Smart Fire Detector

Overview

This repository hosts a YOLO-powered computer vision prototype that spots fire and smoke in real time. It bundles:

• webcam_inference.py – a live preview loop that highlights detections, mirrors consumer smart-home camera behavior, and plays an audible alarm when a confident fire signal appears.
• trainer.py – a GPU-first training harness that resolves dataset paths, resumes Ultralytics runs, and materializes a data.resolved.yaml for deterministic experiments.
• Roboflow-exported dataset stubs (train/, valid/, test/ and data.yaml) plus example checkpoints (yolo11n.pt, yolov8s.pt) so you can fine-tune or evaluate immediately.

Purpose

The codebase is meant as one of the early prototypes for a smart-home fire sensor:

• Validate whether RGB cameras can complement traditional flame/heat sensors before investing in custom hardware.
• Provide a reference implementation teams can wrap into a microservice, Home Assistant add-on, or edge gateway.
• Accelerate experimentation on alarm logic (e.g., what confidence threshold triggers sirens, when to mirror frames, how to log alerts).

Usage

1. Environment setup

1. Ensure Python 3.10+ and a CUDA-capable GPU (trainer requires CUDA; inference can fall back to CPU).
2. Install dependencies:

   pip install -r requirements.txt

3. (Optional) Drop any additional YOLO .pt checkpoints beside this README or under runs/*/weights/.

2. Run the webcam prototype

python webcam_inference.py \
  --source 0 \
  --mirror \
  --fire-alert-threshold 0.85 \
  --alarm-sound firealarm.mp3

• --source accepts a camera index, video path, or RTSP/HTTP stream.
• The script auto-selects the freshest runs/*/weights/best.pt; override with --weights path/to/model.pt.
• Audio alerts use playsound; set --alarm-sound to a custom MP3/WAV or --alarm-sound "" to disable playback.
• Press q or Esc to exit the preview window.

3. Train or fine-tune a model

python trainer.py \
  --data data.yaml \
  --weights yolov8s.pt \
  --epochs 50 \
  --batch 16 \
  --project runs \
  --name smart-flame-detector

• The trainer resolves the relative paths inside data.yaml and writes data.resolved.yaml automatically for Ultralytics.
• --device defaults to GPU 0; pass --device 0,1 for multi-GPU training.
• Checkpoints land under runs/<name>/weights/; pick best.pt for deployment and copy it beside webcam_inference.py if desired.

4. Integrate with a smart-home stack

While this repository focuses on detection, the outputs are designed for downstream automation:

• Listen to the log stream (stdout) or extend AlarmPlayer to call MQTT/webhooks when fire_alert_threshold is met.
• Wrap webcam_inference.py in a service that exposes the annotated frames over RTSP/WebRTC for monitoring panels.
• Export Ultralytics results to ONNX/TensorRT if the final hardware target is an embedded edge device.

Dataset note

data.yaml points to the Roboflow dataset firesensor-zwryw/fire-h2gkf-6mcdq (CC BY 4.0). Replace the train/, valid/, and test/ directories with your own captures to adapt the prototype to new environments or camera placements.

---

Use this README as the living blueprint for transforming the prototype into a production-ready smart-home fire sensor. Update it as you iterate on alarm logic, deployment targets, or dataset sources.