How does Fall detection work in a Fitness watch?
Introduction
Countless people’s lives have been saved due to Fall Detection algorithms in their fitness watches. Well, here is how it works! The following is the algorithm patented by Apple Watch, but it could vary from model to model.
Conditions to start detecting a fall
The Fall detection algorithm is triggered if the acceleration and jerk values exceed certain thresholds. The device then collects gyroscopic measurements, elevation or altitude measurements, and heart rate information, to determine whether a fall has occurred.
An accelerometer is used to detect hard impacts in real-time. They are detected based on the magnitude of acceleration and jerk (rate of change of acceleration) measured by the accelerometer over a sliding window.
The magnitude of the acceleration is calculated using:
Acceleration = sqrt ( max(|AxAx|) + max(|AyAy|) + max(|Az*Az|) )
Ax, Ay, and Az are the x, y, and z components of the acceleration signal respectively, and the maximum value is taken in a certain time window, 0.2 seconds for example.
The magnitude of the jerk is calculated using:
Jerk = max ( sqrt ( dx*dx + dy*dy + dz*dz ) )
dx, dy, and dz are the derivatives of the acceleration signals x, y, and z components in the same time window.
Fall Classifier — Is it Likely a Fall?
The first step in the algorithm is for the Fitness Watch to find out if it could be a fall before getting help from other sensors to confirm it.
- The accelerometer and gyroscope data are used in conjunction by sensor fusion and a feature extraction module uses this data to determine the wrist angle of the user. (Example — By determining the pose angle of the fitness watch).
- Also, the behavioral information of the user is determined using this data, such as whether the user is performing a bracing motion (eg — Thrusting their arms out to arrest forward momentum), a balancing motion (eg: Throwing their arms out to regain balance), a flailing motion (eg: Fluttering their arms during and after an impact), etc. Bracing motion, for example, can be detected based on features such as the wrist traversing negative arc length before an impact, and the wrist pointing toward the ground at the moment of impact.
- The acceleration data is also used to determine the impact information, such as the magnitude of an impact experienced by the user, and motions made by the user after the impact. This is used to determine if the user actually experienced an impact.
- Based on all this, a module classifies if the user has slipped, tripped, rolled, or experienced some other type of fall. This also determines if the user has had an impact but has not fallen. But for the algorithm to confirm that it is a fall, it needs help from other sensors to go through the last set of steps.
Is it a Confirmed Fall?
All the keywords highlighted in bold are blocks represented in Figure 1.
- Figure 1 shows a Fall classifier block. This is basically what we saw in the earlier section that uses accelerometer and gyroscope data to understand the user’s actions.
- Location-based Threshold — By collecting the GPS, and WiFi information, the fitness watch can detect if the user is driving a car, biking, at their home, etc. The likelihood of falling in our home is more than in our car. Depending on this information, the Fall classifier increases the sensitivity of its fall detection if the person is at home, for example. Based on this data and from Step 1, the Fall Classifier decides if it was likely a fall or not. Finally, this information is then sent out to the Sensor Fusion module.
- Altimeter sensor — The altitude data from the altimeter is analyzed to find out if the user is experiencing a significant elevation change (For example- Falling several feet, stories, etc.). If so, the Sensor Fusion may determine that a false positive is less likely (A false positive in this scenario means detecting that a person fell when the user actually didn’t)
- Heart Rate sensor — For example, if the user’s heart rate increases, the Sensor fusion module may determine that a false positive was less likely. This is because our heart rate tends to increase during and right after the fall.
- These are processed by the Sensor fusion model to confirm if the user actually fell, thereby confirming if it was a fall or not. Sensor fusion is combining data from multiple sensors to provide a more accurate and dependable understanding of the system. While an individual sensor may provide useful data on its own, the data that could be extracted from combining output from multiple sensors at once would give us a much better model. To fully understand the sensor fusion, spend 5 minutes reading my newsletter issue on the same here.
Conclusion
If it is confirmed to be a fall, an emergency call can be set up to alert 911 ambulance and a family member or friend. Thereby, probably saving the user.
This is how the Fall Detection algorithm works!
More From Me!
Subscribe to my email newsletter One Gadget at a Time where I explain how a gadget works on Wednesdays, no matter your background. I bet this will be your favorite newsletter :)
