Fall Detector

Type: component
PROBLEM

A fall needs to be detected automatically without the operator having to do anything.

SOLUTION

The Fall Detector is a three-axis MEMS accelerometer and gyroscope fusion module that continuously samples the operator's kinematic state at 100 Hz and runs an on-device edge inference model trained on the specific acceleration and rotational velocity signature of a human fall event as distinct from sitting down quickly, dropping the unit, vigorous dancing, and the roman candle launch scenario which required its own exclusion zone in the training data. A fall event is confirmed when the freefall signature is followed by an impact signature followed by a sustained stationary low-elevation posture reading, at which point a 30-second auto-cancel window opens during which the operator can indicate they are fine by pressing the button. If the button is not pressed, the alert escalates. The model was trained to minimize false positives because false positives erode trust and an operator who dismisses too many false alerts will eventually dismiss a real one, which is the exact outcome the Fall Detector exists to prevent.

Inputs
Outputs
TREE
AVAILABLE INPUTS
AVAILABLE OUTPUTS
Schema
                            {
    "id": "comp_fall_detector",
    "name": "Fall Detector",
    "type": "component",
    "inputs": [],
    "outputs": [],
    "passthru": [],
    "children": [],
    "extends": null,
    "problem": "A fall needs to be detected automatically without the operator having to do anything.",
    "description": "The Fall Detector is a three-axis MEMS accelerometer and gyroscope fusion module that continuously samples the operator's kinematic state at 100 Hz and runs an on-device edge inference model trained on the specific acceleration and rotational velocity signature of a human fall event as distinct from sitting down quickly, dropping the unit, vigorous dancing, and the roman candle launch scenario which required its own exclusion zone in the training data. A fall event is confirmed when the freefall signature is followed by an impact signature followed by a sustained stationary low-elevation posture reading, at which point a 30-second auto-cancel window opens during which the operator can indicate they are fine by pressing the button. If the button is not pressed, the alert escalates. The model was trained to minimize false positives because false positives erode trust and an operator who dismisses too many false alerts will eventually dismiss a real one, which is the exact outcome the Fall Detector exists to prevent.",
    "template": "component"
}

Universe Alpha