Every second counts.

Turn ordinary cameras into vigilant firefighters with AlexNet-powered computer vision that detects flames and smoke in real time and immediately alerts authorities.

Edge or Cloud24/7 Monitoring< 500 ms LatencyRetrofits CCTV
Real‑time Fire/Smoke Detection · Alerts via API/SMS · Temporal Smoothing

The cost of delay

Traditional smoke alarms or human monitoring often detect fires too late. Cameras already capture early signs, but no one watches them 24/7 — leading to delayed responses, false alarms, and avoidable losses.

  • $50B+ annual fire damage globally
  • 15 min typical response time
  • 350K U.S. house fires each year

What success looks like

Early visual detection + instant alerts → faster dispatch, fewer losses, safer communities.

The solution: FireSentinel

AlexNet-based computer vision detects flames and smoke in live video feeds. When confidence crosses a safety threshold, it instantly notifies authorities and facility staff.

Standalone device

Plug‑and‑play camera unit for new installs and remote sites.

Software integration

Drop‑in module for existing CCTV and smart camera systems.

Smart alerting

APIs, SMS/Email, and optional on‑site siren/strobe. Tuned for ≥95% recall.

System architecture

  1. Frame extraction & preprocessing
  2. Fire detection via AlexNet CNN
  3. Confidence scoring & temporal smoothing
  4. Alert trigger via dispatch API/SMS

Deploy anywhere: Edge device (Jetson/RPi) for low latency, or Cloud API for centralized CCTV.

Data logging & auditing

Each detection stores timestamp, location, and a frame snippet for review and model improvement.

Experiments & results

  • Transfer learning depth: head vs partial vs full fine‑tune
  • Augmentation ablations & class imbalance handling
  • Cross‑entropy vs Focal loss (γ∈{1,2})
  • Resolution/latency trade‑offs (224↘128)
  • Threshold tuning for high‑recall operation
  • Temporal smoothing (EMA / majority vote)
  • Grad‑CAM explainability on detections

Metrics (placeholder)

Primary: PR‑AUC / AUROC. Secondary: F1@operating threshold, latency (ms/frame), false alarms/hour.

PR curve placeholder ROC curve placeholder

Real‑time inference

< 500 ms per frame @224×224 on Jetson Nano / standard GPU; ~20 FPS per feed.

Edge + Cloud

On‑device detection with optional cloud console for fleet monitoring, history, and remote thresholds.

Deployment‑ready

Export via TorchScript/ONNX; optional INT8 quantization for edge accelerators.

Security, privacy & reliability

  • TLS/SSL encrypted traffic
  • On‑device inference keeps raw video local
  • Fail‑safes: offline buffers & redundant alert relays
  • Self‑diagnostics & health checks
  • Designed toward NFPA 72 & GDPR readiness

Integrations

RTSP · ONVIF · HTTP · Vendor SDKs · REST APIs for dispatch and third‑party alerting.

Phase 1 — Prototype

Model training & offline testing.

Phase 2 — Pilots

Field deployments with municipal fire departments and facility partners.

Phase 3 — Rollout

Scale manufacturing and integrations with camera vendors.

See it in action

Request a demo or collaborate on a pilot deployment.