A multi-model computer vision pipeline that detects fire, smoke, and unsafe conditions in real time using YOLOv8 and a FastAPI backend.
Fire detection today is still dominated by traditional smoke and heat sensors that react only after the situation becomes dangerous. They also provide no understanding of what type of fire is happening, how severe it is, or whether people nearby are safe. At the same time, computer-vision-based fire detectors that exist often trigger false positives because they cannot differentiate between harmless flames and actual hazards. This makes real-world deployment difficult.
ARIES (Advanced Real-Time Intelligence & Emergency System) aims to solve this gap by using computer vision to provide early, accurate, and context-aware fire detection. Instead of depending on specialized sensors, ARIES analyzes live video from any camera—CCTV, webcams, mobile phones—and identifies fire, smoke, and hazardous flame behavior in real time. The main goal is to make computer-vision-based fire detection practical by reducing false positives and introducing severity-based logic.
The idea behind ARIES is simple: use the cameras that already exist in buildings and infrastructure, and upgrade them with an intelligent, multi-model computer vision system that understands what kind of fire is happening and how dangerous it is. Unlike CV systems that treat all fire the same, ARIES distinguishes between:
Contained fire (low risk)
Hazardous fire (HZRD) with violent flames and heavy smoke
Smoke-only early warnings
This makes the system more reliable and significantly reduces unnecessary alarms.
ARIES uses a multi-model inference pipeline built on YOLOv8, optimized with a custom “gating” logic. The models include:
smokenfire.pt: A lightweight model that detects normal fire and smoke.
HZRD_fire_detection_best.pt: A custom-trained model focused on detecting high-risk flame behavior.
PPE.pt: Activated only when a hazard is present, used to detect firefighters and unprotected workers.
This design ensures the system stays fast and avoids running heavy models unnecessarily.
Frame Capture
Live video frames are captured using the browser (webcam or uploaded video).
Initial Detection
The lightweight model checks for fire and smoke.
Severity Analysis (Gating)
If fire or smoke is present, ARIES activates the HZRD model to check if the fire is dangerous or just contained.
Human-Hazard Correlation
If a hazardous fire is detected, ARIES activates the PPE model to see if:
People are nearby
They are wearing protective gear
A firefighter or response unit has arrived
Result Packaging
The backend returns bounding boxes, detected classes, severity alerts, and the percentage of the frame covered by HZRD/smoke.
Frontend Visualization
The UI overlays bounding boxes, displays hazard popups, and logs timestamps for every detection.
The frontend is built using HTML, JavaScript, and Canvas, ensuring low-latency rendering without heavy frameworks.
The browser sends frames to the FastAPI backend, which performs inference and sends back structured results.
The UI shows:
Live video stream
Processed output with bounding boxes
HZRD/smoke alerts
Area coverage
A real-time detection log panel
This design keeps the system responsive even when running on CPU-only environments, such as Hugging Face Spaces.
ARIES solves multiple real-world problems:
Faster detection compared to smoke/heat sensors
Lower false positives compared to standard CV fire detectors
Understands severity, not just presence of fire
Identifies if people are nearby and whether they are protected
Works with existing camera infrastructure, reducing costs
Edge-ready, capable of running on CPUs, embedded devices, and industrial hardware
Provides actionable insights rather than simple alarms
The system is particularly useful in:
Warehouses
Industrial plants
Chemical storage sites
Smart city infrastructure
Construction sites
Public safety facilities