On-device fall detection that alerts carers without recording the room

Client Industry

Healthcare technology · eldercare & assisted living

Key Challenge

Detect falls reliably without recording private spaces

Core Tech Stack

Computer Vision · Pose estimation · Raspberry Pi · IoT

Deployment Model

Fully on-device inference at the edge

SoftmaxAI has designed a privacy-preserving fall-detection device based on a single-board computer equipped with a camera. Each frame is processed entirely on the device itself; specifically, a detector identifies the people in the scene, a pose model analyzes their body keypoints, and a classifier then assesses the trajectory of these keypoints over a short time window in order to tell the difference between a fall and someone intentionally sitting, bending, or lying down. Before reaching the classifier, any motion that is not human is filtered out by means of object class identification. Nothing is recorded and no images are transmitted — the only thing that happens when a fall is confirmed is that an alert is sent containing the event and its timestamp.

• Wearables depend on being worn. A pendant or wristband has no effect unless it is worn by the person at the moment they fall — and this is exactly when it is most commonly left on a bedside table.
• A camera feed is not acceptable in a bedroom. Streaming a private living area to a monitoring service solves detection only by creating a surveillance problem, and families will not allow this to happen in the rooms where falls actually occur.
• A single frame cannot tell you what happened. In a single image a person who is lying on the floor and a person who is lying down appear exactly the same; it is necessary to reason about how they came to that position in order to tell them apart.
• False alarms destroy the system. If a pet crosses the room, a cushion is dropped, or someone sits down quickly, the carer should not be summoned – alerts which people come to ignore are worse than there being no alert at all.

The device operates as a closed loop that never leaves the edge — the camera sends data to a pipeline located on the device, which then detects people, identifies their posture, classifies the motion over a short time period, and issues an alert. It is only the event that crosses the network boundary.

• Spec before code. We established what counts as a fall, the confirmation period before issuing an alert, the escalation procedure, and the strict rule that no imagery is stored or transmitted — all of these having been agreed upon before implementation started.
• Inference at the edge. Since detection, pose estimation, and classification all take place on the single-board computer, the device has no need for a video uplink and will continue to function even if the network is down.
• Motion over time, not a single frame. The classifier looks at a brief trajectory of body keypoints — specifically the speed and direction of the descent together with the posture that comes after — and it is this that distinguishes a fall from someone lying down.
• Filter before you classify. Objects are classified before the fall logic runs, so a pet moving through the frame is discarded rather than evaluated. Human-owned thresholds and a confirmation delay then suppress the remaining noise.

• Detection without a worn device. The person is not monitored but the room is, which means that protection does not rely on remembering to put something on.
• Privacy preserved by architecture. No footage is retained and no imagery leaves the unit — the privacy guarantee is a property of where the computation happens, not a policy promise.
• Pets do not trigger carers. Class filtering ahead of the fall classifier removes the single largest source of false alarms in a real home.
• Alerts within seconds of the event. A confirmed fall notifies family members or care staff immediately, which is the window that determines how serious the consequences become.
• One unit, household or ward. Whether the device is being used by an individual living on their own or in a care home room, the coverage is increased by adding units rather than by altering the system.