurADLs system has good accuracy and is highly portable to the the sensor unit on the subjects’ shoulder, waist, and foot, respectively. We also compare this system components of a small sensor unit and a mobile phone. Furthermore, our system is comfortable to the with other similar works in terms of accuracy, portability, and comfort. Our system has good user even if used for a long time. accuracy and is highly portable due to the components of a small sensor unit and a mobile phone. The main contributions of this study include the following: Furthermore, our system is comfortable to the user even if used for a long time. mainacontributions of this include system, the following: 1. WeThe design highly portable fallstudy monitoring which only consists a small self-packaged sensor unit and a mobile phone to detect falls, and a fall alarm canconsists be automatically issued to the 1. We design a highly portable fall monitoring system, which only a small self-packaged emergency contact seek timely rescue. this system great detection accuracy sensor unit and a to mobile phone to detectFurthermore, falls, and a fall alarm can has be automatically issued to forthe both falling and ADL.to seek timely rescue. Furthermore, this system has great detection emergency contact for both falling and ADL. 2. Weaccuracy re-identify an acceleration threshold of 2.3 g by experimental observation for accurate fall 2. detection, We re-identify acceleration threshold of 2.3 g ADL by experimental for accurate fall which an is effective to avoid determining as falls, andobservation thus improving the accuracy whichBenefitting is effectivefrom to avoid determining as falls,can andachieve thus improving the ofdetection, fall detection. this new threshold,ADL our system better accuracy accuracy of fall detection. Benefitting from this new threshold, our system can achieve better compared with other similar works. compared with other similar works. Weaccuracy verify the best location of the human body to wear the sensor by conducting a series of 3. 3. experiments We verify the best location of the human body of to fall wear the sensor by the conducting a series of and comparison of the performance detection when sensor unit is placed experiments and comparison of the performance of fall detection when the sensor unit is placed on different body segments. The waist is the best one for sensor placement where the detection on different body segments. The waist is the best one for sensor placement where the detection accuracy is 100% without any misjudgment, whereas wearing the sensor in the shoulder and foot accuracy is 100% without any misjudgment, whereas wearing the sensor in the shoulder and may still mistakenly identify ADL as falls. foot may still mistakenly identify ADL as falls. 2. Fall Monitoring System 2. Fall Monitoring System The system consists of a small self-packaged sensor unit and a mobile phone with an Android The system consists of a small self-packaged sensor unit and a mobile phone with an Android operating system (Figure 1). The sensor unit is easily carried by the user, containing a microelectrometrical operating system (Figure 1). The sensor unit is easily carried by the user, containing a system (MEMS) sensor, a kernel processer, and a Bluetooth module. The sensors’ data are transmitted microelectrometrical system (MEMS) sensor, a kernel processer, and a Bluetooth module. The to the Android mobile phone via Bluetooth. developed (APP) runs on application the Android sensors’ data are transmitted to the Android A mobile phone application via Bluetooth. A developed system to receive sensor data from the sensor device and processes the raw data to calculate (APP) runs on the Android system to receive sensor data from the sensor device and processes thethe orientation acceleration. Based onand the acceleration. orientation and RMS, the orientation APP automatically falls raw data and to calculate the orientation Based on the and RMS,detects the APP based on a proposed fallfalls detection Once the fall is detected, an Once alarmthe is issued by the mobile automatically detects based algorithm. on a proposed fall detection algorithm. fall is detected, an phone and a call will automatically connect to the contact. alarm is issued by the mobile phone and a call willemergency automatically connect to the emergency contact.

Figure Highlyportable, portable,sensor-based sensor-based fall MEMS: microelectrometrical system. Figure 1. 1. Highly fall monitoring monitoringsystem. system. MEMS: microelectrometrical system.

Sensors 2017, 17, 2096

4 of 15

Sensors 2017, 17, 2096 Sensors 2017, 2096 Figure 2 17, shows

4 of 15

4 ofsensor 15 the internal details of the self-packaged sensor unit, wherein the MEMS 2 shows the internal details ofat the self-packaged sensor wherein the MEMS module Figure integrates a triaxis accelerometer the range of ±16 g, a unit, triaxis gyroscope at thesensor range of Figure 2 shows the internal details of the self-packaged sensor unit, wherein the MEMS sensor ◦ module integrates a triaxis accelerometer at the range of ±16 g, a triaxis gyroscope at the range ±2000 , and a triaxis magnetometer. The Kernel processor can reduce the measurement noise of from module integrates a triaxis accelerometer at the range of ±16 g, a triaxis gyroscope at the range of ±2000°, and triaxis magnetometer. The in Kernel processor can reduce the measurement noise afrom the ±2000°, sensor and by aaafiltering algorithm setup the processor. The Bluetooth module includes BC417 triaxis magnetometer. The Kernel processor can reduce the measurement noise from the sensor bybased a filtering algorithm setup inthe the processor. The Bluetooth module includes a sampling BC417 Bluetooth chipby on Bluetooth andin of theThe Bluetooth module is includes 15 m. The the sensor a filtering algorithm2.0, setup therange processor. Bluetooth module a BC417 Bluetooth chip based on Bluetooth 2.0, and the range of the Bluetooth module is 15 m. The sampling frequency of the from the human activities is set Hz. After noiseisreduction, sampling Bluetooth chipsensors based on Bluetooth 2.0, and the range of to the100 Bluetooth module 15 m. The the sampling frequency of the sensors from the human activities is set to 100 Hz. After noise reduction, the datafrequency from the of sensors are transmitted to the mobile phone via Bluetooth communication. The power the sensors from the human activities is set to 100 Hz. After noise reduction, the sampling data from the sensors are transmitted to the mobile phone via Bluetooth communication. supply module contains a power chip and to a battery. With the power supply communication. module, the device sampling data from the sensorssupply are transmitted the mobile phone via Bluetooth The power supply module contains a power supply chip and a battery. With the power supply power supply module contains aafter power supply the chip and a Thus, battery. With theispower supply canThe work without any additional power charging battery. the device veryThus, convenient module, the device can work without any additional power after charging the battery. the module, the device can work without any additional power after charging the battery. Thus, the to be carried by the users. device is very convenient to be carried by the users.

device is very convenient to be carried by the users.

Figure2.2.Small, Small, self-packaged self-packaged sensor Figure sensorunit. unit. Figure 2. Small, self-packaged sensor unit.

Figure 3 shows maininterfaces interfaces of of the the Android Android APP. forfor nearby Figure 3 shows thethemain APP. First, First,the theAPP APPsearches searches nearby Figure 3 shows the main interfaces of the Android APP. First, the APP searches for nearby Bluetooth devices and connects to the sensor unit. After establishing the connection and receiving Bluetooth devices and connects to the sensor unit. After establishing the connection and receiving Bluetooth devices and connects to the sensor unit. After establishing the connection and receiving the the sensorthedata, the APP calculates the Euler angle and then visualizes the Euler angle and sensor APPthe calculates the Euler angle and thenand visualizes the Eulerthe angle andangle acceleration the data, sensor data, APP calculates the Euler angle then visualizes Euler and acceleration by data charts. Meanwhile, a fall detection algorithm was used to detect falls in real time. by data charts.by Meanwhile, fall detection algorithm used was to detect falls in real a fall is acceleration data charts.aMeanwhile, a fall detection was algorithm used to detect fallstime. in realIftime. If a fall is detected, the APP will make an alarm sound and issue a call to the emergency contact. detected, APP will and issue a call toathe contact. If a fallthe is detected, themake APP an willalarm make sound an alarm sound and issue callemergency to the emergency contact.

Figure 3. Main interfaces of the Android application: (a) device search; (b) data visualization and Figure 3. Main interfaces of the Android application: (a) device search; (b) data visualization and

fall monitor; (c) makingof alarm and automatic calling. (a) device search; (b) data visualization and fall Figure 3. Main interfaces the Android application: fall monitor; (c) making alarm and automatic calling. monitor; (c) making alarm and automatic calling.

Sensors 2017, 17, 2096 Sensors 2017, 17, 2096

5 of 15 5 of 15

3. 3D Orientation Measurement of the Human Body To described by by yaw, yaw, To describe describe the posture of the human body, we use the Euler angle, which is described pitch, and roll angles to represent the body’s spatial orientation as shown in Figure 4. In the ideal angles to represent the body’s spatial orientation shown in Figure 4. case, aa single single3-axis 3-axisgyroscope gyroscopeplaced placed a rigid body provide the current orientation the onon a rigid body cancan provide the current orientation of theofbody body by performing integration calculation on gyroscope However, in practice, integration by performing integration calculation on gyroscope data.data. However, in practice, integration will will accumulate the white the reading effects, resulting in drift from expected values. accumulate the white noisenoise fromfrom the reading effects, resulting in drift from the the expected values. To To complement drift gyroscope data,sophisticated sophisticatedalgorithms algorithmsoften oftenfuse fuse the the accelerometer accelerometer and complement thethe drift gyroscope data, magnetometer ofof sensor data fusion is magnetometer data datadue dueto totheir theircompensable compensablecharacteristics. characteristics.The Theperformance performance sensor data fusion often determined by the corrector input for the accelerometer and magnetometer. is often determined by the corrector input for the accelerometer and magnetometer.

Figure the human human body. body. Figure 4. 4. Pitch, Pitch, roll, roll, and and yaw yaw angle angle of of the

As shown in Figure 5, the Euler angle of a human segment is estimated based on the gyroscope As shown in Figure 5, the Euler angle of a human segment is estimated based on the gyroscope data, accelerometer data, and magnetometer data. The estimation can be achieved based on Yean’s data, accelerometer data, and magnetometer data. The estimation can be achieved based on Yean’s model [28]. A Kalman filter is designed to fuse the sensor data mainly through two stages, namely, model [28]. A Kalman filter is designed to fuse the sensor data mainly through two stages, namely, applying the filter to the gyroscope data to predict the variables describing the next state and then applying the filter to the gyroscope data to predict the variables describing the next state and then estimating the orientation by using the accelerometer and magnetometer data to correct the estimating the orientation by using the accelerometer and magnetometer data to correct the prediction prediction from the gyroscope. In the estimation stage, extracting gravity from the accelerometer from the gyroscope. In the estimation stage, extracting gravity from the accelerometer doesn’t suffer doesn’t suffer from the drift issue, meanwhile, the data from the magnetometer measures the from the drift issue, meanwhile, the data from the magnetometer measures the calibrated magnetic calibrated magnetic field values for all three physical axes (x, y, z), especially in the circumstance of field values for acceleration all three physical (x, y, z),movement. especially inThese the circumstance of the acceleration the dynamic andaxes direction facts enable thedynamic correction for the and direction movement. These facts enable the correction for the predicted state. The of predicted state. The output of the Kalman filter is further used for quaternion estimation tooutput calculate the Kalman filter further segment used for and quaternion estimation to calculate the Euler and angle of a human the Euler angle of is a human the corrector input by the magnetometer accelerometer segment and the corrector input by the magnetometer and accelerometer data. Here, quaternion data. Here, quaternion was selected because of its lower computation cost. The quaternion was q is selected because cost. The quaternion q isqexpressed as (qw , qx , qy , and qz ), expressed as (qw, of qx, its qy, lower and qzcomputation ), where qw represents its scalar, and x, qy, and qz represent its vector. where its scalar, and qx ,calculating qy , and qz represent vector.ofBased on thesegment estimated w represents Based qon the estimated quaternion, the Euleritsangles the human isquaternion, easy using calculating the Euler angles of the human segment is easy using the following equation: the following equation:   Yaw = atan2(2q q − 2q q , 2q2 2 + 2q22− 1)  Yaw  atan2(2qxx qyy  2qww qzz , 2qww  2qxx  1) Pitch = arcsin(2qw qy − 2qx qz ) (1) Pitch  arcsin(2qw q y  2qx qz )  (1)    Roll = atan2(2qy qz − 2qw qx , 2q2w2+ 2q2z 2− 1)

 Roll  atan2(2q y qz  2qw qx , 2qw  2qz  1)

Sensors 2017, 17, 2096

6 of 15

Sensors 2017, 17, 2096 Sensors 2017, 17, 2096

6 of 15 6 of 15

Gyroscope Gyroscope

Magnetometer Magnetometer

Kalman Filter Kalman Filter

Accelerometer Accelerometer

Corrector input Corrector input

Quaternion estimation Quaternion estimation

Body segment Euler angle estimation Body segment Euler angle estimation

Figure 5. Estimation of the Euler angle of aofhuman segment. Figure 5.5.Estimation angle a human segment. Figure Estimationofofthe theEuler Euler angle of a human segment.

4. Fall Detection Algorithm 4.4.Fall Detection Algorithm Fall Detection Algorithm During a fall activity, thethe human body first loses balance and falls in in aincertain direction During a afall activity, human body first loses balance and falls a acertain direction During fall activity, the human body first loses balance and falls certain direction weightlessly, then impacts against the low objects, such as the ground. In this process, the human weightlessly, then impacts against the low objects, such as the ground. In this process, the weightlessly, then impacts against the low objects, such as the ground. In this process, thehuman human posture generally changes from standing to to lying or or nearly flat. Thus, thethe process of of falls cancan bebebe posture generally changes from standing flat. Thus, process posture generally changes from standing tolying lying ornearly nearly flat. Thus, the process offalls falls can divided into three phases as follows: (1) start, the human body lost balance and descents toward the divided dividedinto intothree threephases phasesasasfollows: follows:(1) (1)start, start,the thehuman humanbody bodylost lostbalance balanceand anddescents descentstoward towardthe the ground; (2) impact, the human body falls to the ground and comes in contact with the ground; and ground; (2) impact, the human body falls to the ground and comes in contact with the ground; and (3) ground; (2) impact, the human body falls to the ground and comes in contact with the ground; and (3)posture, posture, thehuman human body remains lying onthe the ground foratat least ashort short period time. the body remains lying on for least period ofoftime. (3) posture, the human body remains lying onground the ground for at aleast a short period of time. During these three phases, the supporting force for the human body from the ground changes During changes Duringthese thesethree threephases, phases,the thesupporting supportingforce forcefor forthe thehuman humanbody bodyfrom fromthe theground ground changes sharply. Thus, the acceleration of the body also changes sharply. Given that the acceleration sensor sharply. accelerationofofthe thebody body also changes sharply. Given the acceleration sharply. Thus, Thus, the acceleration also changes sharply. Given that that the acceleration sensor is sensor triaxial, acceleration of the human (RMS) can can be(RMS) calculated from triaxial acceleration isthe triaxial, the acceleration of body the body human body can be calculated from triaxial is triaxial, the acceleration of the human (RMS) be calculated from triaxial acceleration components. acceleration components. components. The triaxis accelerometer measures AA x, Ay, and Az based on forces acting on the accelerometer The accelerometer measures , and Az A based on on forces acting on the in Thetriaxis triaxis accelerometer measures x,y A y, and z based forces acting on accelerometer the accelerometer x ,AA in the the three axes. Thus, RMS will reach 1 g even if the is stationary toinfluence the of of axes.axes. Thus,Thus, RMS RMS will reach 1 g even the device stationary due due to the of gravity. in three the three will reach 1 gifeven if device the is device is stationary due to influence the influence gravity. As shown in Figure 6,during RMS during the start phase isclose nearly tochanges 0 to g and changes Asgravity. shown in Figure 6,inRMS the start phase nearly to 0close g and rapidly torapidly therapidly peak As shown Figure 6, RMS during the is start phase is nearly close 0 g and changes to due the peak due to the supporting force from the ground during the impact phase. Then, RMS supporting force from the ground during impact phase.the Then, RMSphase. remains almost to to thethe peak due to the supporting force from the the ground during impact Then, RMS remains almost stable withwith a slight fluctuation during the posture phase. stable withalmost a slight fluctuation during the posture phase. remains stable a slight fluctuation during the posture phase. Impact Impact

Start Start

Posture Posture

Figure 6. Root mean square (RMS) during a fall activity. Figure 6. Root mean square (RMS) during a fall activity. Figure 6. Root mean square (RMS) during a fall activity.

Given thatthat the start phase is not distinction between fallsfalls and and other activities, suchsuch Given the start phase is the not evident the evident distinction between other activities, Given that the start phase is not the evident distinction between falls and other activities, such as diving, the fall detection algorithm is based on the recognition of impact and posture phases of asof as diving, the fall detection algorithm is based on the recognition of impact and posture phases diving, the fall detection algorithm is based on the recognition of impact and posture phases of the the falls,falls, which can can be achieved by threshold comparisons of some features. The The RMSRMS displays which be achieved by threshold comparisons of some features. displays the falls,change which can beimpact achieved by threshold comparisons of can some displays distinctive in the phase of the Thus, RMSRMS can be selected as one of RMS the distinctive change in the impact phase of falls. the falls. Thus, be features. selected asThe one of features the features for fall detection. When RMS exceeded the acceleration threshold for impact (A ti), the impact phase for fall detection. When RMS exceeded the acceleration threshold for impact (Ati), the impact phase is detected. However, onlyonly based on the the fall detection algorithm mistakenly determines is detected. However, based on RMS, the RMS, the fall detection algorithm mistakenly determines

Sensors 2017, 17, 2096

7 of 15

distinctive change in the impact phase of the falls. Thus, RMS can be selected as one of the features for fall2017, detection. Sensors 17, 2096 When RMS exceeded the acceleration threshold for impact (Ati ), the impact phase 7 of 15 is detected. However, only based on the RMS, the fall detection algorithm mistakenly determines some ADLs ADL, such as as jumping andand crouching, maymay also also obtain a similar RMS ADLs as asfalls fallseasily. easily.Some Some ADL, such jumping crouching, obtain a similar distribution with falls, as shown in Figure 7. Thus, we detect the posture phase by the Euler angle to RMS distribution with falls, as shown in Figure 7. Thus, we detect the posture phase by the Euler achieve after the after impact is recognized. Given that the positive negative angle to precise achievedetection precise detection thephase impact phase is recognized. Given that the and positive and values of roll and pitch only represent the direction of the human body, the absolute value of roll negative values of roll and pitch only represent the direction of the human body, the absolute value and pitch usedisto represent the posture of the of human body. body. As shown in Figure 7a, during falls, of roll andispitch used to represent the posture the human As shown in Figure 7a, during the posture of the of human body changes from standing to lyingto (orlying near (or lying). the absolute falls, the posture the human body changes from standing nearTherefore, lying). Therefore, the ◦ to approximately 90◦ . value of roll (abs (Roll)) or pitch (abs (Pitch)) angles change from nearly 0 absolute value of roll (abs (Roll)) or pitch (abs (Pitch)) angles change from nearly 0° to Meanwhile, during the absolute value of roll pitch angles a small change, as shown approximately 90°. ADLs, Meanwhile, during ADLs, theand absolute value have of roll and pitch angles have in a Figure 7b. Thus, the roll and pitch can be used for posture phase detection through the comparison with small change, as shown in Figure 7b. Thus, the roll and pitch can be used for posture phase orientationthrough threshold posture (Otpwith ). If the absolute threshold value of roll pitch exceeds Ttpvalue after tp within detection theofcomparison orientation of or posture (Otp). If O the absolute theroll impact phaseexceeds (the time posture), the posture is detected. After recognition of or pitch Otpthreshold within Tfor tp after the impact phasephase (the time threshold forthe posture), the of the impact and posture phases, a fall of the human body is detected. Given that the recognition of posture phase is detected. After the recognition of the impact and posture phases, a fall of the human impactisand posture phases is based on threshold comparisons, the fallphases detection algorithm displays body detected. Given that the recognition of impact and posture is based on threshold good performance in detection real-time algorithm fall detection. Oncegood a fallperformance occurs and the RMS and fall Euler angle exceed comparisons, the fall displays in real-time detection. Once the thresholds, the fall is detected immediately. a fall occurs and the RMS and Euler angle exceed the thresholds, the fall is detected immediately.

Impact Posture

Start

(a)

(b)

Figure 7. RMS, roll, and pitch during falling and activities of daily living (ADL): (a) forward fall Figure 7. RMS, roll, and pitch during falling and activities of daily living (ADL): (a) forward fall ending ending up lying; (b) jumping. up lying; (b) jumping.

Meanwhile, the determination of the threshold for recognizing the impact and posture phases Meanwhile, theAs determination the threshold recognizing thephase impact posture phases is is very important. proposed byofPierleoni et al.for [24], the impact is and detected when RMS very important. As proposed Pierleoni et the impact phase detected when exceeds 2.5 g. However, theby sensitivity of al. fall[24], detection based on isthe 2.5 g is not RMS goodexceeds in the 2.5 g. However, the sensitivity of fall detection based on the 2.5 g is not good in the reference [24]. reference [24]. Thus, we conducted a set of preliminary experiments with 5 subjects to determine Thus, we conducted a set of preliminary experiments with 5 subjects to determine the threshold, which is the threshold, which is then used for validating the detection algorithm. The preliminary then used forhas validating the protocol detectionwith algorithm. The preliminary experiment has the samefrom protocol experiment the same the validation experiment. Our observation the with the validation experiment. Our observation from the preliminary experimental data found that preliminary experimental data found that the peak RMS of the impact phase is more than 2 g. Atithe is peak RMS of the impact phase is more than 2 g. A is re-identified by our pre-experiments and set as ti re-identified by our pre-experiments and set as 2.3 g in our system. The value of 2.3 g is evaluated 2.3 the g in data our system. The value ofand 2.3 gwaist is evaluated on the datahas from the shoulder and waist segments, on from the shoulder segments, which similar distribution during the fall which has similar distribution during the fall and ADL activities. For the foot segment, we and ADL activities. For the foot segment, we found that Ati during the range of 2.2 g~4.0 g canfound reach that A during the range of 2.2 g~4.0 g can reach the best detection performance. To facilitate the ti the best detection performance. To facilitate the experiment implementation, we choose 2.3 g as the experiment implementation, we choose 2.3 g as the same A for the shoulder, waist, and foot segments. ti same Ati for the shoulder, waist, and foot segments. The thresholds for posture phase recognition is Theas thresholds for posture set because as Otp = they 50 within Ttp = observed 2 s by referring [24] set Otp = 50 within Ttp = phase 2 s by recognition referring tois[24] have been to givetogood because they have observed to give good performance in our experiments. performance in ourbeen experiments. 5. Fall Detection Experiment 5. Fall Detection Experiment We design and conduct a set of experiments by employing subjects to perform specific activities We design and conduct a set of experiments by employing subjects to perform specific to test the performance of our system and also verify the best human body location to place the sensor activities to test the performance of our system and also verify the best human body location to unit. In literature, the previous methods have tried the sensor on different body segments individually, place the sensor unit. In literature, the previous methods have tried the sensor on different body segments individually, such as neck [26], waist [24,29,30], chest and thigh [31], and ankle [30]. Kangas et al. [32] compared the fall detection performance on the head, waist, and wrist and found obvious differences between them. Because the detection performance differs greatly when the sensor is placed on different body segments, our experiments tested the body segments of shoulder, waist, and foot, concerning the facts of easy portability and low disturbance during the fall activities.

Sensors 2017, 17, 2096

8 of 15

such as neck [26], waist [24,29,30], chest and thigh [31], and ankle [30]. Kangas et al. [32] compared the fall detection performance on the head, waist, and wrist and found obvious differences between them. Because the detection performance differs greatly when the sensor is placed on different body segments, our experiments tested the body segments of shoulder, waist, and foot, concerning the facts of easy portability and low disturbance during the fall activities. The activities designed in the experiment are based on the reality of the elderly fall in daily life, which includes two categories, namely, falls in daily life and ADL designed as interference activities for falls. As shown in Table 1, we divided 16 activities into two categories as follows: 7 falls and 9 ADLs. In daily life, most falls occur during stumbling on things or slipping on a smooth surface, and then the body falls forward or backward. Alternatively, being hit by something from the left and right sides then losing balance causes the body to fall sideways. Thus, the seven simulated falls are grouped into four subcategories, namely, backward fall, forward fall, lateral left fall, and lateral right fall. Furthermore, sometimes the body would not be lying on the ground horizontally, and they are further subcategorized by concerning the different ending-up conditions in backward and forward falls. ADL is designed to verify whether the fall detection system can evoke a fall alarm correctly. In an ideal case, an ADL should not be detected by the system as a fall. However, in some ADLs, the RMS and Euler angle of a certain body segment may have the same features as falls and then were mistaken as falls, decreasing the accuracy of fall detection. Thus, we need to verify the best body location to place the sensor to avoid misjudgment. Table 1. Falls and ADL designed in the experiment. Activities Backward fall

Ending up lying Ending up falling to the ground with elbows Ending up sitting after lying for 2 s

Forward fall

Ending up lying Falling on the knees, ending up lying

Lateral left fall Lateral right fall

Ending up lying Ending up lying

Fall Activities(FA)

Activities of Daily Living (ADL)

Standing-sitting and lying on a mattress-standing Standing-sitting on a chair-standing Taking stairs Bending over Walking Stretching shoulder Staggering forward Staggering backward Twisting waist

A positive or a negative outcome is expected in all the activities in the experiment protocol. If a fall activity evokes the fall alarm, it will be specified as a true positive outcome, and if an ADL evokes the fall alarm, it will be specified as a false positive outcome. Similarly, it will be specified as a true negative outcome or a false negative outcome if no alarm occurs for ADL or a fall activity. An ideal fall detection system should identify all simulated falls as positive outcomes and all ADLs as negative outcomes. Conducting the experiment would be too risky for elderly users. Thus, a group of 15 healthy subjects aged from 21 to 25 were asked to perform the simulated falls and ADLs (Figure 8). When subjects simulate falls, falling to the ground directly could possibly injure them. Thus, we placed an inflatable mat (thickness 15 cm) on the ground for cushion. Meanwhile, due to the good elasticity of the mat, the subjects would uncontrollably bounce up at a short distance when they fall to the mat. Therefore, the jitter of RMS and Euler angle of the impact are greater than that in the ideal situation.

Sensors 2017, 17, 2096 Sensors 2017, 17, 2096

9 of 15 9 of 15

Figure8.8.Conducting Conductingfalling fallingand andADL. ADL. Figure

Each performed all thethe 16easy activities and repeated them the sensor on In our subject experiment, considering portability, the sensor unitafter was placing respectively placed unit on the the shoulder, waist, andand foot, and activity each activity was Scotch repeated thrice for each subjects’ shoulder, waist, footrespectively, to conduct each repeatedly. tape was used to fixbody the segment. preparation stage, was set andhorizontally the sensor unit sensor unitInonthe different segments of the the experimental human body. environment In detail, we fixed theup, device on was charged to ensure that the battery was adequate during the experiment. Then, the sensor unit the shoulder (or the foot) by the tape and fixed the device in the coat pocket for the waist segment. wasEuler fixedangle to thedisplayed participant’s body segment, andthe thetheoretical APP wasvalue readybecause to be connected to the sensor The a small deviation from of the irregularities of unit. In the experiment’s conduction stage, the mat was fixed tightly by two persons to ensure the human body parts. safety, and the participant front of therepeated mat andthem thenafter performed simulated Each subject performed stood all thein 16the activities and placing all thethe sensor unit onfalls. the After thewaist, falls, and the participant performed all the ADLs. The resultsthrice during experiment were shoulder, foot, respectively, and each activity was repeated forthe each body segment. recorded by the APP automatically. In the preparation stage, the experimental environment was set up, and the sensor unit was charged We did data of the same activity sample onThen, the three humanunit segments dueto to to ensure thatnot the collect batterythe was adequate during the experiment. the sensor was fixed theparticipant’s following concerns: (1) the and sensor on the three human segments the body segment, theunits APP simultaneously was ready to beplaced connected to the sensor unit. In the during activities can easily havethe errors the issues and experiment’s conduction stage, matdue wastofixed tightlyofbystabilization two persons to inter-disturbance; ensure safety, and(2) theit is difficult stood to distinguish which sensor unit then issues the alarmallwhen more thanfalls. one After occur,the and we participant in the front of the mat and performed the simulated falls, need further analysis to verify it. ADLs. ThoughThe there may during be difference between the three repeats an the participant performed all the results the experiment were recorded byofthe activity on the three human segments, it does not influence the detection results since the different APP automatically. is slight when subject difference between due samples We did notacollect therepeats data ofthe thesame sameactivity. activity Furthermore, sample on thethis three human segments to theis allowed by our detection algorithm. following concerns: (1) the sensor units simultaneously placed on the three human segments during activities can easily have errors due to the issues of stabilization and inter-disturbance; (2) it is difficult 6. distinguish Experimental Results to which sensor unit issues the alarm when more than one occur, and we need further analysis to verify it. Though thereroll, mayand be difference three repeats of three an activity onbody the In the experiment, the RMS, pitch data between of all the the participants on the human three human segments, it does not influence the detection results since the different is slight when segments (shoulder, waist, and foot) were captured and recorded, respectively, during all the aspecified subject repeats the same activity. Furthermore, this difference allowedcorrectly by our simulated falls and ADLs. We found that all the fallbetween activitiessamples can be is detected detection when thealgorithm. sensor unit was placed on the three human body segments. However, some ADLs were detected incorrectly when the sensor unit was placed on the shoulder and foot segments. 6. Experimental Results In the experiment, the RMS, roll, and pitch data of all the participants on the three human body segments (shoulder, waist, and foot) were captured and recorded, respectively, during all the specified simulated falls and ADLs. We found that all the fall activities can be detected correctly when the sensor unit was placed on the three human body segments. However, some ADLs were detected incorrectly when the sensor unit was placed on the shoulder and foot segments. Figure 9 shows the RMS and absolute value(a)of roll and pitch distribution on the three body segments during typical fall activities, including the following: (a) backward fall ending up lying; (b) forward fall ending up lying; (c) lateral right fall ending up lying; and (d) lateral left fall ending up lying. The peaks of RMS distribution on all three body segments exceeds 2.3 g, and the peaks of absolute of roll or pitch exceeded 50◦ during the simulated falls. Thus, these activities were detected as falls correctly based on the threshold comparison. (b)

In the experiment, the RMS, roll, and pitch data of all the participants on the three human body segments (shoulder, waist, and foot) were captured and recorded, respectively, during all the specified simulated falls and ADLs. We found that all the fall activities can be detected correctly when the sensor unit was placed on the three human body segments. However, some ADLs were Sensors 2017, 17, 2096 10 of 15 detected incorrectly when the sensor unit was placed on the shoulder and foot segments. Sensors 2017, 17, 2096

10 of 15

(a) (c)

Sensors 2017, 17, 2096

(b)

10 of 15

(d) Figure 9. RMS and absolute value of roll and pitch distribution on the shoulder, waist, and foot during typical fall activities: (a) Backward fall ending up lying; (b) Forward fall ending up lying; (c) Lateral right fall ending up lying; (d) Lateral left fall ending up lying.

(c) Figure 9 shows the RMS and absolute value of roll and pitch distribution on the three body segments during typical fall activities, including the following: (a) backward fall ending up lying; (b) forward fall ending up lying; (c) lateral right fall ending up lying; and (d) lateral left fall ending up lying. The peaks of RMS distribution on all three body segments exceeds 2.3 g, and the peaks of absolute of roll or pitch exceeded 50° during the simulated falls. Thus, these activities were detected (d) as falls correctly based on the threshold comparison. Figure shows RMS and absolute value ofdistribution roll and on pitch distribution the three body Figure 9.10RMS andand absolute value of roll the on shoulder, waist,on and foot during Figure 9. RMS absolute value ofand rollpitch and pitch distribution the shoulder, waist, and foot segments during some ADLs: (a) bending over, (b) staggering forward, (c) walking, and (d) taking typical falltypical activities: Backward endingfall up lying; fallForward ending up (c)up Lateral during fall (a) activities: (a) fall Backward ending(b) upForward lying; (b) falllying; ending lying; stairs.right Figure 10a,b are detected as falls incorrectly on the shoulder segment because the shoulder fall ending up lying; (d) Lateral left fall ending up lying. (c) Lateral right fall ending up lying; (d) Lateral left fall ending up lying. sharply changes from near vertical to near horizontal status when the human body is bending over Figure 10 shows RMS absolute value ofvalue roll10c, and distribution on the three segments rapidly or staggering forward. In and addition, Figure are and detected falls incorrectly onthree the foot Figure 9 shows theand RMS absolute ofdpitch roll pitchasdistribution onbody the body during some ADLs: (a) bending over, (b) staggering forward, (c) walking, and (d) taking stairs.(b) segment because thetypical movement of the feet during the walking or taking stairs is large compared with segments during fall activities, including following: (a) backward fall ending up lying; Figure are detected assegments. falls(c) incorrectly on fall the shoulder because the shoulder sharply that of 10a,b other human body Theright peak of ending RMS and the peak of(d)absolute value pitchup forward fall ending up lying; lateral upsegment lying; and lateral left fall of ending changes from near vertical to near horizontal status when the human body is bending over rapidly exceed 2.3 g and 50°, respectively, on the shoulder and foot body segments. Meanwhile, in Figure lying. The peaks of RMS distribution on all three body segments exceeds 2.3 g, and the peaks of or absolute staggering forward. In addition, Figure are detectedthe as falls incorrectly on the2.3 foot segment 10a–d, during the same ADLs performed on10c,d other segments, RMS peaks areactivities below g and the of roll or pitch exceeded 50° during the simulated falls. Thus, these were detected because the movement of the feet during walking or taking stairs is large compared with that of other peaks of roll and pitch are below 50°. The ADLs were detected correctly. as falls correctly based on the threshold comparison. human Figure body segments. peak of absolute RMS and value the peak value of pitch exceed 2.3 three g and body 50◦ , 10 showsThe RMS and of of rollabsolute and pitch distribution on the respectively, on the shoulder and(a) foot body segments. Meanwhile, in Figure during segments during some ADLs: bending over, (b) staggering forward, (c)10a–d, walking, and the (d) same taking ADLs performed on other segments, the RMS peaks are below 2.3 g and the peaks of roll and pitch are stairs. Figure 10a,b are detected as falls incorrectly on the shoulder segment because the shoulder ◦ . The ADLs were detected correctly. below 50 sharply changes from near vertical to near horizontal status when the human body is bending over rapidly or staggering forward. In addition, Figure 10c, d are detected as falls incorrectly on the foot segment because the movement of the feet during walking or taking stairs is large compared with that of other human body segments. The peak of RMS and the peak of absolute value of pitch exceed 2.3 g and 50°, respectively, on the shoulder and foot body segments. Meanwhile, in Figure 10a–d, during the same ADLs performed on other segments, the RMS peaks are below 2.3 g and the peaks of roll and pitch are below 50°. The ADLs (a)were detected correctly.

(b)(a) Figure 10. Cont.

(b)

Sensors 2017, 17, 2096 Sensors 2017, 17, 2096

of 15 11 11 of 15

(c)

(d) Figure 10. RMS absolute roll anddistribution pitch distribution on the shoulder, waist, and foot Figure 10. RMS and and absolute valuevalue of rollof and pitch on the shoulder, waist, and foot during (c) Walking ; (d) Taking stairs . during ADLs: (a) Bending over; (b) Staggering forward; ADLs: (a) Bending over; (b) Staggering forward; (c) Walking; (d) Taking stairs.

correct detection of these 16 activities including fallsADL andon ADL thehuman three body human TheThe correct detection rate rate of these 16 activities including falls and the on three body segments is summarized in Table 2. This rate is obtained through dividing the number segments is summarized in Table 2. This rate is obtained through dividing the number of corrected of corrected by the total detection for an activitybyimplemented by As all mentioned the subjects. detections by detections the total detection samples for an samples activity implemented all the subjects. As mentioned in Section positive or true isnegative detection into account as one in Section 5, the true positive5,orthe truetrue negative detection taken into accountisastaken one result of corrected result of corrected detection. The correct detection rate for all the fall activities is 100%, whereas detection. The correct detection rate for all the fall activities is 100%, whereas the detection correct ratethe detection rate in some ADL cases are variable, rate for bending over in some ADLcorrect cases are variable, namely, 60% correct rate fornamely, bending60% overcorrect and staggering forward on staggering forward on the shoulder segment. In 0% addition, rate and correct rate theand shoulder segment. In addition, 6.7% correct rate and correct6.7% rate correct are observed for 0% walking and are observed for walking and taking stairs, respectively, on the foot segment. taking stairs, respectively, on the foot segment. Table 2. Detection accuracy of activities 16 activities three human body segments in the experiment. Table 2. Detection accuracy of 16 on on thethe three human body segments in the experiment. 16 Activities 16 Activities

Shoulder Waist Waist Correct Correct Correct Rate (%) Correct Rate (%) Rate (%) Rate (%) Shoulder

Backward fall Backward Ending up lying fall Ending up falling to the ground with elbows Ending up lying Ending up siting after lying for 2 s Forward fallto the ground with Ending up falling Falls activities Ending up lying Falling on the kneeselbows ending up lying left fallafter lying for 2s EndingLateral up siting Ending up lying Forward fall Falls activities Lateral right fall Ending up lying Ending up lying

100 100 100 100 100 100 100

100 100

Foot Correct Correct Rate (%) Rate (%) Foot

100 100 100 100

100 100 100 100

100100 100

100 100 100

100

100

100

100

100 100100

100 100 100

Falling on Walking the knees ending up lying 100 100 100 100100 6.7 Bending over 60 100 100 Lateral left fall Taking stairs 100 100 100100 0 100 Ending up lying Standing-sitting on a chair-standing 100 100 100 Standing-sitting and right lying on a ADL activities Lateral fall 100 100 100 100 100 100 mattress-standing Ending up lying Stretching shoulder 100 100 100 StaggeringWalking forward 60 100100 100 100 6.7 Staggering backward 100 100 100 Bending over 60 100 100 Twisting waist 100 100 100 Taking stairs 100 100 0 Standing-sitting on a chair-standing 100 100 100 We can further analyze the detection performance of all activities (accuracy), simulated falls Standing-sitting and lying on a 100sensor ADL activities (sensitivity), and the ADLs (specificity) to demonstrate the best 100 body location100 for placing the mattress-standing unit. The accuracy means the total correct detection rate of all the 16 activities on a human segment, Stretching shoulder 100 100 100 in which the sensitivity is the detection correct rate of the 7 simulated falls activities, and the specificity Staggering forward 60 100 100 is the detection correct rate ofStaggering the 9 ADLs. As shown in Figure 11, the sensitivity method backward 100 100of detection100 is 100% for all three body segments, whereas the specificity of detection method for shoulder, Twisting waist 100 100 100waist, and foot are 91.1%, 100%, and 78.5%, respectively. These values mean that our system has high accuracy We can further the detection performance all activities (accuracy), simulated falls of fall detection, and theanalyze waist segment is the best location toofplace the sensor unit, wherein both the (sensitivity), and the ADLs (specificity) to demonstrate the best body location for placing the sensor

Sensors 2017, 17, 2096 Sensors 2017, 17, 2096

12 of 15 12 of 15

detection sensitivity can reach 100%. and During the experiment, all100%. the subjects sensor unit, wherein and bothspecificity the detection sensitivity specificity can reach Duringreach the an agreement that the waist is also the most comfortable location to place the sensor unit, which is experiment, all the subjects reach an agreement that the waist is also the most comfortable location simply in the jacket pocket. to place placed the sensor unit, which is simply placed in the jacket pocket.

Falls detection performance on different human body segments 120 100

95

100

100

100

100

87.9

91.1

100 78.5

80 60 40 20 0 Accuracy(%)

Sensitivity(%) Shoulder

Waist

Specificity(%) Foot

Figure Figure11. 11.Fall Falldetection detectionperformance performanceon ondifferent differenthuman humanbody body segments. segments.

7. 7.Discussion Discussionand andConclusions Conclusions In the literature, literature,there there already for human fall detection. Totheaddress the In the areare already manymany worksworks for human fall detection. To address advantages advantages of our method, we make a comparison with some previous methods in terms of sensor of our method, we make a comparison with some previous methods in terms of sensor location, location, portability, and comfort. Though the experiment of these methods are accuracy,accuracy, portability, and comfort. Though the experiment protocolsprotocols of these methods are different different and the datasets for detection are based on their experiments, all of them considered the and the datasets for detection are based on their experiments, all of them considered the distinction distinction between falls simulated falls In and He et[26], al.’sthe method [26], the and between simulated and ADLs. HeADLs. et al.’s In method acceleration andacceleration angular velocity angular databyare measured by fixed a sensor board the vest andneck closed the neck and data arevelocity measured a sensor board in the vestfixed and in closed to the andtotransmitted to transmitted to a smart phone. They detect falls based on a Bayes network classifier with an accuracy a smart phone. They detect falls based on a Bayes network classifier with an accuracy of 95.67%. of InetPierleoni et al.’s method [24], thedevice wearable device placed waist abybelt, wearing a In 95.67%. Pierleoni al.’s method [24], the wearable is placed onisthe waiston bythe wearing and the belt, and the fall detection based oncomparison the threshold comparisonand of Euler acceleration and Euler angle, fall detection is based on theisthreshold of acceleration angle, which is similar to which is similar to our method. However, our method improves the detection accuracy on the waist our method. However, our method improves the detection accuracy on the waist segment to 100%, segment to 100%, benefitting from the threshold of RMS re-identified our experiment. Erdogan benefitting from the threshold of RMS re-identified in our experiment.inIn Erdogan et al.’sInstudy [29], et al.’s study [29], the data were also measured by wearing a belt on the waist, but a computer is the data were also measured by wearing a belt on the waist, but a computer is needed to receive needed to receive and process the data. They used a k-nearest neighbor algorithm to detect falls and process the data. They used a k-nearest neighbor algorithm to detect falls based on acceleration based at an 89.4%. In Ojetola et al.’s method [31], the acceleration data atonanacceleration accuracy ofdata 89.4%. Inaccuracy Ojetola etofal.’s method [31], the acceleration and angular velocity and angular velocity data are also measured by wearing a belt on both the chest and right thigh and data are also measured by wearing a belt on both the chest and right thigh and are then transmitted are then transmitted to a computer. The detection accuracy range is from 98.91% to 99.45%, which is to a computer. The detection accuracy range is from 98.91% to 99.45%, which is dependent on the dependent the training of a decision machinetree. learning decision tree.method In Sorvala al.’s method [30], the training of on a machine learning In Sorvala et al.’s [30],et the data is measured by data is measured by wearing a wireless inertial measurement unit (IMU) on the waist and an wearing a wireless inertial measurement unit (IMU) on the waist and an accelerometer on an ankle. accelerometer on an ankle. The with detection 98.69% 95.6% Yu sensitivity and 99.6% The detection accuracy is 98.69% 95.6% accuracy sensitivityisand 99.6%with specificity. et al.’s system [33] is specificity. Yu et al.’s system [33] is vision-based by using a camera and a personal computer, vision-based by using a camera and a personal computer, indicating that the fall detection system is indicating thatvision the field fall detection system is limited in is thenotvision of the camera the limited in the of the camera and the portability good. field The detection accuracyand of their portability is not good. accuracy of their methodaccuracy is 98.13%. Thus, compared our method hasother the method is 98.13%. Thus,The ourdetection method has the highest detection of 100% with highest detection accuracy of 100% compared with other methods. Meanwhile, our system has methods. Meanwhile, our system has good portability because the small packaged sensor unit is easily good because the the small packaged sensor is easilyaccelerometer fixed in a pocket of a top.methods For the fixed portability in a pocket of a top. For wearing comfort, all unit the existing sensor-based wearing all the existing accelerometer sensor-based methods including [24,26] and or [29–31] includingcomfort, [24,26] and [29–31] depend on the wearing time because long wearing time of a vest a belt depend on the wearing time because long wearing time of a vest or a belt on the waist, thigh, and on the waist, thigh, and ankle will lead to discomfort for the user, whereas our method is comfortable ankle will lead to discomfort for the user, whereas our method is comfortable to the user, even to the user, even when porting the sensor in the pocket for a long time. when porting the sensor in the pocket for a long time. Our method proposes an easy-to-use, portable human fall monitoring system using only a small self-packaged sensor unit and a mobile phone. We designed 16 simulated falls and ADLs and

Sensors 2017, 17, 2096

13 of 15

Our method proposes an easy-to-use, portable human fall monitoring system using only a small self-packaged sensor unit and a mobile phone. We designed 16 simulated falls and ADLs and conducted experiments by placing the sensor units on different body segments, such as the shoulder, waist, and foot of participants. Although fall activities can be correctly detected in all the three body segments, ADL is still possibly mistaken as falls when the sensor unit is placed on the shoulder and foot, but not when placed on the waist. The monitoring system can achieve 100% accuracy, sensitivity, and specificity on the waist segment, verifying that the waist is the best location for placing the sensor. The sensor unit can be placed in the pocket of a top at waist-height without needing to be fixed on the skin. Furthermore, our system is highly portable and comfortable because the sensor unit can be fixed in the pocket of a top and works only with a mobile phone, which offers great portability and will not lead to discomfort during usage. Although we designed the experiment based on the reality of the elderly fall in daily life, the subjects in the experiment are young people to avoid any potential risk of injury. We still need to consider the possible differences in fall activities between the elderly people and young people. Evaluation and validation of the practical application of the proposed system for elderly people is our next objective. Acknowledgments: The work of this paper is financially supported by The National Natural Science Foundations of China (No. 61502112), NSF of Guangdong Province (No. 2016A030313499), The Science and Technology Planning Project of Guangdong Province (No. 2015A030401030, No.2015A020219014), Guangzhou Municipal University Research Projects (No. 1201630279), and the Fundamental Research Funds for the Central Universities (No. 2017ZD054), The Guangdong Province Universities and Colleges Excellent Youth Teacher Training Program and the Student Research Project (SRP) of South China University of Technology. Author Contributions: Xuedong Ma contributes on realizing the fall monitor system, collecting and interpreting the experimental data; Aihua Mao contributes on conception and design of the work, writing and revising the paper; Yinan He contributes on design and conducting the experiment; Jie Luo contributes on conception and design of the work, writing and revising the paper. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3.

4. 5. 6. 7. 8.

9. 10.

World Health Organization (WHO). Available online: http://www.who.int/mediacentre/factsheets/fs344/ en/ (accessed on 12 September 2017). Heinrich, S.; Rapp, K.; Rissmann, U.; Becker, C.; König, H.H. Cost of falls in old age: A systematic review. Osteoporos. Int. 2009, 21, 891–902. [CrossRef] [PubMed] Tran, T.H.; Le, T.L.; Hoang, V.N.; Vu, H. Continuous detection of human fall using multimodal features from Kinect sensors in scalable environment. Comput. Methods Progr. Biomed. 2017, 146, 151–165. [CrossRef] [PubMed] Bian, Z.P.; Hou, J.; Chau, L.P.; Magnenat-Thalmann, N. Fall detection based on body part tracking using a depth camera. IEEE J. Biomed. Health Inform. 2015, 19, 430–439. [CrossRef] [PubMed] Tao, S.; Kudo, M.; Nonaka, H. Privacy-preserved behavior analysis and fall detection by an infrared ceiling sensor network. Sensors 2012, 12, 16920–16936. [CrossRef] [PubMed] Stone, E.E.; Skubic, M. Fall detection in homes of older adults using the Microsoft Kinect. IEEE J. Biomed. Health Inform. 2015, 19, 290–301. [CrossRef] [PubMed] Liu, S.H.; Cheng, W.C. Fall detection with the support vector machine during scripted and continuous unscripted activities. Sensors 2012, 12, 12301–12316. [CrossRef] [PubMed] Karantonis, D.M.; Narayanan, M.R.; Mathie, M.; Lovell, N.H.; Celler, B.G. Implementation of a real-time human movement classifier using a triaxial accelerometer for ambulatory monitoring. IEEE Trans. Inform. Technol. Biomed. 2006, 10, 156–167. [CrossRef] Lai, C.F.; Chang, S.Y.; Chao, H.C.; Huang, Y.M. Detection of cognitive injured body region using multiple triaxial accelerometers for elderly falling. IEEE Sens. J. 2011, 11, 763–770. [CrossRef] Kangas, M.; Konttila, A.; Winblad, I.; Jamsa, T. Determination of Simple Thresholds for Accelerometry-Based Parameters for Fall Detection. In Proceedings of the EMBS 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Lyon, France, 22–26 August 2007; pp. 1367–1370.

Sensors 2017, 17, 2096

11. 12.

13.

14. 15. 16. 17. 18.

19.

20. 21. 22.

23. 24. 25.

26. 27.

28.

29. 30.

14 of 15

Luque, R.; Casilari, E.; Morón, M.J.; Redondo, G. Comparison and characterization of android-based fall detection systems. Sensors 2014, 14, 18543–18574. [CrossRef] [PubMed] Li, Q.; Stankovic, J.A.; Hanson, M.A.; Barth, A.T.; Lach, J.; Zhou, G. Accurate, Fast Fall Detection Using Gyroscopes and Accelerometer-Derived Posture Information. In Proceedings of the BSN 2009 Sixth International Workshop on Wearable and Implantable Body Sensor Networks, Berkeley, CA, USA, 3–5 June 2009; pp. 138–143. Nari, M.I.; Suprapto, S.S.; Kusumah, I.H.; Adiprawita, W. A Simple Design of Wearable Device for Fall Detection with Accelerometer and Gyroscope. In Proceedings of the 2016 International Symposium on Electronics and Smart Devices (ISESD), Bandung, Indonesia, 29–30 November 2016; pp. 88–91. Andò, B.; Baglio, S.; Lombardo, C.O.; Marletta, V. A multisensor data-fusion approach for ADL and fall classification. IEEE Trans. Instrum. Meas. 2016, 65, 1960–1967. [CrossRef] Yang, L.; Ren, Y.; Hu, H.; Tian, B. New fast fall detection method based on spatio-temporal context tracking of head by using depth images. Sensors 2015, 15, 23004–23019. [CrossRef] [PubMed] Gasparrini, S.; Cippitelli, E.; Spinsante, S.; Gambi, E. A depth-based fall detection system using a Kinect®sensor. Sensors 2014, 14, 2756–2775. [CrossRef] [PubMed] Lee, Y.S.; Chung, W.Y. Visual sensor based abnormal event detection with moving shadow removal in home healthcare applications. Sensors 2012, 12, 573–584. [CrossRef] [PubMed] Anderson, D.; Keller, J.M.; Skubic, M.; Chen, X.; He, Z. Recognizing Falls from Silhouettes. In Proceedings of the 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY, USA, 30 August–3 September 2006; pp. 6388–6391. Rougier, C.; Meunier, J.; St-Arnaud, A.; Rousseau, J. Monocular 3D Head Tracking to Detect Falls of Elderly People. In Proceedings of the 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, New York, NY, USA, 30 August–3 September 2006; pp. 6384–6387. Munoz-Organero, M.; Lotfi, A. Human movement recognition based on the stochastic characterization of acceleration data. Sensors 2016, 16, 1464. [CrossRef] [PubMed] Hsieh, C.Y.; Liu, K.C.; Huang, C.N.; Chu, W.C.; Chan, C.T. Novel hierarchical fall detection algorithm using a multiphase fall model. Sensors 2017, 17, 307. [CrossRef] [PubMed] Kurniawan, A.; Hermawan, A.R.; Purnama, I.K. E. A Wearable Device for Fall Detection Elderly People Using Tri Dimensional Accelerometer. In Proceedings of the 2016 International Seminar on Intelligent Technology and Its Applications (ISITIA), Lombok, Indonesia, 28–30 July 2016; pp. 671–674. Albert, M.V.; Kording, K.; Herrmann, M.; Jayaraman, A. Fall classification by machine learning using mobile phones. PLoS ONE 2012, 7, e36556. [CrossRef] [PubMed] Pierleoni, P.; Belli, A.; Palma, L.; Pellegrini, M.; Pernini, L.; Valenti, S. A high reliability wearable device for elderly fall detection. IEEE Sens. J. 2015, 15, 4544–4553. [CrossRef] Wibisono, W.; Arifin, D.N.; Pratomo, B.A.; Ahmad, T.; Ijtihadie, R.M. Falls Detection and Notification System Using Tri-axial Accelerometer and Gyroscope Sensors of a Smartphone. In Proceedings of the 2013 Conference on Technologies and Applications of Artificial Intelligence (TAAI), Taipei, Taiwan, 6–8 December 2014; pp. 382–385. He, J.; Bai, S.; Wang, X. An unobtrusive fall detection and alerting system based on Kalman filter and Bayes network classifier. Sensors 2017, 17, 1393. [CrossRef] [PubMed] Ntalampiras, S.; Roveri, M. An Incremental Learning Mechanism for Human Activity Recognition. In Proceedings of the 2016 IEEE Symposium Series on Computational Intelligence (SSCI), Athens, Greece, 6–9 December 2016; pp. 1–6. Yean, S.; Lee, B.S.; Yeo, C.K.; Vun, C.H. Algorithm for 3D Orientation Estimation Based on Kalman Filter and Gradient Descent. In Proceedings of the 2016 IEEE 7th Annual Information Technology on Electronics and Mobile Communication Conference (IEMCON), Vancouver, BC, Canada, 13–15 October 2016; pp. 1–6. Erdogan, S.Z.; Bilgin, T.T. A data mining approach for fall detection by using k-nearest neighbour algorithm on wireless sensor network data. IET Commun. 2012, 6, 3281–3287. [CrossRef] Sorvala, A.; Alasaarela, E.; Sorvoja, H.; Myllyla, R. A Two-Threshold Fall Detection Algorithm for Reducing False Alarms. In Proceedings of the 6th International Symposium on Medical Information and Communication Technology (ISMICT), La Jolla, CA, USA, 25–29 March 2012; pp. 1–4.

Sensors 2017, 17, 2096

31.

32. 33.

15 of 15

Ojetola, O.; Gaura, E.I.; Brusey, J. Fall Detection with Wearable Sensors-SAFE (Smart Fall Detection). In Proceedings of the 7th International Conference on Intelligent Environments, Nottingham, UK, 25–28 July 2011; pp. 318–321. Kangas, M.; Konttila, A.; Lindgren, P.; Winblad, I.; Jämsä, T. Comparison of low-complexity fall detection algorithms for body attached accelerometers. Gait Posture 2008, 28, 285–291. [CrossRef] [PubMed] Yu, M.; Rhuma, A.; Naqvi, S.M.; Wang, L.; Chambers, J. A posture recognition-based fall detection system for monitoring an elderly person in a smart home environment. IEEE Trans. Inf. Technol. Biomed. 2012, 16, 1274–1286. [PubMed] © 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).