reTerminal Fall Detection System Across Industries
Built on reTerminal, the system raises an alarm when a fall is detected, with monitoring and setup on the touchscreen; used in healthcare, elderly care and industrial sites.
Elderly Care Monitoring
Fall alerts reach carers in seconds; video stays local
In-room cameras detect falls on-device and send sub-second alerts; supports mic arrays for distress calls and abnormal sound detection.
On-Site Decision
1.4 s mean alert on the test set, even offline
Video Stays Local
Alerts carry time, zone and type, no images
Fits Your System
Alerts go straight to your nurse-call or care system
Scenario details
In-room AI hardware does two things: detects falls and captures audio for call-for-help detection.
Pain Points & Challenges
Relying on manual nursing rounds in senior care cannot guarantee 24/7 observation, leaving elders unable to call for help after sudden falls; streaming live camera footage to cloud servers causes severe personal privacy violations and regulatory risks.
Edge AI Solution
Deploy edge AI cameras or local edge hosts to run human pose estimation and fall detection algorithms on-premise, sending lightweight alert payloads only upon confirmed incidents while keeping video frames on-device.
Core Advantages
- On-device inference
- Video stays local
- Per-zone timeouts
Fall Detection
Detects falls from body pose; 1.4 s mean to event on the held-out test set, and it keeps working offline.
Zones and Timeouts
Draw zones on the picture, each with its own no-person and no-motion timeout; keep beds outside the zones.
Camera Placement
Mount side-on or at an angle, 2–3 m away, with shoulders and hips visible; no top-down views.
Pain Points & Challenges
Bathrooms, restrooms, and dark nighttime bedrooms cannot be monitored by optical cameras due to strict privacy taboos and lack of illumination, creating dangerous blind spots when elders slip or face sudden medical distress.
Edge AI Solution
Install far-field 4-mic arrays paired with edge acoustic processing units to cancel ambient echoes and run local detection of groans, blunt floor impacts, and distress wake-words, triggering nurse calls within milliseconds.
Core Advantages
- Far-field 4-mic pickup
- On-device wake-word & acoustic anomaly detection
- Covers spots cameras cannot see
Far-Field Audio & Noise Cancellation
Hardware AEC and 4-mic beamforming eliminate room echo and background noise.
On-Device Distress Call & Event Detection
Edge hosts run keyword spotting (such as "help") and acoustic event models, streaming alerts to nurse stations and care systems in milliseconds.
Deployment & Selection
Choose the in-room devices first, then pick a detection host by room count.
Not sure which to choose? See the full selection guide →
Room Devices
Is there a camera in the room already? Choose the last option for audio capture only.
In-room edge sensing and alarms, aggregated locally into nurse-call systems
The camera provides the picture, the detection host decides, and events go to your nurse-call or care system. With reCamera, the camera decides on its own and no detection host is needed.
Core Advantages
- In-room capture
- On-site decision
- Event push
| Image | Product | Role | Integration |
|---|---|---|---|
 | reCamera Pro | Camera + on-device fall detection | In the reference design |
 | reCamera 2002 | Camera + on-device fall detection | In the reference design |
 | reSpeaker XVF3800 | Far-field 4-mic array for calls and sounds | On-device acoustic AI supported |
Edge cameras in bedrooms and living areas, far-field audio in private zones
Rooms that can take a camera use an all-in-one AI camera. Where cameras cannot be placed (bathrooms, private areas), far-field mic arrays detect sounds and calls for help.
Core Advantages
- All-in-one camera
- Far-field microphone
| Image | Product | Accelerator | When to choose |
|---|---|---|---|
 | reComputer R2035-12 | Hailo-8 | Many rooms on one host; 16 sub-streams measured |
 | reComputer J4012 | Jetson Orin NX | 16 streams 1080p full HW decode; local VLM headroom |
 | reComputer J3011 | Jetson Orin Nano | 8 streams 1080p full HW decode; high energy efficiency |
 | reComputer RK3588 | Rockchip NPU | Already on Rockchip; 5 streams measured |
 | reComputer RK3576 | Rockchip NPU | Already on Rockchip; 1 stream measured |
Multi-room streams aggregated on edge hosts with multi-view inference
Existing IP cameras need a detection host. The decision logic is the same on every host; they differ in accelerator and stream count.
Core Advantages
- Many rooms, one host
- Several views
- Existing Rockchip boards
FAQ
No internet is needed. Pose estimation, tracking and the fall decision run on the reCamera or the detection host; the internet is involved only if you forward events to an outside system. Video stays on the device, and what crosses the network is a JSON message of a few hundred bytes per frame.
Measured at 15 FPS: reComputer R2000 (Hailo-8) 16 streams, J40 9, J30 8, RK3588 5, RK3576 1; a reCamera handles only its own stream. The deployment presets are single-stream, so test with your own cameras before fixing the count.
The detector publishes one JSON message per frame over MQTT and discovery configs under homeassistant/, so Home Assistant only needs to connect to the same broker. The alarm panel can also send a webhook POST with alarm id, type, zone and time, without video; with reCamera packages the alarm panel needs an extra host.
Yes. The solution supports far-field acoustic sensing using reSpeaker XVF3800 mic arrays paired with edge hosts, running on-device keyword spotting (such as "help" or "fall") and acoustic event detection (thuds, groans) to provide 24/7 protection in private areas where cameras are not allowed. The alarm panel also supports two-way voice check-ins: when a fall is detected, it broadcasts an inquiry to the room, escalating the alert if distress is heard or if there is no response.
No. Fall detection is not a medical device or a certified emergency-response product, and alarms need a person to assess and respond. On a 27-clip held-out test set, six configurations averaged 95.8% fall recall; on an external dataset with long shots and occlusion, recall was 52.9%–58.8%.
Yes. The Seeed devices used in this solution can be customized with your own logo, enclosure, packaging and pre-loaded firmware, and existing models can be adapted to add or drop interfaces and features. See Customization Service for scope and process, or .
Didn't find your answer?
Contact Us
We Are Glad to Be Your Hardware Partner !