Fall and man-down detection

Many of the tags people already wear in Omaya have a small motion sensor inside. Omaya can use it to notice when the person wearing the tag falls, or has stopped moving for a long time (man-down), and then:

• ask that person "Are you OK?" on their phone;
• tell the guards if they say they need help, do not answer, or cannot be asked;
• record every event in the Fall & Man-down report, so you can see how often it was right.

Nothing has to be set up per person. Each tag learns by itself how it is carried and how it usually rests.

Which tags can do it

Only tags whose motion readings Omaya can decode. System › Config › Fall detection lists them under Tags that can do this, with how many of each you have. Today they are:

Tag	Worn	Notes
Minew MWC01	badge, lanyard, belt clip	The best-tested tag.
MOKO B2, H8, H3 Pro buttons	lanyard, keyfob	Same button firmware; also carry a panic button.
MOKO M3 Mini, H2, H5, BXP-C, BXP-D, BXP-DH	asset tag, keyfob	Only when the tag's motion slot is switched on in the MOKO app.
Minew MWL01	wristband	Works, but a wrist hides many falls. See Where the tag is worn.

Any other tag is never watched for falls.

How it decides

A turn on its own is not enough (people sit down), and stillness on its own is not enough (badges get put down). It takes all three, in order.

Every few seconds the tag tells Omaya which way gravity points through it. A possible fall is three things in a row:

1. The tag was moving. Someone was walking about with it. A tag lying in a drawer cannot fall.
2. Its direction changed sharply, within a few seconds.
3. It then stayed still for about a minute, in a posture clearly different from the way it is usually carried while moving.

All three are needed. People sit down, bend and turn all day; on their own those are not falls.

Man-down is simpler and slower: the tag was moving, then did not move at all for 15 minutes.

Rests it already knows

Badges get taken off and put down: on a desk, in a charging tray, on a shelf. Each tag remembers the ways it usually comes to rest. When it comes to rest in one of those again, the event is still written to the report as habitual, but nobody is asked and nobody is paged. A new tag has nothing learned yet, so its first few rests on a desk can raise a question; after a few days they stop.

Turning it on

Open System › Config and the Fall detection card.

1. Switch Fall detection on.
2. Choose Watch every supported tag, or leave it off and choose with rules.
3. Pick a Sensitivity:
   • Normal for most sites;
   • Low for fewer alerts, if you would rather miss a short fall than be asked often;
   • High to react sooner, accepting more false alarms.
4. Press Save.

With fall detection on and Alert people off, events are only recorded. That is the recommended first week: look at the report, label what it found, and switch alerts on when you trust it.

The guards are told through Security Watch, so that app must be enabled for anyone to be paged. Events are recorded and reported either way; the Config page warns you if Security Watch is off.

Where the tag is worn

Where the tag sits matters more than any setting:

• On a lanyard at the chest, or clipped at the waist: the tag moves with the body, so a fall shows clearly. This is what fall detection is built for.
• On the wrist: the arm keeps moving after a fall, or comes to rest in an ordinary position, so many falls look like nothing happened. Expect fewer falls caught and more false alarms.
• In a bag, a pocket or on a desk: the tag does not move with the person and cannot feel a fall at all.

The same advice, with the accuracy for each position, is in the Wear it right card on the Config page and in the report, and the "Are you OK?" page the person opens carries a one-line reminder.

The Your tags card lists every tag that can do fall detection, with:

• Worn on: taken from the model (a wristband is the wrist, everything else is on the body) until you choose one of:

  • Lanyard (chest), Clip at the waist, On the body (lanyard or waist) or Wrist. These record how the tag is worn, for the warnings on this page, the Moves like check and the report's accuracy by position. Detection judges them all the same way.
  • Carried in a bag or pocket — not worn. Fall detection is off for that tag: a tag that is not worn cannot feel a fall, and its alarms would only be noise. Use it for tags that are carried, not worn.

  A tag set to Body before these choices existed keeps that setting, shown as On the body (lanyard or waist).

• Moves like: after the tag has moved for a while (about ten minutes of walking), Omaya says whether its own movement looks like a body or a wrist. If that disagrees with Worn on, the row is highlighted: check how that person actually wears it. This is a check, not a setting; it is right about 85 % of the time.

• Looks worn?: whether the tag seems to be on a person at all, with the reason in words, for example Lay still for 3 h or more at a stretch on 6 of the last 8 working days (longest 5 h 15 min).

If any watched tag is worn on the wrist, or looks not worn, the page shows a warning.

Looks worn?

Someone wearing a tag stands up, turns round or walks somewhere well within a few hours. A tag left on a desk, in a drawer or in a bag lies perfectly still all afternoon. Omaya keeps, for each watched tag, which quarter-hours of the last two weeks it was heard in and which it moved in, and judges:

1. Only your open days and opening hours count (set under System › Config; every hour of every day if none are set). A badge left on a desk overnight says nothing about the working day.
2. A day is judged once the tag was heard for at least 4 hours of those hours.
3. On a judged day the tag was left lying when it stayed still for 3 hours or more at a stretch: heard all the while, never moving.
4. With at least 3 judged days, the tag looks not worn when it was left lying on at least half of them, and is worn otherwise. With fewer days it says Not enough data. Times it was put down the same way it usually rests are added to the reason, but never decide on their own: a badge worn all day and put down every evening is still worn.

Only quarter-hours are kept, only for watched tags, and only for two weeks: enough to see a badge lying on a desk all afternoon, never what anyone did.

A tag that looks not worn is shown:

• on the Config page, with a warning to remind the person or set the tag as not worn;
• in the report's By tag table, with the reason;
• on the person's panel in the Entity Tracker: Tag looks not worn — fall detection cannot protect them;
• in the OAIA briefing, which can suggest reminding the person.

Rest places

Some places are where people lie down on purpose: a sick bay, a rest room, a nursing room, a prayer room, a bed. To a motion sensor, lying down on purpose looks exactly like a fall. Tell Omaya where those places are in the Rest places card:

• tick zone types (every zone of that type), or individual zones;
• choose what happens there:
  • Wait, and ask only if they still have not moved after a number of minutes (10 by default). If they get up in that time, the event is closed and nobody is asked. If they are still lying there, the normal "Are you OK?" starts, because people can collapse in a rest room too.
  • Only record it in the report.

Toilets, washrooms and showers can never be rest places. They are greyed out and refused on save: bathrooms are one of the most common places people fall, and someone lying still on a toilet floor needs help.

Fall rules (below) stay quiet in rest places.

Alerts

The guards' side is Security Watch's: the same board, the same notifications and re-alerts as a panic button.

Switch on Alert people in the Alerts card when you are ready. Then:

1. The person is asked. The app user linked to the tag's person (in People and assets) gets a push notification, and an SMS or WhatsApp if their account has a phone number: "Are you OK? Your tag detected a fall." One tap on I'm OK closes it as a false alarm. I need help tells the guards at once.
2. The guards are told if there is no answer within Time to answer (60 seconds by default), if the person asked for help, or if nobody can be asked (no app user is linked to that person). Telling them means a critical Possible fall incident opens in Security Watch at once, and Security Watch does the rest the way it does for a panic button: every guard with Security Watch gets a push, its escalation list gets email, SMS or WhatsApp, and it is sent again if nobody picks it up.
3. A guard responds on the Security Watch board: Acknowledge, Assign, then Resolve with what it was. Genuine is recorded as a real fall, False alarm and Expected as false alarms, Test as a test. That is what the Fall & Man-down report counts.

Example. Ravi Kumar's lanyard badge falls with him in the level 3 corridor. A minute later his phone asks "Are you OK?". He does not answer. Sixty seconds later a critical Possible fall — Ravi Kumar incident appears on the Security Watch board and the guards on duty get a push; Hafiz Ismail acknowledges it, finds Ravi, and resolves it as Genuine. The Fall & Man-down report shows the fall, the minute before Ravi was asked, and how long Hafiz took to respond.

Choosing who is watched with rules

With Watch every supported tag off, only tags covered by a rule are watched. Make the rule under Rules like any other (Your first rule):

• Watches: a device, a label (for example everyone labelled Night shift), or one person or asset;
• Check: the metric Fall / man-down detected, equal to fall or man_down;
• Trigger: Real-time (a scheduled trigger is refused: a fall is an event, not a level);
• Actions: anything a rule can do: notify someone, call a webhook, open a work order.

The rule's actions run in addition to the "Are you OK?" flow, which is set on the Config page. The Config page lists your fall rules with a link to each.

The Fall & Man-down report

Reports › Analysis › Fall & Man-down opens with the detection status in one line (on or off, how many tags are watched, the sensitivity, whether alerts go out or events are only recorded, and Change), then four figures for the date range:

• Flagged: events found, split into falls and man-down;
• Would alert: how many would have paged someone, per day, with habitual rests and rest places left out;
• Precision: of the events someone labelled, how many were real falls (real · false · not labelled);
• Response: how fast people answered, guards picked up and incidents were closed.

Beside the per-kind table, the accuracy panel has Label the N unlabelled, which filters the events to the ones nobody has labelled yet. When shows events per day and by hour; By tag lists tags, and clicking one filters the events to it. Each event row shows when, who (and the tag), where, the kind, whether it was a rest, its status and the answer.

While alerts are off, label events yourself with the Real / False / Test toggle on each row; it shows the label already given, and labelled rows get a coloured edge. That is what turns the report into an accuracy figure before anyone is paged. Precision only counts false alarms; a fall the detector missed never appears in the report, so it cannot be measured there.

How accurate is it? also splits your labelled events by where the tag is worn (lanyard, waist, on the body, wrist, set as not worn, looked not worn), beside the published figures below. An event from a tag that looked not worn at the time is marked tag looked not worn and left out of precision, because such a tag cannot feel a fall and its alarms say nothing about the detector; switch on Count tags that looked not worn to include them. By tag shows each tag's Looks worn? verdict with its reason, and the CSV export has the position and whether the tag looked not worn for every event.

How accurate is it?

We tested the detector on public research recordings of people falling and going about their day, played to it at the same slow rate a tag reports (one reading every 3 seconds), at Normal sensitivity:

Tag worn on	Falls caught	Everyday movement that raised a false fall
Waist (belt)	96–99 %	about 1 %
Neck (lanyard)	91 %	under 1 %
Wrist	76–82 %	about 1 in 6
Bag, pocket or desk	cannot feel a fall	—

These figures leave out one thing on purpose: lying down on purpose and staying still, which looks the same as a fall to any motion sensor, on any part of the body. That is what rest places and the "Are you OK?" question are for.

Keep in mind:

• the recorded falls were acted by volunteers onto mats; a real fall can look different;
• a fall is called about a minute after it happens, and with alerts on the person then has the time you set to answer before the guards are told;
• your own figures, from your own people, are in the report once you label events.

Questions

It says a fall, but the badge was just put on a desk. New tags have not learned their usual resting places yet. Label it False in the report; after the tag has rested the same way a few times it stops asking. If it keeps happening for one tag, check Worn on.

Someone fell and nothing happened. Check that fall detection is on, that the tag is a supported model and is being watched (all tags, or a rule covering it), that it is not set as not worn, and where it is worn. A wrist tag misses more falls. With Alert people off, the event is in the report but nobody was told.

A tag shows "Looks not worn". It lay still for most of the working day on most of the last two weeks: on a desk, in a drawer or in a bag. Remind the person to wear it on a lanyard at the chest or clipped at the waist. If it is meant to be carried, set Worn on to Carried in a bag or pocket — not worn: fall detection is then off for it and its events stop counting against the detector.

The person got the question, but the guards did not. The guards are only told if the person does not answer, asks for help, or cannot be asked. If they were meant to be told, check that Security Watch is enabled, that the guards have it under Apps, and its notification settings (push to guards, escalation list).

Can we use it for people who do not carry a phone? Yes. If no app user is linked to the person, there is nobody to ask, so the guards are told straight away.