AIoTShipped
Wildfire Early-Warning Thermal Camera
Cảnh báo Cháy rừng qua Thermal Camera
FLIR Lepton 3.5 thermal camera + edge inference — abnormal hotspot alert within 12 seconds.
Raspberry PiFLIR LeptonLoRaWANPythonNumPyOpenCV
Problem
3 hilltop sites in the Central Highlands — no grid power, no reliable 4G. Dry-season avg 2 fires/month in the area. Late detection = tens of hectares lost + huge containment cost.
Architecture
FLIR Lepton 3.5 (160×120 thermal, 9 Hz) → Raspberry Pi 4 → 16×16 grid sliding-baseline (10 min) → ΔT > 8°C → confirm 3 consecutive frames → LoRaWAN alert → base station. 60 W solar + 100 Ah battery survives the wet season.
Stack & rationale
- Lepton 3.5: calibrated thermal — not fooled by overhead sun.
- Sliding baseline grid: auto-adapts to day/night + weather, no ML needed.
- LoRaWAN: 12 km range, no 4G needed, nearest public TTN gateway.
Results (6-month run)
| Metric | Value |
|---|---|
| Fires caught early (<5 ha) | 4/4 |
| False positives | 0 |
| Avg time-to-alert | 12 s |
| Uptime | 98.7% |
Lessons
Simple algorithms (threshold + sliding baseline) on specialised data (calibrated thermal) beat — and cost less than — CNN-on-RGB. Edge case: granite rocks heat up in noon sun → time-filter alerts to non-peak hours.