Feature Request: Incident Detection Sensor rationality to avoid false alarms

It happened to be again yesterday. I was riding on a residential street and came to a four way stop intersection. A car was approaching on the cross street and I braked harder than usual to stop and give them the right-of-way. I couldn't have been traveling at more than 10 -12 mph at the time. It was a normal routine event, not a "holy sh*t" moment.  But my Instinct 2 thought I had an incident and stopped recording my ride and when into incident mode. I wasn't aware of it and two blocks later my phone sounded the alarm that my wife was contacted.

Seriously, it would be real easy to avoid false alarms by performing a simple 'sensor rationality' test. This is very common in robotics and the auto industry. E.g., I continued to pedal (cadence sensor) and travel at a moderate speed (speed sensor and GPS). My Instinct 2 should have been able to infer that no incident had taken place by looking at the data from the other sensors that indicated I was continuing my ride.

Yes, I get it, you don't want to miss an actual incident (True Positive) and overlook it (False Negative). But if your False Positives are so frequent that people disable the feature because they are annoyed by it (like the little boy who cried Wolf!) who benefits?

  • I continued to pedal (cadence sensor) and travel at a moderate speed (speed sensor and GPS).

    Well, you can fall, and the wheel of the bike may continue to turn for some time. If it prevented sending the alert, it  would not be OK. Similarly with GPS - imagine you fall on a slope, and continue sliding for some time. Or the GPS continues to report motion due to very common GPS drift (especially when the signal is weak). The algorithm certainly can be done quite a bit smarter, but on the other hand, the more complicated, the higher chance it won't trigger any alert when really needed.

    Still, feel free to suggest the improvements to Garmin at Share Ideas | Garmin 

  • Why didn't you cancel de incident detection when you got it?

  • I get a lot of notifications on my watch. Text messages and phone calls cause a notification, every 5 miles I get a pace notification, and if I'm using the training feature I get constantly notified if my power is below or above the target range plus a notification when it's back in range in the target range. After a while they get harder and harder to notice. 

  • Ok that sounds like user error in part then, if you have too many notifications and you can afford to stop paying attention to them you should just mute them to leave room for the important ones I guess. In any case, the incident notification has a distinct vibe/tone iirc

  • I'm a former embedded controls engineer and hghly reliable safety critical systems is one of my fortes, including sensor rationality and AI.  In systems like these using sensor rationality INCREASES system accuracy and reliability.  A simple system that relies on a single accelerometer is not robust and is as likely to fall prey to False Negatives as False Positives. Rationalizing the accelerometer data with additional sensor data makes for better True Positive detection rates.

  • Would be nice to be able to adjust the sensitivity I guess but there's not much else to add in terms of sensors or logic with the current hardware. We're talking a simple consumer wrist based device, not some critical device in an expensive system

  • I disagree with the characterization as 'user error' since what I'm suggesting in an improvement to the design system that would help reduce False Positive events and, with thoughtful design and implementation, improve True Positive event rates. 

    Mine only vibrates until the phone responds with the siren. While riding, especially in cold weather with long sleeves and gloves like the other day, it's easy to miss the vibrations.

  • My watch is connected to my speed, cadence, heart rate, and power sensors already, plus it has its own GPS receiver. There's nothing to add to the current hardware and an algorithm including sensor rationality would be simple to implement.  Does everyone use all the sensors I do? No, but an algorithm could be developed to use whatever additional sensor data is available to it, and when no additional data is available it falls back to it's simplest operating mode. This is pretty typical because any sensor in a multi-sensor system can be subject to  failure (including the accelerometer) and it's not unusual to use other sensors to diagnose a third sensor and determine when a fault code or remedial action is appropriate.