elektor
5568

FallGuard is an innovative project designed to detect falls in a privacy-respecting and efficient way using Edge AI technology. The system is built on the STM32N6570-DK development board paired with the MB1854B camera, enabling intelligent local processing without relying private images to the cloud—ensuring maximum privacy for the user.

What Makes This System Special?

The Hardware Foundation
At the heart of FallGuard lies the STM32N6570-DK development board, which represents a significant leap forward in embedded AI capabilities. 
The camera system uses the MB1854B camera module.

Understanding Edge AI
Edge AI refers to running artificial intelligence algorithms directly on local devices rather than sending data to cloud servers. Think of it like having a smart assistant that works entirely within your home rather than needing to call a remote service center for every question.
The benefits are substantial:
  • Instant response - no waiting for internet communication
  • Complete privacy - sensitive video never leaves the device
  • Reliable operation - works even without internet connection
  • Lower costs - no cloud processing fees
  • Reduced bandwidth - only sends simple alerts, not video data
All these advantages are thankfully used in this project.

How Fall Guard Detects Falls: Two AI Approaches

FallGuard demonstrates two distinct AI-powered methods for detecting falls, each with unique advantages:


Approach 1: Object Detection Using Visual Recognition
object-detection-with-camera-a-v3.jpg
This method works like teaching a computer to recognize what a fallen person looks like in a picture. The system uses a SSD MobileNet V1 neural network that has been specially trained to identify two things:
  • Normal standing/sitting people
  • People who have fallen
How it works:
  1. The camera captures video frames
  2. Frames are analyzed at a fixed rate of 20fps by the AI model
  3. The system looks for the distinctive visual patterns of a person on the ground
  4. When a fall is detected consistently over several frames, an alert is triggered
Performance results of training the model:
  • Person detection: 100% precision, 4.9% recall
  • Fall detection: 93.8% precision, 20.4% recall
  • Processing time: 21.34 milliseconds per frame
This approach works by detecting falls after they occur, providing high confidence when someone is found on the ground in the strangest positions.

Approach 2: Pose Estimation Using Body Movement Analysis
pose-estimation-with-camera-a-v3-20250713181221.jpg
This method analyzes human body positions by tracking key points on the body, similar to how motion capture works in movies. The system identifies key points on each person (head, shoulders, elbows, wrists, hips, knees, ankles) and monitors how these points move relative to each other.
The Detection Algorithm:
The system groups keypoints into five body regions:
  • Head region (keypoints 0-4) - colored green on display
  • Shoulders (keypoints 5-6) - colored blue
  • Hips (keypoints 11-12) - colored magenta
  • Knees (keypoints 13-14) - colored orange
  • Ankles (keypoints 15-16) - colored orange
Fall detection logic:
  1. Calculate the average vertical position of each body region
  2. Compare positions between different regions
  3. When any two regions align horizontally (within 30 pixels), this suggests the person is lying down
  4. Trigger a fall alert only when this condition persists for 6 consecutive frames (300 milliseconds)
Key advantages:
  • Detects falls as they happen, not just after
  • Can work even when the person is partially outside the camera view
  • More sensitive to the actual motion of falling
Smart Environmental Monitoring

Beyond fall detection, FallGuard incorporates intelligent environmental sensors to create a comprehensive safety system:

Human Presence Detection with Radar
The HLK-LD2410c sensor uses 24 GHz radar technology to detect human presence with remarkable capabilities:
  • Detects both moving and stationary people up to 5 meters away
  • Works through walls and in darkness - unlike cameras, radar isn't affected by lighting
  • Distinguishes between different types of movement - motion or complete stillness
  • Configurable sensitivity zones - can be tuned for specific room layouts via smartphone app over bluetooth.
Special feature: When someone remains completely still for 20 seconds, the system generates a "freeze detection" alert, which combined with fall detection can indicate the severity of an incident.

Intelligent Lighting Control
The TSL2561 light sensor measures ambient light levels with high precision:
  • Dual photodiode design measures both visible and infrared light
  • Wide dynamic range from 0.1 to 40,000+ Lux
  • Approximates human eye response to lighting conditions
The system automatically controls room lighting based on occupancy and light levels, ensuring the camera can always track people effectively while providing safety lighting when needed.

Comprehensive Alert and Monitoring System

Real-Time Communication with MQTT
MQTT (Message Queuing Telemetry Transport) serves as the communication backbone of Fall Guard. This lightweight messaging protocol is specifically designed for IoT devices and offers:
  • Minimal data usage - perfect for battery-powered devices
  • Reliable delivery with three quality-of-service levels
  • Publish-subscribe architecture - multiple devices can receive the same alerts
The system publishes JSON-formatted messages containing:
 json { "ts": 1752327297 "mac": "0080E11D6279" "nb_detect": 0 "event": 1 "lux": 1.2 "sfal": "normal" "smov": "move" "slum": "light" } 


Automated Response with Node-RED
node-red-flowv2.jpg
Node-RED serves as the automation engine, providing a visual programming interface where users can create complex logic flows by connecting nodes.
Integrations:
  • Instant notifications via Pushbullet and Slack
  • Data logging to InfluxDB for historical analysis
  • Visual dashboards through Grafana for real-time monitoring

Data Storage and Visualization
grafana-a-v3.jpg
InfluxDB handles time-series data storage:
  • Optimized for sensor data - designed specifically for time-stamped measurements
  • High-speed ingestion - can handle millions of data points per second
  • Efficient compression - stores large amounts of historical data economically
Grafana provides comprehensive visualization:
  • Real-time dashboards showing current system status
  • Historical trending of fall incidents and environmental conditions
  • Alert management with threshold-based notifications

Comparing the Two Detection Methods
Object Detection= OD / Pose Estimation= PE
Detection Timing |  OD: After fall occurs |  PE: During fall motion
Processing Speed |  OD: Slower but more certain |  PE: Faster response
Partial Visibility |  OD: Requires full person in frame |  PE: Works with partial visibility
Awkward Positions |  OD: Better at detecting unusual fall positions |  PE: May struggle with complex poses
False Positives |  OD: Lower rate due to temporal filtering |  PE: Higher rate, requires careful tuning
Best Use Case |  OD: Monitoring unconscious individuals |  PE: Active fall prevention

System States and Intelligence

Fall Guard operates using a state machine that combines multiple sensor inputs:

Room State (Environment)
  • Radar sensor: Determines if room is empty, has moving people, or stationary people
  • Light sensor: Tracks lighting conditions with hysteresis to prevent flickering
Object State (AI Analysis)
  • Object detection: Identifies fallen vs. normal postures this is the trained model class
  • Pose estimation: Analyzes body positions and movement vectors
This multi-layered approach allows the system to distinguish between different types of incidents:
  • Normal activity - person moving around normally
  • Potential concern - person stationary but upright
  • Fall with movement - person fell but still moving
  • Fall without movement - person fell and completely still (highest priority)
The complexity is simplified with an easy to understand "traffic-light" dashboard. 
dashboard01.PNG
dashboard02.PNG
dashboard03-20250713182557.PNG
Privacy and Open Source Commitment

Fall Guard's commitment to privacy extends beyond just local processing:
  • No cloud dependencies - all AI processing occurs on-device
  • Open source code - complete transparency in implementation
  • Configurable sensitivity - users control detection parameters
  • Local data storage - historical data remains on local network
The complete source code is available on GitHub:
  • Object Detection version: github.com/ginodecock/fall-guard-od
  • Pose Estimation version: github.com/ginodecock/fall-guard-pe

Real-World Applications

Fall Guard's technology is particularly valuable in:
  • Elderly care facilities - 24/7 monitoring without privacy invasion
  • Independent living - peace of mind for families
  • Rehabilitation centers - tracking patient mobility and safety
  • Smart homes - integration with existing home automation

Conclusion

Fall Guard explores what cutting edge technology has to offer. By combining powerful edge AI processing with intelligent environmental sensors, it creates a comprehensive monitoring system that respects user privacy while providing reliable fall detection.
The dual approach of object detection and pose estimation offers flexibility of choice for different use cases - whether you need the certainty of post-fall detection or the speed of real-time fall prevention. The complete automation stack, from sensors through visualization, demonstrates how modern embedded systems can create sophisticated IoT solutions that operate entirely within your local environment.
Most importantly, Fall Guard shows that advanced AI capabilities don't require sacrificing privacy. By processing everything locally on powerful embedded hardware, it provides the intelligence of cloud-based systems with the security and reliability of local processing.