AQUASCOUT Get the APK

About the project

A fire alarm that is already in your pocket.

BSCS thesis · Saint Columban College, Pagadian City · 2026

The one-line version. Commercial smoke alarms cost money and need batteries. AQUASCOUT uses the camera and the model runtime that are already in the phone in your pocket, and spends them watching for flame and smoke.

Why it exists

Household fire is usually detected by a device that has to be bought, installed, tested and kept alive with a battery. In the Philippines most of them are simply not there. Meanwhile almost every adult in the country carries a phone with a camera, a capable CPU and a network connection.

That phone can do the watching. The hard part is not the alert, it is deciding whether a frame actually contains fire, on hardware with no GPU, no network round trip, and no appetite for a 300 MB model file.

How it was built

The model is YOLOv8n, quantised to dynamic INT8, which is what brings it down to 3.3 MB. It was trained and evaluated on D-Fire, a dataset of 21,527 real fire and smoke photographs, with a held-out test split of 4,306 images that the model never saw.

The reason D-Fire matters is that fire photographs are hard. In a typical D-Fire frame a flame covers a median of only 0.54% of the image, and 64% of all flames are under 1% of the frame. An alerting system that needs a big obvious fire to work is not an alerting system.

  • mAP50 of 0.748 on the held-out split, smoke 0.806, fire 0.689
  • 2.0 frames per second on a mid-range handset, entirely on-device
  • 21,527 images total, 17,221 for training and 4,306 held out

What it does not do

We would rather state the limits here than have you find them.

  • It does not predict a fire. It detects visible flame and smoke. The research literature is clear that the more desirable goal is sensing combustion products before they accumulate, not after.
  • It does not see an overheating appliance that has not yet produced flame or smoke. That is a thermal problem, and a thermal sensor such as an MLX90640 would cover it. That is designed, not built.
  • It cannot see through a blocked lens. Camera visibility is the whole input.
  • It is not a certified life-safety device. Do not replace working smoke alarms with it.

The rover

The optional rover is a LAFVIN 2WD chassis carrying an Arduino, a JDY-16 radio link and two ESP32-CAM heads, so the phone can keep watch from a desk while the camera sits in another room.

It is specified and partly wired, but not finished. The motor driver failed and is awaiting replacement, so no movement is claimed, and the water pump is not built. None of this is needed for the alert to work.

The team

Tampugao

Programmer · model, Android app, system design

Peñaranda

Project manager · alerting, escalation

Balbarino

Quality assurance · testing, results verification