Fall Detection

Detailed explanation of the SmartFall fall detection algorithm.

Algorithm Overview

Fall detection uses accelerometer and gyroscope data to identify sudden motion patterns characteristic of falls.

Confidence Calculation

confidence = freefall_score + tumbling_score + vertical_score
confidence = min(confidence, 1.0)

Component Scores

Free Fall Detection (0.0-0.3)

Sudden drop in acceleration indicating loss of support:

totalAccel = sqrt(accel_x² + accel_y² + accel_z²)
accelChange = totalAccel - previousAvgAccel

if accelChange < -5.0 m/s²:
  freefall_score = 0.3
else:
  freefall_score = 0.0

Tumbling Detection (0.0-0.3)

High angular velocity indicating rotation:

rotationRate = sqrt(gyro_x² + gyro_y² + gyro_z²)

if rotationRate > 500 °/s:
  tumbling_score = 0.3
else:
  tumbling_score = 0.0

Vertical Component (0.0-0.2)

Significant vertical motion:

verticalVelocity = accel_z - 9.8

if abs(verticalVelocity) > 3.0:
  vertical_score = 0.2
else:
  vertical_score = 0.0

Confidence Levels

Level	Range	Status	Action	Color
NO_FALL	0.00-0.30	No detection	None	Green
SUSPICIOUS	0.31-0.50	Possible	Log event	Yellow
POTENTIAL	0.51-0.70	Likely	Alert caregiver	Orange
HIGH	0.71-0.89	Strong	Notify + alert	Red
CONFIRMED	0.90-1.00	Definite	Emergency	Dark Red

Example Scenarios

Normal Movement (Confidence: 0.15)

accel: [0.5, 0.3, 9.8]  (normal)
gyro: [0.1, 0.2, -0.05] (slow rotation)
result: NO_FALL

Tripping (Confidence: 0.45)

accel: [-2.5, -3.0, 8.0]  (acceleration drop)
gyro: [100, 80, 50]        (moderate rotation)
result: SUSPICIOUS

Severe Fall (Confidence: 0.88)

accel: [-15.0, -12.0, -5.0]  (major acceleration drop)
gyro: [500, 450, -300]        (high rotation)
result: HIGH

Time Windows

Detection Windows

Immediate: 0.5 seconds (500ms)
Confirmation: 2 seconds (2000ms)
Recovery: 5 seconds (5000ms)

Algorithm must detect fall pattern within 500ms to trigger immediate alert.

Environmental Factors

Factors that can affect detection:

Factor	Impact	Mitigation
User height	Different gravity profile	Calibration per user
Age/mobility	Slower fall patterns	Adjustable thresholds
Fall type	Forward vs. sideways	Multi-axis detection
Device position	Wrist vs. chest	Device position awareness
Surface type	Carpet vs. tile	Impact detection

Post-Fall Analysis

After detecting a fall:

1. Check for recovery motion
   - Standing up within 2 minutes?
   - Normal acceleration patterns?

2. Assess fall severity
   - Peak acceleration values
   - Duration of abnormal patterns
   - Device orientation changes

3. Request user confirmation
   - Toast notification on device
   - "Did you fall?" prompt
   - Timeout: 30 seconds

4. Escalate if needed
   - No response → notify caregivers
   - User confirms → immediate alert
   - User dismisses → log event

Machine Learning Enhancement

Optional ML-based refinement:

// Classify fall with ML model
const mlScore = await fallDetectionModel.predict({
  accelData: last500ms,
  gyroData: last500ms,
  previousState: deviceHistory
});

// Combine with rule-based score
confidence = 0.6 * ruleBasedScore + 0.4 * mlScore;

Tuning Parameters

Sensitivity Levels

# Conservative (fewer false positives)
FALL_ACCEL_THRESHOLD=-3.0
FALL_GYRO_THRESHOLD=300

# Balanced (default)
FALL_ACCEL_THRESHOLD=-5.0
FALL_GYRO_THRESHOLD=500

# Sensitive (more detections)
FALL_ACCEL_THRESHOLD=-7.0
FALL_GYRO_THRESHOLD=700

User Calibration

Calibrate per user:

# Submit baseline motion data
POST /api/device/calibrate
{
  "type": "sitting",
  "samples": [/* 100 readings */]
}

Testing Fall Detection

Simulated Fall Data

{
  "device_id": "AA:BB:CC:DD:EE:FF",
  "accel_x": -15.0,
  "accel_y": -12.0,
  "accel_z": -5.0,
  "gyro_x": 500,
  "gyro_y": 450,
  "gyro_z": -300,
  "uptime_ms": 3600000
}

Expected Response

{
  "fallDetected": true,
  "fallConfidence": 0.88,
  "status": "HIGH"
}

Common Issues

False Positives

Reducing false positives:

Cause: Jumping, running, vigorous exercise
Solution: Check for sustained normal patterns after spike
          Require 1+ second of normal accel before clearing

Missed Falls

Improving sensitivity:

Cause: Slow falls, falling on low surfaces
Solution: Increase detection window
          Lower acceleration thresholds
          Add impact detection

Calibration Drift

Managing sensor drift:

Solution: Recalibrate quarterly
          Track baseline changes
          Adaptive thresholds per user