Computer Vision + Environmental Sensors + AI Agent for Museum Asset Protection
Roni Bandini June 2026 MIT License
Traditional asset protection systems often rely on a single sensing modality, such as a Passive Infrared (PIR) motion detector. While inexpensive and widely deployed, PIR-based systems are prone to false positives caused by temperature fluctuations, reflections, electrical noise, and environmental conditions.
This project explores a different approach: combining computer vision, environmental sensing, and an AI reasoning agent running entirely on a low-cost edge AI computer.
The prototype continuously monitors a valuable object—in this case, an Omega wristwatch—and evaluates sensor data, object position, operating hours, and environmental conditions to determine the most appropriate response.
No cloud processing is required.
- 🎥 Object detection using Edge Impulse
- 👁️ Continuous asset monitoring
- 📍 Position tracking inside the camera frame
- 🌡️ Non-contact temperature monitoring
- 🚶 Human presence detection
- 🤖 AI-based reasoning with OpenClaw
- 🔒 Fully local processing
- ⚡ Runs on Qualcomm AI hardware
- 📲 Telegram and WhatsApp notifications
┌─────────────┐
│ PIR Sensor │
└──────┬──────┘
│
┌─────────────┐
│ RCWL-0516 │
└──────┬──────┘
│
┌─────────────┐
│ MLX90614 │
└──────┬──────┘
│
┌─────────────┐
│ USB Camera │
└──────┬──────┘
│
┌────────────────────┐
│ Rubik Pi 3 │
│ Edge Impulse Model │
└─────────┬──────────┘
│
┌────────────────────┐
│ OpenClaw │
│ AI Agent │
└─────────┬──────────┘
│
┌────────────┬────────────┬────────────┬────────────┐
│ Log Event │ Maintenance│ Security │ Authorities│
└────────────┴────────────┴────────────┴────────────┘
| Component | Purpose |
|---|---|
| Edge Impulse | Object detection |
| OpenClaw | AI reasoning and decision making |
| Rubik Pi 3 | Edge AI execution |
| PIR Sensor | Motion detection |
| RCWL-0516 | Microwave presence detection |
| MLX90614 | Infrared temperature monitoring |
| USB Camera | Visual monitoring |
| Specification | Value |
|---|---|
| SoC | Qualcomm Dragonwing QCS6490 |
| CPU Architecture | ARM64 / AArch64 |
| AI Accelerator | Hexagon NPU |
| GPU | Adreno 643 |
| AI Performance | Up to 12 TOPS |
| RAM | 8 GB LPDDR4x |
| Storage | 128 GB UFS 2.2 |
| Dimensions | 100 × 75 mm |
| Qty | Item |
|---|---|
| 1 | Thundercomm Rubik Pi 3 |
| 1 | Active Cooler |
| 1 | USB-C PD Power Supply |
| 1 | Logitech USB Camera |
| 1 | PIR Motion Sensor |
| 1 | RCWL-0516 Microwave Sensor |
| 1 | MLX90614 Temperature Sensor |
| 20 | Dupont Jumper Wires |
Detects motion by measuring changes in infrared radiation emitted by surrounding objects.
Unlike PIR sensors, the RCWL-0516 is an active sensor that emits microwave energy and detects changes in the reflected signal.
Advantages:
- Works in hot environments
- Detects movement regardless of temperature
- Less affected by ambient heat
Provides:
- Object temperature
- Ambient temperature
The sensor can detect:
- Human contact
- Object removal
- Tampering attempts
by comparing the watch surface temperature against ambient conditions.
OpenClaw receives:
- Asset presence
- Asset coordinates
- Motion events
- Human presence detection
- Temperature changes
- Museum operating hours
Example reasoning:
IF
Watch missing
AND museum closed
AND motion detected
AND human presence detected
THEN
Potential security incident
Possible actions:
| Action | Description |
|---|---|
| 📝 Log Event | Record incident |
| 🔧 Notify Maintenance | Possible equipment issue |
| 👮 Notify Security | Security review required |
| 🚨 Contact Authorities | High-confidence incident |
| Parameter | Value |
|---|---|
| Images | 80+ |
| Resolution | 96×96 |
| Labels | Bounding Boxes |
| Training Split | 90 / 10 |
| Epochs | 70 |
| Learning Rate | 0.001 |
- Create project
- Upload images
- Label watch
- Create Impulse
- Train model
- Test performance
- Deploy to Rubik Pi 3
sudo apt update
sudo apt-get install \
libportaudio2 \
libportaudiocpp0 \
portaudio19-dev \
--break-system-packages
sudo apt install python3-pyaudio
sudo apt install selinux-utils
sudo apt install fswebcam -y
sudo apt install -y sox libsox-fmt-all
sudo apt install python3-smbus
sudo apt install gpiodpip3 install edge_impulse_linux \
-i https://pypi.python.org/simple \
--break-system-packages
pip3 install "opencv-python>=4.5.1.48,<5" \
--break-system-packageslsusbls /dev/video*fswebcam -d /dev/video0 \
-r 1280x720 \
--no-banner test.jpgi2cdetect -a -y -r 1sudo edge-impulse-linux-runnerSelect the quantized model.
Qualcomm's Hexagon NPU supports quantized models. Float32 models will execute on the CPU.
Typical performance:
boundingBoxes 2ms. []
boundingBoxes 3ms. [{"label":"watch","value":0.70}]
Install:
curl -fsSL https://openclaw.ai/install.sh | bashConfigure:
- OpenAI
- Anthropic
- Ollama
- Telegram
Telegram pairing:
openclaw pairing approve telegram XXXXXWATCH_DEFAULT_X = 32
WATCH_DEFAULT_Y = 40
WATCH_POSITION_THRESHOLD_PCT = 20
CONFIDENCE_THRESHOLD = 0.85
MUSEUM_OPEN_HOUR = 9
MUSEUM_CLOSE_HOUR = 18
DEFAULT_OBJECT_TEMP = 22.0
DEFAULT_AMBIENT_TEMP = 22.0
TEMP_THRESHOLD_PCT = 15AI Asset Protection
Roni Bandini, Oct 2025, Argentina
Museum hours → 09:00 – 18:00
Watch default position → X:32 Y:40
Confidence threshold → 0.85
Object temp default → 22.0°C
Ambient temp default → 22.0°C
Stop with CTRL-C
A camera alone can be fooled.
A PIR sensor alone can generate false positives.
A temperature sensor alone lacks context.
By combining:
- Computer Vision
- Temperature Monitoring
- Microwave Detection
- Motion Detection
- AI Reasoning
the system can make significantly more informed decisions than any individual sensor.
- Edge Impulse
- OpenClaw
- Rubik Pi 3
- MLX90614
- RCWL-0516
MIT License
Copyright (c) 2026 Roni Bandini