Edge-based elderly fall detection using Arm Ethos-U NPU

The problem: Falls are the leading cause of injury death in older adults

Every year, millions of older adults suffer fall-related injuries. Many falls go unwitnessed or discovered too late for timely help. Traditional solutions such as wearable panic buttons often go unused. They fail to detect falls automatically, or not provide enough context for an effective response.

Cloud-based camera monitoring raises privacy concerns and adds latency. It also depends on internet connectivity. In older adult care, these limitations can mean the difference between rapid intervention and tragedy.

The challenge is clear. We need a fall detection system that works in real-time, preserves privacy, and runs fully offline. All without compromising accuracy or affordability.

Figure 1: System running on NXP i.MX93 FRDM board with HDMI display showing live fall detection.

Figure 2: Fall detected alert with confidence score

Why it matters: Privacy, speed, and always-on protection

For older adults living independently, reliable fall detection offers peace of mind. It also reassures their families. But current solutions fall short.

The Arm Ethos-U NPU changes what is possible at the edge. It enables complex pose estimation models to run in milliseconds.

Comparison of cloud-based systems vs. Edge AI solution

Challenge	Cloud-Based Systems	Our Edge AI Solution
Privacy	Video streams to remote servers	All processing on-device
Latency	500ms–2s network delay	<30ms end-to-end
Reliability	Fails during outages	Works 100% offline
Cost	Monthly subscription fees	One-time hardware cost
Data Security	Exposed to breaches	Local network only

The solution: Real-time fall detection powered by Arm Ethos-U NPU

This open-source project shows how the Arm Ethos-U65 Neural Processing Unit (NPU) enables real-time pose estimation and fall detection entirely on-device. No cloud required. No compromise on performance or accuracy. No privacy concerns.

Key implementation highlights

• Ethos-U NPU acceleration: The MoveNet pose estimation model is quantized (INT8) and compiled using Arm's Vela compiler for optimal Ethos-U65 execution. Inference runs in 8ms—fast enough for 30+ FPS real-time processing.
• Multi-factor fall detection algorithm: Falls are detected by analyzing multiple biomechanical indicators simultaneously:
  • Body angle deviation from vertical
  • Downward velocity during fall motion
  • Bounding box aspect ratio changes
  • Upper/lower body position differential
  • Horizontal spread (sprawled position detection)
• Privacy-first architecture: All video processing occurs on-device. No frames are transmitted externally. The system can operate completely air-gapped from any network.
• Multiple output modes: Direct framebuffer display for maximum performance, Web dashboard for remote monitoring (local network), REST API for integration, and configurable notifications.

Figure 3: Data flow diagram showing camera → NPU inference → fall detection → alert flow.

Hardware setup

The Ethos-U65 NPU provides up to 1 TOPS (Tera Operations Per Second) of neural network inference performance. It uses minimal power which makes it ideal for always-on monitoring applications.

Figure 4: Hardware components

Hardware specifications

Component	Specification
NPU	Arm Ethos-U65 (256 MAC units)
Processor	Arm Cortex-A55
Board	NXP i.MX93 FRDM
Camera	USB Webcam (640×480)
Display	HDMI Monitor (optional)
Memory	2GB RAM
Storage	MicroSD or eMMC

Arm Ethos-U NPU: The key enabler

• The Ethos-U65 NPU is designed for efficient neural network inference at the edge. Key advantages for this application include the following:
• Optimized operator support: The Vela compiler maps supported TensorFlow Lite operations directly to NPU hardware (Depthwise Convolutions, Standard Convolutions, Fully Connected layers, etc.).
• Memory efficiency: The Ethos-U architecture minimizes data movement between memory and compute units.
• Power efficiency: Neural network inference on dedicated NPU hardware consumes significantly less power than CPU or GPU alternatives.

Model optimization with Vela

The MoveNet model is prepared for Ethos-U execution using Arm's Vela compiler:

vela movenet_lightning_int8.tflite \
--accelerator-config ethos-u65-256 \
--optimise Performance

Vela analyzes the model graph and fuses operations where possible. It then generates optimized command streams for the NPU.

Figure 6: Vela compilation output showing operator delegation.

Performance metrics

Metric	Value	Notes
NPU Inference Time	≈8ms	MoveNet Lightning INT8
End-to-End Latency	<30ms	Camera to detection result
Frame Rate	30+ FPS	Framebuffer display mode
Model Size	2.5MB	Vela-optimized
Power Consumption	≈2W	Typical operation
Detection Accuracy	>65%	Controlled testing
False Positive Rate	<5%	Normal activity filtering

Fall detection algorithm

The algorithm analyzes pose keypoints to detect falls using multiple indicators. A fall is confirmed when either of the following conditions is met:

• (Angle exceeded AND Velocity exceeded) or
• (Angle exceeded AND Y-distance collapsed) OR (3+ factors simultaneously exceeded).

Detection logic thresholds

Factor	Detection Logic	Threshold
Body Angle	Torso deviation from vertical	>50°
Velocity	Downward movement speed	>0.15 units/frame
Aspect Ratio	Height/width of bounding box	<0.8 (horizontal)
Y-Distance	Upper vs lower body height	<0.1 (same level)
X-Spread	Horizontal keypoint spread	>0.4 (sprawled)

Get started

Repository: https://github.com/fidel-makatia/Elderly_Fall_Detection_on_ETHOS_NPU 

Quick start

1. Clone the repository to your Ethos-U enabled board.
2. Install dependencies: pip3 install numpy opencv-python flask.
3. Verify NPU is available: ls /dev/ethosu0.
4. Run the system: python3 gui/fb_display.py

Resources

• Arm Ethos-U NPU documentation
• Vela Compiler for Ethos-U
• TensorFlow Lite for Microcontrollers
• Arm NN SDK
• ML Inference Advisor

About the author

Fidel Makatia, PhD, Texas A&M University Distinguished Arm Ambassador, specializes in integrated Circuits design and the deployment of optimized machine learning models on Arm-based hardware. A former Autodesk Software Engineer, he possesses deep technical expertise in hardware IP , including SerDes and NPUs, alongside a strong background in embedded systems.