Keywords

1 Introduction

It is the worldwide trend that more and more people live alone at home, specific for old people. For those people who live alone, sometime they may meet dangerous things happened to them at home, like fall or sick and outside people does not know that to prove help in time. According to World Health Organization (WHO) [1]” data shows that Falls are the second leading cause of accidental or unintentional injury deaths worldwide.” When fall accident happened, outside people may not know and cannot provide help in real time. It is high possibility that to cause serious hurt to the people and let them cannot move to use phone to call outside for help. Like broken their bones, brain or spinal trauma, and internal bleeding [2], it is serious to impact old people health and life. According to the accident data from Taiwan related organization report out, fall is the second high accident in Taiwan. From Taiwan Patient safety Reporting system [3], Medication errors are the most frequent event reported to TPR (Taiwan Patient-safety Reporting system) in 2016 (20,245 reported), and second are falls (16,635), as shown in Fig. 1.

Fig. 1.
figure 1

The types of TPR Reporting Events in 2016 [3]

Full size image

To protect the people from fall accident and get help in time to save their life or mitigate the damage. Many researches use different methods and technologies for fall detection [4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19]. Fall detection can be grouped into two categories. That is environmental detection system [4,5,6,7,8,9] and wearable sensor detection system [11,12,13,14,15,16,17,18,19]:

  1. 1.

    Environmental detection system

    This method is placed in the predefined environmental to monitor people activity without wear any type of sensors [4, 5]. Such as image recognition (camera) [6, 7], floor pressure sensor [8, 9], etc. The system can setup in users’ home in advance to detect fall accident. The main advantage is that user may not feel uncomfortable since they do not need to wear any sensors with them [10]. But environment detection system needs to design for specific home users with different angels to cover everywhere in the home at different rooms and it has private problem for user [6]. This system only can protect user in predefined space, like home.

  2. 2.

    Wearable sensor detection system

    This method just simple to us use sensors like microphone [11], barometric pressure sensor [12], accelerometer sensor [12,13,14,15,16,17], gyroscope sensor [17], electronic compass etc. Those sensors can be made very small and easy to hand carry by users. Those wearable devices are low power can continue working for whole day [18]. It is suitable for a mobile detection system. It can be used in various environment. The devices can be placed on leg, hand, chest and waist. But wearing those devices may cause user feel uncomfortable and inconvenience [10].

Smart phone is a mobile device and almost everyone has it now, including old people, it is easy to carry and use [18, 19]. More and more people use smart phone to make phone call or access internet to get the information they need. The mobile phone cost down to everyone can offer to buy it [20]. With the quick development smart phone has several sensors in smart phone, such as the tri-axial accelerometer, electronic compass, global positioning system (GPS), etc. [13,14,15,16,17,18,19,20]. With those sensors, we can analysis the motion signal from human. In this paper, we will use ASUS ZenFone 2 Laser (ZE50kl) mobile phone device, Android 6.0.1 OS. This paper will analysis of five different normal movements [14] and fall event. We purpose a fall detection system based on smart phone angle and acceleration features.

2 Motion Signal Process and Feature Select

In this Chapter, we will introduce the motion signal process and features selection. From Fig. 2, it lists six activities including fall, sit down, jump, lying down, walk and run. When people carry smart phone and enable the detection system, it will record every movement to identify the activity in line with which motion signal in the database. From the Fig. 2, we can see the fall feature is very different from other normal movements. We also can identify that if the same motion signal pattern repeated in short time, then system will identify it is strenuous exercise instead of fall. In order to identify clearly if it is fall event or not, after system recording the motion signals and then it will be compared with database to choose which signal feature is in line with signals system recorded to decide which activity happens.

Fig. 2.
figure 2

Six different kinds of activities. (a) Fall. (b) Sit. (c) Jump. (d) Lying down. (e) Walk. (f) Run.

Full size image

2.1 Tri-Axial Accelerometer

Accelerometer (G-Sensor), also called gravity sensor, it can do the acceleration detection from any directions. The expression is the axial acceleration magnitude and direction (XYZ). It can detect the forces from both of the gravity and the forces when moving the device. The center of gravity is the balance point for human body. When human body is moving, the center of gravity is near the waist. That’s why we will suggest the detection system (smart phone) should be placed in the waist pocket instead of hand or leg. The data signals is a three-dimensional signal X-axis, Y-axis, and Z-axis (as in Fig. 3). The sampled signal is [ax [n], ay [n], az [n]]. To predigest the three-dimensional signal, we will use the one-dimensional signal magnitude vector (SMV) [14] by

$$ \text{SMV} = \sqrt {{\text{a}}_{\text{x}}^{2} \left[ {\text{n}} \right] + {\text{a}}_{\text{y}}^{2} \left[ {\text{n}} \right] + {\text{a}}_{\text{z}}^{2} \left[ {\text{n}} \right]} , $$
(1)

to find out the fall feature. We will observe the fall event motion signal. In the Fig. 2(a), we can see the motion signal at 55th date point, the SMV value is smaller than 1 g. At this time the user body is in a state of weightlessness before landing. In this time period SMV is close to 0.5G. We observe other fall events and find the same characteristic. So we define when SMV smaller than 0.6 g is the weightlessness, and it is the detection system first feature.

Fig. 3.
figure 3

Tri-axial Accelerometer of Smart phone [21]

Full size image

2.2 Weighted Moving Average

Now we have the first feature ready to make sure the detection system is more accurate. For next step, we need to find out the second feature. We notice that the fall event happens after weightlessness. The peak values will appear in motion signal after the body is impacted by ground. But it does have problems exist potentially to confuse system to judge if it is fall or not, like the peak values are very often to appear when the body is doing intense exercise to cause the motion signals will be very messy. It increases the burden of detection system operations and increases the possibility of misjudgment. To offload detection system operation burden and increase the efficiency. We will use weighted moving average (WMA) to process the motion signal,

$$ {\text{WMA}}_{{{\text{t}}1,{\text{n}}}} = {\text{WMA}}_{{{\text{t}}0,{\text{n}}}} - \frac{{{\text{Gain*SMV}}_{1} }}{\text{n}} + \frac{{{\text{Gain*SMV}}_{{{\text{n}} + 1}} }}{\text{n}} $$
(2)
$$ \begin{aligned} {\text{if SMV}} > 2.0\,{\text{g}}\; then\;Gain\; = {\text{Impact}} - {\text{weightlessness}} \hfill \\ \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{\text{else if SMV}} < 2.0\,{\text{g Gain}} = 1. \hfill \\ \end{aligned} $$

Using the moving average process, it can make the motion signal more accurately and can exclude small motion signal noise. But the weakness is that the motion signal peak values can be distortion and lose the fall feature what we are looking for. To retain fall peak values feature, we will give a weighted value, if the SMV value is more than 2.0 g, to give a gain value by impact subtract the weightlessness. The result shows in Fig. 4(b). It just retains fall peak values and make the motion signals cleaner, and then we can compare that with normal movements in Fig. 4(b) and (d). We can define that when the SMV after WMA more than 2.5 g means that it is impacted. And this will be the second feature for the detection system.

Fig. 4.
figure 4

Comparison motion signal after weighted moving average. (a) Fall original data. (b) Fall process data. (c) Sit original data. (d) Sit process data.

Full size image

2.3 Electronic Compass

After getting above two features, we notice that the fall event after weightlessness and impact. The motion signal will appear horizontal stationary state for a period of time. But this feature has some problems by only using the accelerometer. Detection system can’t identify fall or general activity static like jump and run. So we add angle recognition to assist detection system. Smart phone’s angle detection uses the electronic compass, we will combine both motion signal and angle information together to detect if is body stillness after the fall event happened. If the motion signal SMV standard deviation equals to 0.1 for 2.5 s after first and second feature and the phone angle smaller than 30°, we can see the SMV and angle comparison in Fig. 5(a) (b). We can use it to identify that the body is at the stillness. And this will be 3rd feature from the detection system.

Fig. 5.
figure 5

Comparisons fall event between SMV and angle. (a) Fall event SMV. (b) Fall event angle.

Full size image

3 Fall Detection Algorithm and Result

We have defined three features to identify fall event versus normal movements. Figure 6 shows the detection flow chart for three features weightlessness, impact and stillness. The first step is to make sure the user body is under weightlessness by detecting SMV value smaller than 0.6 g or not. The second step is to determine the impact on the floor after step 1, it is judged by the peak of the SMV after motion signal data processing the value is over the 2.5 g or not. If the peak over default threshold, the system will identify it is an impact. The third step is to use accelerometer and angle recognition to detect if it is at horizontal stationary state or not after two features were met the conditions. During this step, it also will distinguish the fall event from strenuous exercise. After body impacted within 2.5 s sit will calculate motion signal standard deviation of variation. If standard deviation is smaller than 0.1 and angle recognition is smaller than 30°, the system will define it is stillness. When the three features are all match, the detection system will determine it is a fall event. This algorithm will discriminate fall event from normal movements, including misjudge sitting and lying down. This algorithm can detect fall event effectively. Figure 7 as show the different kinds of test results.

Fig. 6.
figure 6

Fall detection system flowchart.

Full size image
Fig. 7.
figure 7

Fall detect results. (a) Fall. (b) Left fall. (c) Right fall. (d) Sit. (e) Jump. (f) Lying down. (g) Walk. (h) Run.

Full size image

4 Conclusions

In this paper, we have proposed a fall detection based on smart phone, by installing the system software without adding or modifying hardware. This fall detection algorithm has been implemented to identify three features. Our testing results have shown that this detection system can distinguish people daily living activities and successfully detected three different direction fall events for the forward and lateral (right and left). This design can be used in various environments including indoor and outdoor. The acceptance from general people to use this system should be very high due to no extra cost, easy to handy and easy to use it. For future works, we want to collect more human motion signals from real environment and link the data with machine learning to improve the accuracy of the fall detection system.