Skip to content

mun3im/argus

Repository files navigation

ARGUS

Autonomous Real-time Guardian for Ultra-low-power Species Monitoring

ARGUS is an ultra-low-power, two-tier acoustic biologger designed for long-term wildlife monitoring in remote environments.

Inspired by Argus Panoptes --- the guardian who never slept --- ARGUS listens continuously at low power and activates high-intelligence processing only when a biological event occurs.

Goal: >98% recall of bird activity while enabling months of deployment on a single battery.


Why ARGUS?

Field monitoring systems face a difficult trade-off:

  • Continuous listening drains batteries
  • Aggressive power saving risks missing rare events
  • Storing everything creates massive data volume

ARGUS solves this with asymmetric intelligence:

  • A lightweight Sentinel listens continuously
  • A powerful Classifier wakes only when needed

No missed birds. No wasted power.


System Architecture

ARGUS runs on a dual-core microcontroller (Cortex-M4 + Cortex-M7).

Tier 1 --- Sentinel (Cortex-M4)

  • Always listening at ultra-low power
  • Lightweight bird activity detection
  • 3-second wake cycles via RTC
  • Activates Tier 2 only when bird presence is detected

Priority: Maximize recall.


Tier 2 --- Classifier (Cortex-M7)

  • Wakes only upon gatekeeper trigger
  • Performs species-level classification
  • Logs metadata or stores audio
  • Returns system to low-power state

Priority: Accuracy and efficiency.


ARGUS System Architecture


Core Components

1. MyBAD Dataset

  • 50,000 labeled 3-second clips (16 kHz)
  • Binary: Bird / No Bird
  • Optimized to prevent false negatives

https://github.com/mun3im/mybad


2. MyBADnet (Tier 1 Bird Activity Detector)

  • Custom 4-layer CNN (<8 KB INT8)
  • Input: 16×184 mel spectrogram (n_fft=512)
  • <8 ms inference on Cortex-M4
  • ≥99% recall target

3. MyGardenBird Dataset

  • 6,000 labeled clips
  • 10 Southeast Asian species
  • Ornithologist-verified

https://github.com/mun3im/mygardenbird


4. MynaNet (Tier 2 Classifier)

  • Depthwise Separable CNN + SE + Attention
  • Input: 80×300 mel spectrogram (n_fft=1024)
  • 92% top-1 accuracy
  • <45 ms inference on Cortex-M7
  • Unknown species → audio saved for review

https://github.com/mun3im/mynanet


5. Runtime System

  • Platform: Arduino Portenta H7
  • FreeRTOS-based scheduling
  • RPC inter-core communication
  • DMA audio pipeline
  • Average current: <85 µA (99.9% sleep duty cycle)

Deployment Context

Designed for Southeast Asian rainforest monitoring where:

  • Power is limited
  • Long-term autonomy is required
  • High recall is critical

ARGUS performs edge inference directly on embedded hardware --- no cloud required.


Advantages

  • Two-tier ultra-low-power design
  • ≥99% bird activity recall
  • On-demand species classification
  • Unknown species capture for dataset growth
  • Hardware-aware ML architecture

License

MIT


Citation

Publication forthcoming.

About

ARGUS: Autonomous Real-time Guardian for Ultra-low-power Species monitoring

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors