Fire and Smoke Detection from CCTV using CNN

Most fire damage happens in the minutes before anyone notices — warehouses, parking basements and factory floors burn while CCTV cameras watch silently, because nobody stares at every feed. This project gives those cameras eyes: a YOLOv8 detector fine-tuned on D-Fire (about 21,500 images with ~26,500 fire and smoke bounding boxes, including surveillance-camera footage) scans each frame for flames and smoke. A 3-frame persistence check suppresses flicker false-alarms; confirmed fire triggers the escalation chain — fire-panel signal, SMS to the supervisor list, and a logged event with the snapshot. The demo dashboard (the screenshots) shows a priority feed with fire/smoke boxes and confidence, a camera wall, a live event feed and a 24-hour alert history with response actions. Built with Ultralytics YOLOv8 and OpenCV, every stage — detection, confirmation, escalation — is explainable in the viva.

1. Each CCTV frame is resized to 640×640 and normalized for the YOLOv8-s network.
2. The network predicts bounding boxes for two classes — fire and smoke — with NMS (IoU 0.5) removing duplicates.
3. Detections below the confidence threshold (default 0.40) are discarded.
4. A detection must appear in 3 consecutive frames before it counts as confirmed, filtering transient glare and flicker.
5. Confirmed fire raises a critical alert: fire-panel signal, SMS dispatch and an event-log entry with the snapshot frame.
6. Smoke-only detections raise a warning-level alert for early investigation.
7. The dashboard polls the event log and renders boxes, the camera wall, the live feed and the alert history.

Tech stack:

• Python 3, PyTorch, Ultralytics YOLOv8
• OpenCV (frame capture, persistence logic)
• D-Fire dataset (~21,500 images, GAIA)
• ONNX Runtime (edge deployment)
• NumPy, Matplotlib (evaluation plots)
• HTML/CSS/JS single-file dashboard (this demo)

Dataset & model details

• Dataset: D-Fire (Sharma et al., GAIA) — 21,527 images labeled fire / smoke / both / neither, with 14,692 fire and 11,865 smoke bounding boxes in YOLO format; imagery spans regular, aerial and surveillance cameras.
• Task: Two-class object detection; input = 640×640×3 frame, output = fire/smoke boxes with confidence, then temporal confirmation and escalation.
• Model: YOLOv8-s (approximately 11.2M parameters), fine-tuned; NMS IoU 0.5; confidence threshold 0.40; 3-frame persistence rule.
• Metrics: mAP@0.5 0.80, fire-class precision 0.92, smoke recall 0.85, under 2 false alarms/day — design targets for the built-to-order training run. The report documents the actually achieved figures.

• [Fire + smoke detection] YOLOv8 fine-tuned on D-Fire detects flames and smoke separately — smoke-only alerts give the earliest possible warning.
• [3-frame confirmation logic] A detection must persist across 3 consecutive frames before escalation, suppressing single-frame flicker false-alarms.
• [Escalation chain] Confirmed fire triggers a fire-panel signal, SMS to the supervisor list and a logged event with snapshot — the full response workflow.
• [Live monitoring dashboard] Priority camera feed with detection boxes, a camera wall, and a scrolling event feed in one single-file web UI.
• [Alert history & analytics] 24-hour event log with severity, confidence and recorded response actions; fire/smoke counts and detect-to-SMS latency.
• [Tunable sensitivity] Confidence and persistence thresholds are operator-adjustable; the false-alarm trade-off is measured in the report.
• [ONNX edge export] Weights export to ONNX for pole-side or on-premise edge inference without cloud dependency.

• Fine-tuned YOLOv8-s weights and the complete training notebook
• Detection + persistence + escalation Python pipeline
• Single-file HTML monitoring dashboard wired to pipeline outputs
• Training curves, mAP plots and false-alarm analysis for the report
• Project report PDF (background, dataset, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (detection metrics, temporal confirmation, false-positive control)