PyroGuardian is an end-to-end computer vision ecosystem designed for real-time fire detection, risk assessment, and emergency reporting. It integrates high-performance AI at the edge with a user-friendly cloud-connected dashboard.
graph TD
%% Actors
User((Public User))
Admin((Admin / Firefighter))
%% Edge System (UAV / Fixed)
subgraph Edge ["Edge Computing"]
Source[Drone Camera] --> AI{AI Engine}
AI -->|Baseline| YOLO["YOLOv5, ( MobileNetV2 -User Uploads classification)"]
AI -->|Research| RT[RT-DETR]
YOLO & RT --> Stream[Live Annotated Feed]
YOLO & RT --> Logic[Severity & Analytics]
Logic -->|High Risk| SNS[AWS SNS Alert]
Logic -->|Telemetry| Data[Analytics & Evidence]
end
%% Cloud Infrastructure
subgraph Cloud ["Management Platform"]
Data ==>|HTTP POST| API[Node.js API]
User -->|Manual Report| API
API --> DB[(MongoDB)]
DB --> Dash[Integrated Dashboard]
end
%% Integrated Communications
SNS -.->|SMS / Email| Admin
Stream ==>|RTSP| Admin
Dash -->|Monitor & Control| Admin
-
1st Runner-Up – Honeywell Building Technologies Drone Hackathon 2024.
- Recognized for "Mission-Critical Fire Detection" at the Honeywell Bangalore Campus.
- Awarded for excellence in UAV-Edge integration and real-time severity classification.
| Metric | Baseline (YOLOv5) | Optimized (RT-DETR) |
|---|---|---|
| Inference Latency | 250ms | 33ms |
| Throughput | 4 FPS | 30 FPS |
| Precision | FP32 | FP16 / INT8 |
| Edge Hardware | Desktop CPU/GPU | NVIDIA Jetson Nano |
detection-engine/: Core AI modules optimized for Jetson.core/: Alerting logic & AWS SNS/Twilio integration.deployment/: DeepStream GStreamer pipelines.training/: NVIDIA TAO specifications for RT-DETR.data-engineering/: Dataset curation & augmentation tools.research-baseline/: Original YOLOv5 reference implementation.
web-dashboard/: Full-stack Node.js reporting application.
- Platform: NVIDIA Jetson Nano
- Pipeline: GStreamer (DeepStream SDK), Python bindings (
pyds) - Models: RT-DETR (TensorRT Optimized), YOLOv5 (Baseline)
- Tools: OpenCV, PyTorch
- Runtime: Node.js (v18), Express.js
- Database: MongoDB, Mongoose ODM
- Frontend: Vanilla JS, Leaflet.js, Chart.js
- Security: Helmet, Rate-Limiting, Bcrypt
- Containerization: Docker (Multi-stage, Non-root security)
- CI/CD: GitHub Actions (Automated Testing & Linting)
- Cloud: AWS SNS, Twilio
This project adheres to production-grade software engineering practices:
- CI/CD Pipeline: Automated testing for both Python and Node.js environments on every push.
- Code Quality: Enforced via
ruff,black(Python) andeslint,prettier(JavaScript). - Security: Hardened Docker containers, API rate limiting, and strict environment variable management.
For component-specific details, development guides, and architecture deep-dives, please consult the respective documentation:
- 🤖 Detection Engine Documentation
- For: AI Engineers, Edge Deployment, Model Training (TAO), GStreamer Pipelines.
- 💻 Web Dashboard Documentation
- For: Full-stack Developers, API Specs, Database Schema.
The fastest way to spin up the entire PyroGuardian ecosystem (Dashboard + Database + Detection Engine) is via Docker Compose.
- Docker Desktop installed and running.
Create the necessary environment files from templates:
# Root environment (optional override)
cp detection-engine/.env.example detection-engine/.env
cp web-dashboard/.env.example web-dashboard/.envdocker-compose up --build- Web Dashboard: http://localhost:8000
- MongoDB:
mongodb://localhost:27017
If you prefer to run components individually for development, please refer to the specific installation guides:
- AI Engine: Follow the Detection Engine Setup.
- Web Dashboard: Follow the Dashboard Setup.
This project implements a Targeted Augmentation strategy to handle high-occlusion fire scenarios (smoke, fog). By distilling knowledge from an 86.7M parameter RT-DETR model into a TensorRT-optimized engine, we achieve near-lossless accuracy at a 90% speed improvement on edge hardware.