Fall Detection — Deployment Guide

Preset

reCamera 2002

A reCamera 2002 decides on-device whether someone fell and publishes the event over MQTT.

Important: this is an assistive alert, not a certified medical or life-safety system. Long shots, occlusion, low light and fall-like floor activities remain weak cases.

Update the reCamera Console

Install console 0.5.5, which manages the camera apps. Already current? It's skipped.

Prerequisites

Connect the reCamera over USB, or put it on the same network as this computer.
Over USB the address is 192.168.42.1; over Wi-Fi use the IP your router shows.
The default password is recamera (older units use recamera.2).
The next step, installing fall detection, needs this console version.

Troubleshooting

Symptom	Action
Cannot connect	Confirm SSH is enabled and the IP and password are correct
Console page does not load after install	Give it 30 seconds to restart, then reload `http://<camera-ip>/`
Password rejected	Try `recamera.2`; units shipped with older firmware use it

Install Fall Detection

Install the pose model and the fall detector, then start it on the camera.

Wiring

Camera placement: a side or corner view at 2–3 m works; straight-down and long-shot or occluded views do not.

Mount the camera rigidly with a clear, wide view of the area you want covered.
Keep the whole person — especially shoulders and hips — visible along the path where a fall would happen.
Aim for a side or corner view of the floor area rather than looking straight down.
Point it at circulation space, not primarily at a bed or an exercise area — everyday floor activities there read as falls until you have validated them separately.
Expect it to detect the fall itself, not the aftermath: starting it while someone is already lying down reports the posture but raises no event.

Troubleshooting

Symptom	Action
Service exits immediately	Another camera app is still running; only one app can hold the camera, so reboot and retry
Node-RED stopped working after install	Expected — installing takes the camera from Node-RED and any other vision app
Falls are missed	Widen the view, improve lighting, keep shoulders and hips visible before and after impact
Push-ups trigger an alert	A known fall-like activity; change the view or add downstream human confirmation
No MQTT messages	Confirm port 1883 is reachable from your computer and the topic is `recamera/fall-detection/results`

Watch Fall Status

Click Connect to see the skeleton, the state and the event number live.

Troubleshooting

Symptom	Action
Overlay appears before video	MQTT connects faster than RTSP; wait a few seconds
Skeleton disappears near the floor	Reframe the camera; a short post-impact gap is tolerated, long occlusion cannot be classified
No overlay at all	Confirm MQTT port 1883 is reachable and the topic matches

After Deployment

Alerts go to recamera/fall-detection/results, and Home Assistant discovery creates fall state, event ID and person presence entities.

Quick verification

Click Connect and wait for the video to appear.
Walk into view — the skeleton should follow you and the card should read NORMAL.
Lie down deliberately on the floor. Within roughly two seconds the card should turn red and show a new event number.

Next steps

Add the camera to Home Assistant — the entities appear automatically once your broker is shared with it.
Run a site acceptance test with representative falls and normal activity before enabling any notification workflow.

Install the Alarm Panel (optional)

Skip this step and the deployment is finished: the camera keeps publishing MQTT events and nothing else changes.

Take it and you install an alarm panel on a separate host: a site overview, zones drawn on the live picture, a no-person and a no-motion timeout per zone, operator confirm or dismiss for each alarm, an audit trail, and a webhook that carries no video.

Prerequisites

The panel goes on a separate box on the camera's network and needs no AI accelerator: a reComputer R1000 Series, or a Linux machine you already run.

A x86_64 or arm64 Linux host on the camera's network, with Docker and the compose plugin (docker compose version has to succeed) and SSH access.
The exact topic the cameras publish on. Check it from that host before you start: mosquitto_sub -h <camera-or-broker-ip> -t '#' -v. A reCamera 2002 publishes on <device-name>/fall-detection/results, single stream, no stream-id suffix.
Ports 8080 and 1883 free on that host, or different ports entered in the form.

Troubleshooting

Symptom	Action
Deploy stops on "Port 8080 is already in use"	Enter another Panel Port in the form, or stop the service the message names.
"no message on ... within 20 s" warning at the end	The panel is up but has seen no detector result. Re-check the topic against `mosquitto_sub -t '#' -v`, and that the camera publishes to the broker address entered here.
Alarm list stays empty and no-person alarms never fire	The camera has to publish on frames with nobody in view for that timeout to have an input. Falls still work either way.
`pull access denied` on `eldercare-alarm-*`	Check that the host can reach the image registry.

Remote device

Deploy to the panel host over SSH.

What you will be asked for(filled in the SenseCraft Solution app)

Panel Host IP*
The box that will run the panel — a reComputer R1000 or any existing Linux server with Docker
SSH Username*default recomputer
SSH Password*
Camera Model*default fall_result_v1
2 options
- · reCamera 2002
- · reCamera Pro
Which payload the cameras publish. reCamera 2002 sends fall_result_v1; reCamera Pro sends its own state message and needs the Pro adapter.
Detector Result Topic*default recamera/fall-detection/results
The exact topic the cameras publish on. Confirm it with `mosquitto_sub -h <broker> -t '#' -v` before deploying — a 2002 uses <device-name>/fall-detection/results, a Pro uses <base>/fall-detection/state. A + wildcard covers several cameras.
MQTT Broker Host*default mosquitto
Leave as mosquitto to use the broker this stack starts here, and point the cameras at this host. Enter the camera's or your own broker address to subscribe to an existing one instead.
MQTT Broker Port*default 1883
The port the panel connects to. With the bundled broker this is the same value as the published broker port below.
MQTT Username
Leave empty for the bundled broker, which accepts anonymous connections on the LAN
MQTT Password
Quotes and backslashes are not accepted — the value is written into a quoted field in the generated configuration
Broker Port to Publish on This Host*default 1883
The host port the bundled broker binds. Change it if 1883 is already taken on this machine — the deploy fails early with the name of the process holding it.
Panel Port*default 8080
The host port the panel page is served on. Change it if 8080 is already taken.
Zone Name*default bedroom
The area these cameras watch. The deploy creates one zone covering the whole frame; split it into sub-rectangles on the host afterwards if one camera covers two rooms.
No-person Timeout (minutes)*default 60
How long the zone may stay empty before an alarm is raised. Set it above the longest normal absence — a bathroom zone needs a much shorter value than a bedroom.
No-motion Timeout (minutes)*default 30
How long a person may stay in the zone without moving before an alarm is raised. Sleeping hours will trip this unless the zone excludes the bed or the value covers a nap.
Webhook URL
Where a confirmed alarm is POSTed. The payload carries the alarm id, kind, zone, stream, timestamp and operator — no snapshot and no video. Leave empty to record alarms on the panel only.

Open the Alarm Panel

Only relevant if you installed the panel in the previous step.

Troubleshooting

Symptom	Action
Page does not load	Check the Panel Port matches what the deploy step used, and that the host firewall allows it.
A login screen appears	The deployment set `ELDERCARE_API_TOKEN`. Enter that token plus an operator name — the name goes on confirm and dismiss receipts.
A room reads unknown and stream-lost	The panel host cannot reach that camera; check the network. The last-frame time on the card shows when a frame last arrived.

After Deployment

Quick verification

The overview page lists the room you named, with its camera and zone counts.
An empty alarm list on a quiet site is the correct result — it means the service is up and answering.
To prove the ingest path end to end, drop the zone's no-person timeout to one minute on the host (config/alarm-panel.yaml, then docker compose restart alarm-panel), leave the area empty, and confirm an alarm appears. Put the real value back afterwards.

Next steps

Draw the zones on the live picture instead of keeping the single whole-frame rectangle the deploy created.
Point the webhook at your own alerting system if you left it empty.
Voice check-in is off by default and needs an OpenVoiceStream instance plus a USB microphone and speaker on this host; see Step 6.

Voice Check-in (optional)

Optional, off by default. After a fall alarm is raised, the service asks the resident out loud whether they are all right and acts on the answer; turning it off leaves the alarm path unchanged.

What the answer does:

Answer	Result
A call for help ("救命", "help", "I can't get up")	Confirmed immediately, skipping the rest of the operator window
No answer at all	Confirmed immediately
Something unreadable	Confirmed immediately
"I'm fine"	Default `on_ok: needs_review` — the alarm keeps its normal timing and is flagged for a person to look at. Set `on_ok: dismiss` to close it instead

Prerequisites

An OpenVoiceStream instance on the same LAN, with a USB microphone and a speaker plugged into the box running it, not into the camera.

Troubleshooting

Symptom	Action
Every alarm gets `no_answer`	Either the prompt is inaudible, or the microphone is not being captured. Check the speaker first, then `arecord -l` on the alarm host.
Every alarm gets `unclear`	ASR is returning text the keyword lists do not match. Read the transcript in the console and add the phrasing the resident actually uses to `ok_keywords` / `help_keywords`.
Alarms close by themselves	`on_ok` is set to `dismiss`. Put it back to `needs_review` unless a person really is reviewing the dismissals.
The service starts but never speaks	Confirm the alarm host's audio device is passed into the container, and check `docker compose logs eldercare-alarm` for TTS or playback errors.

After Deployment

A distress word beats a safe word in the same sentence.

Privacy: audio is never written to disk. The audit trail keeps the verdict, confidence, latency and transcribed text; store_transcript: false drops the text. Notifications carry no snapshot and no video.

Quick verification

curl -sf http://<ovs-host>:8621/readyz returns 200.
The synthesized prompt is audible from where a fall would happen.

Preset

reCamera Pro

A reCamera Pro tracks several people and decides on-device whether someone fell, then publishes the event over MQTT.

Important: this is an assistive alert, not a certified medical or life-safety system. Long shots, occlusion, low light and fall-like floor activities remain weak cases.

The detector ships in the device's own App Center; this preset configures the installed app and makes it the running app. If your device does not have it yet, install it from the App Center first.

Update the Camera Firmware

Only needed once, and only if your camera has no App Center yet.

Device Management, the Embedded tab, and the reCamera Pro entry with its address and ADB port

What to check

Open the camera's page first — if App Center is there with Fall Detection, skip this step.
In this app: Device Management → Embedded → reCamera Pro, fill in the camera's address, then Check for device updates.
It uses ADB on port 5555, not SSH, so the camera must be on the network — USB alone is not enough.
The update reboots the camera and takes a few minutes. Do not power it off.
It keeps a copy of the factory files, so Factory reset on the same page can roll it back.

Troubleshooting

Symptom	Action
Test connection fails	Check the address, and that port 5555 is reachable from this computer
Nothing happens after Check for device updates	The camera may already be up to date — look for the App Center on its page
App Center still missing afterwards	Reload the page; the camera needs a moment after its reboot

Configure Fall Detection

Point the app at your MQTT broker and make it the running app.

What to check

The device runs one app at a time, so activating this one stops whatever is currently running.
MQTT is optional. Leave the broker address empty and you watch results on the device's own page; fill it in to forward events to Home Assistant. This camera ships no broker, unlike the 2002.
The credentials are the web console's, not SSH.

reCamera Pro

Troubleshooting

Symptom	Action
"not installed on this device"	Install Fall Detection from the device's App Center, then run this step again
Login rejected	Repeated failures lock your IP for an increasing delay — confirm the password in the console before retrying
Nothing arrives on MQTT	Check the broker address is reachable from the camera, not just from your computer. If you left it empty, results only appear on the device page — that is by design

Watch Fall Status

Open the device console and watch the live view while someone walks in front of the camera.

reCamera Pro

Troubleshooting

Symptom	Action
The camera's page does not open	The firmware has an HTTPS switch; port 80 answers with a redirect to 443. Follow it, or open the `https://` address directly
Live view works but no fall events reach the broker	Check the broker address and port in the previous step; the summary topic only appears once a person is detected
Fall Detection is listed but will not start	Its AI model is missing. Reinstall it from the App Center on the camera — the install downloads the model with the app
A fall is missed near the floor	Reframe the camera; a short post-impact gap is tolerated, long occlusion cannot be classified

After Deployment

The camera is now publishing fall events to your broker.

What it publishes

Topic	Content
`<device-name>/fall-detection/summary`	`person_count`, `fallen_count`
`<device-name>/fall-detection/fall`	`fall_event` on the transition

These topics carry no skeleton data; the live view with skeletons is on the device console page, opened by this step.

Install the Alarm Panel (optional)

Skip this step and the deployment is finished: the camera keeps publishing MQTT events and nothing else changes.

Prerequisites

The panel goes on a separate box on the camera's network and needs no AI accelerator: a reComputer R1000 Series, or a Linux machine you already run.

A x86_64 or arm64 Linux host on the camera's network, with Docker and the compose plugin (docker compose version has to succeed) and SSH access.
The exact topic the cameras publish on. Check it from that host before you start: mosquitto_sub -h <camera-or-broker-ip> -t '#' -v. A reCamera Pro publishes on <base>/fall-detection/state; pick reCamera Pro as the Camera Model in the form so the Pro adapter is used.
Ports 8080 and 1883 free on that host, or different ports entered in the form.

Troubleshooting

Symptom	Action
Deploy stops on "Port 8080 is already in use"	Enter another Panel Port in the form, or stop the service the message names.
"no message on ... within 20 s" warning at the end	The panel is up but has seen no detector result. Re-check the topic against `mosquitto_sub -t '#' -v`, and that the camera publishes to the broker address entered here.
Alarm list stays empty and no-person alarms never fire	The camera has to publish on frames with nobody in view for that timeout to have an input. Falls still work either way.
`pull access denied` on `eldercare-alarm-*`	Check that the host can reach the image registry.

Remote device

Deploy to the panel host over SSH.

What you will be asked for(filled in the SenseCraft Solution app)

Panel Host IP*
The box that will run the panel — a reComputer R1000 or any existing Linux server with Docker
SSH Username*default recomputer
SSH Password*
Camera Model*default fall_result_v1
2 options
- · reCamera 2002
- · reCamera Pro
Which payload the cameras publish. reCamera 2002 sends fall_result_v1; reCamera Pro sends its own state message and needs the Pro adapter.
Detector Result Topic*default recamera/fall-detection/results
The exact topic the cameras publish on. Confirm it with `mosquitto_sub -h <broker> -t '#' -v` before deploying — a 2002 uses <device-name>/fall-detection/results, a Pro uses <base>/fall-detection/state. A + wildcard covers several cameras.
MQTT Broker Host*default mosquitto
Leave as mosquitto to use the broker this stack starts here, and point the cameras at this host. Enter the camera's or your own broker address to subscribe to an existing one instead.
MQTT Broker Port*default 1883
The port the panel connects to. With the bundled broker this is the same value as the published broker port below.
MQTT Username
Leave empty for the bundled broker, which accepts anonymous connections on the LAN
MQTT Password
Quotes and backslashes are not accepted — the value is written into a quoted field in the generated configuration
Broker Port to Publish on This Host*default 1883
The host port the bundled broker binds. Change it if 1883 is already taken on this machine — the deploy fails early with the name of the process holding it.
Panel Port*default 8080
The host port the panel page is served on. Change it if 8080 is already taken.
Zone Name*default bedroom
The area these cameras watch. The deploy creates one zone covering the whole frame; split it into sub-rectangles on the host afterwards if one camera covers two rooms.
No-person Timeout (minutes)*default 60
How long the zone may stay empty before an alarm is raised. Set it above the longest normal absence — a bathroom zone needs a much shorter value than a bedroom.
No-motion Timeout (minutes)*default 30
How long a person may stay in the zone without moving before an alarm is raised. Sleeping hours will trip this unless the zone excludes the bed or the value covers a nap.
Webhook URL
Where a confirmed alarm is POSTed. The payload carries the alarm id, kind, zone, stream, timestamp and operator — no snapshot and no video. Leave empty to record alarms on the panel only.

Open the Alarm Panel

Only relevant if you installed the panel in the previous step.

Troubleshooting

Symptom	Action
Page does not load	Check the Panel Port matches what the deploy step used, and that the host firewall allows it.
A login screen appears	The deployment set `ELDERCARE_API_TOKEN`. Enter that token plus an operator name — the name goes on confirm and dismiss receipts.
A room reads unknown and stream-lost	The panel host cannot reach that camera; check the network. The last-frame time on the card shows when a frame last arrived.

After Deployment

Quick verification

The overview page lists the room you named, with its camera and zone counts.
An empty alarm list on a quiet site is the correct result — it means the service is up and answering.
To prove the ingest path end to end, drop the zone's no-person timeout to one minute on the host (config/alarm-panel.yaml, then docker compose restart alarm-panel), leave the area empty, and confirm an alarm appears. Put the real value back afterwards.

Next steps

Draw the zones on the live picture instead of keeping the single whole-frame rectangle the deploy created.
Point the webhook at your own alerting system if you left it empty.
Voice check-in is off by default and needs an OpenVoiceStream instance plus a USB microphone and speaker on this host; see Step 6.

Voice Check-in (optional)

Optional, off by default. After a fall alarm is raised, the service asks the resident out loud whether they are all right and acts on the answer; turning it off leaves the alarm path unchanged.

What the answer does:

Answer	Result
A call for help ("救命", "help", "I can't get up")	Confirmed immediately, skipping the rest of the operator window
No answer at all	Confirmed immediately
Something unreadable	Confirmed immediately
"I'm fine"	Default `on_ok: needs_review` — the alarm keeps its normal timing and is flagged for a person to look at. Set `on_ok: dismiss` to close it instead

Prerequisites

An OpenVoiceStream instance on the same LAN, with a USB microphone and a speaker plugged into the box running it, not into the camera.

Troubleshooting

Symptom	Action
Every alarm gets `no_answer`	Either the prompt is inaudible, or the microphone is not being captured. Check the speaker first, then `arecord -l` on the alarm host.
Every alarm gets `unclear`	ASR is returning text the keyword lists do not match. Read the transcript in the console and add the phrasing the resident actually uses to `ok_keywords` / `help_keywords`.
Alarms close by themselves	`on_ok` is set to `dismiss`. Put it back to `needs_review` unless a person really is reviewing the dismissals.
The service starts but never speaks	Confirm the alarm host's audio device is passed into the container, and check `docker compose logs eldercare-alarm` for TTS or playback errors.

After Deployment

A distress word beats a safe word in the same sentence.

Quick verification

curl -sf http://<ovs-host>:8621/readyz returns 200.
The synthesized prompt is audible from where a fall would happen.

Preset

IP Camera + reComputer J30 / J40

A reComputer J30 / J40 pulls the RTSP streams of your existing IP cameras and tracks several people per stream independently.

Camera: any ONVIF or RTSP IP camera.

Important: this is an assistive alert, not a certified medical or life-safety system. Long shots and occlusion lower recall.

Deploy Fall Detection

Deploy the detector and build its inference engine on the Jetson. Allow 10–20 min.

Prerequisites

The Jetson runs JetPack 6.x with the NVIDIA container runtime available.
At least 10 GB free disk.
Your IP camera's RTSP URL, including credentials if it requires them, for example rtsp://admin:password@192.168.1.64:554/Streaming/Channels/101.
The first deploy spends most of its time building the inference engine on the device; later deployments reuse it.
Match the pose model to the board: YOLO11s for Orin Nano, YOLO11m for Orin NX.

Troubleshooting

Symptom	Action
Engine build fails	Confirm `/usr/src/tensorrt/bin/trtexec` exists and the disk has 10 GB free
No video from the camera	Test the RTSP URL in VLC first; most failures are a wrong path or wrong credentials
Container restarts repeatedly	Check the logs for the engine path; a half-built engine from an interrupted run must be deleted
Deploy cannot connect	Confirm SSH is reachable and the username is right — Seeed images use `recomputer` or `nvidia`

Deployment method

Deploy to the Jetson over SSH from this computer.

What you will be asked for(filled in the SenseCraft Solution app)

Jetson IP*
SSH Username*default recomputer
Common defaults: recomputer (Seeed reComputer J30 / J40 series), nvidia, or ubuntu.
SSH Password*
Device Name*default recomputer
First segment of the MQTT topic — name this device however you like (product line, room, or site) so several devices stay separable
Camera RTSP URL*
Test this address in VLC first — a wrong path or wrong credentials is the most common failure
Stream ID*default cam-01
Appended to the MQTT topic so several cameras stay separable
Pose Model*default yolo11s-pose
2 options
- · YOLO11s — faster, more streams
- · YOLO11m — more accurate
YOLO11s suits Orin Nano; YOLO11m is more accurate and suits Orin NX
Alarm Panel Port*default 8080
The port the alarm panel page is served on. The stack uses host networking, so change this if 8080 is already taken on the device.
Zone Name*default bedroom
The area this camera watches. The deploy creates one zone covering the whole frame; split it into sub-rectangles on the device afterwards if one camera covers two rooms.
No-person Timeout (minutes)*default 60
How long the zone may stay empty before an alarm is raised. Set it above the longest normal absence — a bathroom zone needs a much shorter value than a bedroom.
No-motion Timeout (minutes)*default 30
How long a person may stay in the zone without moving before an alarm is raised. Sleeping hours will trip this unless the zone excludes the bed or the value covers a nap.
Webhook URL
Where a confirmed alarm is POSTed. The payload carries the alarm id, kind, zone, stream, timestamp and operator — no snapshot and no video. Leave empty to record alarms on the panel only.

Watch Fall Status

Click Connect to see each tracked person boxed, labelled and state-coloured.

Troubleshooting

Symptom	Action
Video but no overlay	The preview reads MQTT separately; confirm port 1883 on the Jetson is reachable
Overlay but no video	The preview pulls RTSP straight from the camera; confirm this computer can reach it too
Boxes flicker between people	Raise the tracker IoU threshold, or reframe so people overlap less

After Deployment

Results go to recamera/fall-detection/results/<stream-id>, one topic per camera.

Quick verification

Click Connect and wait for the camera's video to appear.
Walk into view — a box should follow you, labelled with a track number and NORMAL.
Lie down deliberately. The box should turn red and the card should show a new event number.

Adding more cameras

The detector handles several streams at once. Add them to the streams list in the configuration on the device and restart the container; each stream keeps its own tracking state and gets its own MQTT topic.

Next steps

Point your alerting system at the MQTT topic, or add the broker to Home Assistant to pick up the discovery entities.
Measure the actual frame rate on site before adding streams.

Open the Alarm Panel

The alarm panel was deployed with the detector in Step 1 on the same device, port 8080 by default. It provides a site overview, zones drawn on the live picture, a no-person and a no-motion timeout per zone, confirm and dismiss recorded against an operator, and a webhook that carries no video.

Troubleshooting

Symptom	Action
Page does not load	Check the Alarm Panel Port matches what the deploy step used, and that the device firewall allows it.
A login screen appears	The deployment set `ELDERCARE_API_TOKEN`. Enter that token plus an operator name — the name goes on confirm and dismiss receipts.
Falls raise alarms but no-person alarms never do	The detector config ships `publish_empty_frames: true`, which that timeout depends on. If you replaced `config/config.json` with the device's own copy, set the key again.

After Deployment

Quick verification

The overview page lists the zone you named in the deploy form.
An empty alarm list on a quiet site is the correct result — it means the service is up and answering.
To prove the ingest path end to end, drop the zone's no-person timeout to one minute on the device (config/alarm-panel.yaml, then docker compose restart alarm-panel), leave the area empty, and confirm an alarm appears. Put the real value back afterwards.

Next steps

Draw the zones on the live picture instead of keeping the single whole-frame rectangle the deploy created.
Point the webhook at your own alerting system if you left it empty.
Voice check-in is off by default and needs an OpenVoiceStream instance plus a USB microphone and speaker on this device; see Step 4.

Voice Check-in (optional)

Optional, off by default. After a fall alarm is raised, the service asks the resident out loud whether they are all right and acts on the answer; turning it off leaves the alarm path unchanged.

What the answer does:

Answer	Result
A call for help ("救命", "help", "I can't get up")	Confirmed immediately, skipping the rest of the operator window
No answer at all	Confirmed immediately
Something unreadable	Confirmed immediately
"I'm fine"	Default `on_ok: needs_review` — the alarm keeps its normal timing and is flagged for a person to look at. Set `on_ok: dismiss` to close it instead

Prerequisites

An OpenVoiceStream instance on the same LAN, with a USB microphone and a speaker plugged into the box running it, not into the camera.

Troubleshooting

Symptom	Action
Every alarm gets `no_answer`	Either the prompt is inaudible, or the microphone is not being captured. Check the speaker first, then `arecord -l` on the alarm host.
Every alarm gets `unclear`	ASR is returning text the keyword lists do not match. Read the transcript in the console and add the phrasing the resident actually uses to `ok_keywords` / `help_keywords`.
Alarms close by themselves	`on_ok` is set to `dismiss`. Put it back to `needs_review` unless a person really is reviewing the dismissals.
The service starts but never speaks	Confirm the alarm host's audio device is passed into the container, and check `docker compose logs eldercare-alarm` for TTS or playback errors.

After Deployment

A distress word beats a safe word in the same sentence.

Quick verification

curl -sf http://<ovs-host>:8621/readyz returns 200.
The synthesized prompt is audible from where a fall would happen.

Preset

IP Camera + reComputer RK3576 / RK3588

The detector runs on the NPU of a reComputer RK3576 / RK3588, with the same MQTT output as the other presets.

Camera: any ONVIF or RTSP IP camera.

Important: this is an assistive alert, not a certified medical or life-safety system.

Deploy Fall Detection

Deploy the detector to your Rockchip board. Allow about 5 minutes.

Prerequisites

The board runs a vendor image with the NPU driver and librknnrt.so present, plus Docker.
At least 6 GB free disk for the runtime image and the pose model.
Your IP camera's RTSP URL, with credentials if it needs them.
Choose the deployment target that matches your board; RK3588 and RK3576 models are not interchangeable. This deployment installs one camera stream.

Troubleshooting

Symptom	Action
`librknnrt.so not found`	Install the board's `rknpu2` runtime package
Model fails to load	The model must match the board — re-run the step with the correct board selected
No video from the camera	Test the RTSP URL in VLC first; most failures are a wrong path or wrong credentials
Low frame rate	Other NPU workloads compete for the accelerator; check what else is running on the board

Deployment method

Device

What you will be asked for(filled in the SenseCraft Solution app)

Device IP*
SSH Username*default radxa
Depends on the board image — commonly radxa, cat, linaro or ubuntu.
SSH Password*
Device Name*default recomputer
First segment of the MQTT topic — name this device however you like (product line, room, or site) so several devices stay separable
Camera RTSP URL*
Test this address in VLC first — a wrong path or wrong credentials is the most common failure
Stream ID*default cam-01
Appended to the MQTT topic so several cameras stay separable
Alarm Panel Port*default 8080
The port the alarm panel page is served on. The stack uses host networking, so change this if 8080 is already taken on the device.
Zone Name*default bedroom
The area this camera watches. The deploy creates one zone covering the whole frame; split it into sub-rectangles on the device afterwards if one camera covers two rooms.
No-person Timeout (minutes)*default 60
How long the zone may stay empty before an alarm is raised. Set it above the longest normal absence — a bathroom zone needs a much shorter value than a bedroom.
No-motion Timeout (minutes)*default 30
How long a person may stay in the zone without moving before an alarm is raised. Sleeping hours will trip this unless the zone excludes the bed or the value covers a nap.
Webhook URL
Where a confirmed alarm is POSTed. The payload carries the alarm id, kind, zone, stream, timestamp and operator — no snapshot and no video. Leave empty to record alarms on the panel only.

Watch Fall Status

Click Connect to see each tracked person boxed, labelled and state-coloured.

Troubleshooting

Symptom	Action
Video but no overlay	The preview reads MQTT separately; confirm port 1883 on the board is reachable
Skeleton offset from the person	Report it to Seeed
Boxes flicker between people	Raise the tracker IoU threshold, or reframe so people overlap less

After Deployment

The board is publishing to recamera/fall-detection/results/<stream-id>, one topic per camera.

Quick verification

Click Connect and wait for the camera's video to appear.
Walk into view — a box should follow you, labelled with a track number.
Lie down deliberately. The box should turn red and the card should show a new event number.

Next steps

Point your alerting system at the MQTT topic, or add the broker to Home Assistant.
Run a site acceptance test before relying on it.

Open the Alarm Panel

Troubleshooting

Symptom	Action
Page does not load	Check the Alarm Panel Port matches what the deploy step used, and that the device firewall allows it.
A login screen appears	The deployment set `ELDERCARE_API_TOKEN`. Enter that token plus an operator name — the name goes on confirm and dismiss receipts.
Falls raise alarms but no-person alarms never do	That timeout needs the detector to publish on frames with nobody in view; confirm the MQTT topic still gets messages when nobody is in front of the camera.

After Deployment

Quick verification

The overview page lists the zone you named in the deploy form.
An empty alarm list on a quiet site is the correct result — it means the service is up and answering.
To prove the ingest path end to end, drop the zone's no-person timeout to one minute on the device (config/alarm-panel.yaml, then docker compose restart alarm-panel), leave the area empty, and confirm an alarm appears. Put the real value back afterwards.

Next steps

Draw the zones on the live picture instead of keeping the single whole-frame rectangle the deploy created.
Point the webhook at your own alerting system if you left it empty.
Voice check-in is off by default and needs an OpenVoiceStream instance plus a USB microphone and speaker on this device; see Step 4.

Voice Check-in (optional)

Optional, off by default. After a fall alarm is raised, the service asks the resident out loud whether they are all right and acts on the answer; turning it off leaves the alarm path unchanged.

What the answer does:

Answer	Result
A call for help ("救命", "help", "I can't get up")	Confirmed immediately, skipping the rest of the operator window
No answer at all	Confirmed immediately
Something unreadable	Confirmed immediately
"I'm fine"	Default `on_ok: needs_review` — the alarm keeps its normal timing and is flagged for a person to look at. Set `on_ok: dismiss` to close it instead

Prerequisites

An OpenVoiceStream instance on the same LAN, with a USB microphone and a speaker plugged into the box running it, not into the camera.

Troubleshooting

Symptom	Action
Every alarm gets `no_answer`	Either the prompt is inaudible, or the microphone is not being captured. Check the speaker first, then `arecord -l` on the alarm host.
Every alarm gets `unclear`	ASR is returning text the keyword lists do not match. Read the transcript in the console and add the phrasing the resident actually uses to `ok_keywords` / `help_keywords`.
Alarms close by themselves	`on_ok` is set to `dismiss`. Put it back to `needs_review` unless a person really is reviewing the dismissals.
The service starts but never speaks	Confirm the alarm host's audio device is passed into the container, and check `docker compose logs eldercare-alarm` for TTS or playback errors.

After Deployment

A distress word beats a safe word in the same sentence.

Quick verification

curl -sf http://<ovs-host>:8621/readyz returns 200.
The synthesized prompt is audible from where a fall would happen.

Preset

IP Camera + reComputer R2000 (Hailo)

The detector runs on a reComputer R2000 with a Hailo-8 accelerator, with the same MQTT output as the other presets.

Camera: any ONVIF or RTSP IP camera.

Important: this is an assistive alert, not a certified medical or life-safety system.

Deploy Fall Detection

Deploy the detector to your Hailo-equipped device. Allow about 5 minutes.

Prerequisites

A Hailo-8 accelerator present as /dev/hailo0, with HailoRT 4.21 installed (GStreamer plugin, user library and kernel driver all at that version).
Docker, and at least 4 GB free disk.
Your IP camera's RTSP URL, with credentials if it needs them.
The pose model downloads automatically during deployment.
Stop any other application using the Hailo accelerator first.

Troubleshooting

Symptom	Action
`No /dev/hailo0`	The accelerator is not seated or its driver is not loaded; check `hailortcli fw-control identify`
`libhailort.so.4.21.0 not found`	Install HailoRT 4.21, with plugin, library and driver all at that version
Container starts then exits	Stop the other process using the accelerator
No video from the camera	Test the RTSP URL in VLC first; most failures are a wrong path or wrong credentials
Deployment stops before verification	Check the detector log for the `HAILO_BATCH` line, container health, and an MQTT result on the configured topic

Deployment method

Deploy to the device over SSH from this computer.

What you will be asked for(filled in the SenseCraft Solution app)

Device IP*
SSH Username*default pi
Depends on the image — commonly pi, recomputer or ubuntu.
SSH Password*
Device Name*default recomputer
First segment of the MQTT topic — name this device however you like (product line, room, or site) so several devices stay separable
Camera RTSP URL*
Test this address in VLC first — a wrong path or wrong credentials is the most common failure
Stream ID*default cam-01
Appended to the MQTT topic so several cameras stay separable
Alarm Panel Port*default 8080
The port the alarm panel page is served on. The stack uses host networking, so change this if 8080 is already taken on the device.
Zone Name*default bedroom
The area this camera watches. The deploy creates one zone covering the whole frame; split it into sub-rectangles on the device afterwards if one camera covers two rooms.
No-person Timeout (minutes)*default 60
How long the zone may stay empty before an alarm is raised. Set it above the longest normal absence — a bathroom zone needs a much shorter value than a bedroom.
No-motion Timeout (minutes)*default 30
How long a person may stay in the zone without moving before an alarm is raised. Sleeping hours will trip this unless the zone excludes the bed or the value covers a nap.
Webhook URL
Where a confirmed alarm is POSTed. The payload carries the alarm id, kind, zone, stream, timestamp and operator — no snapshot and no video. Leave empty to record alarms on the panel only.

Watch Fall Status

Click Connect to see each tracked person boxed, labelled and state-coloured.

Troubleshooting

Symptom	Action
Video but no overlay	The preview reads MQTT separately; confirm port 1883 on the device is reachable
Overlay but no video	The preview pulls RTSP straight from the camera; confirm this computer can reach it too
`inference_time_ms` reads 0	Expected; the Hailo runtime does not report this value

After Deployment

The device is publishing to recamera/fall-detection/results/<stream-id>, one topic per camera.

Quick verification

Click Connect and wait for the camera's video to appear.
Walk into view — a box should follow you, labelled with a track number.
Lie down deliberately. The box should turn red and the card should show a new event number.

Next steps

Point your alerting system at the MQTT topic, or add the broker to Home Assistant.
Run a site acceptance test before relying on it.

Open the Alarm Panel

Troubleshooting

Symptom	Action
Page does not load	Check the Alarm Panel Port matches what the deploy step used, and that the device firewall allows it.
A login screen appears	The deployment set `ELDERCARE_API_TOKEN`. Enter that token plus an operator name — the name goes on confirm and dismiss receipts.
Falls raise alarms but no-person alarms never do	Check that the zone's stream id matches the Stream ID from the deploy form.

After Deployment

Quick verification

The overview page lists the zone you named in the deploy form.
An empty alarm list on a quiet site is the correct result — it means the service is up and answering.
To prove the ingest path end to end, drop the zone's no-person timeout to one minute on the device (config/alarm-panel.yaml, then docker compose restart alarm-panel), leave the area empty, and confirm an alarm appears. Put the real value back afterwards.

Next steps

Draw the zones on the live picture instead of keeping the single whole-frame rectangle the deploy created.
Point the webhook at your own alerting system if you left it empty.
Voice check-in is off by default and needs an OpenVoiceStream instance plus a USB microphone and speaker on this device; see Step 4.

Voice Check-in (optional)

Optional, off by default. After a fall alarm is raised, the service asks the resident out loud whether they are all right and acts on the answer; turning it off leaves the alarm path unchanged.

What the answer does:

Answer	Result
A call for help ("救命", "help", "I can't get up")	Confirmed immediately, skipping the rest of the operator window
No answer at all	Confirmed immediately
Something unreadable	Confirmed immediately
"I'm fine"	Default `on_ok: needs_review` — the alarm keeps its normal timing and is flagged for a person to look at. Set `on_ok: dismiss` to close it instead

Prerequisites

An OpenVoiceStream instance on the same LAN, with a USB microphone and a speaker plugged into the box running it, not into the camera.

Troubleshooting

Symptom	Action
Every alarm gets `no_answer`	Either the prompt is inaudible, or the microphone is not being captured. Check the speaker first, then `arecord -l` on the alarm host.
Every alarm gets `unclear`	ASR is returning text the keyword lists do not match. Read the transcript in the console and add the phrasing the resident actually uses to `ok_keywords` / `help_keywords`.
Alarms close by themselves	`on_ok` is set to `dismiss`. Put it back to `needs_review` unless a person really is reviewing the dismissals.
The service starts but never speaks	Confirm the alarm host's audio device is passed into the container, and check `docker compose logs eldercare-alarm` for TTS or playback errors.

After Deployment

A distress word beats a safe word in the same sentence.

Quick verification

curl -sf http://<ovs-host>:8621/readyz returns 200.
The synthesized prompt is audible from where a fall would happen.