🚁 RescueVision Edge

Lightweight Sovereign AI for On-Device Victim Localization
in Post-Disaster Aerial Assessment

Universitas Darussalam Gontor (UNIDA)

What if a 7.8 magnitude earthquake levels a city — towers down, internet dead, no cloud, no API, no maps.
Search teams arrive with nothing but drones and a laptop.
They fly. The drone streams footage. A laptop runs our model.
Within seconds, every victim is located — offline, on-device, no connection needed.
That's what RescueVision Edge was built for.

---

📋 Ringkasan

RescueVision Edge adalah sistem deteksi korban bencana dari citra udara drone yang berjalan sepenuhnya luring tanpa cloud/API eksternal. YOLOv8n → ONNX (11.70 MB), CPU-only, latensi <40 ms.

Aspek	Detail
Model	YOLOv8n → ONNX (11.70 MB)
Inference	CPU-only via CPUExecutionProvider
Latensi	~30 ms rata-rata (max 38.1 ms)
Akurasi	[email protected] 0.5280 (pedestrian)
Frontend	React + Vite + Leaflet
Backend	FastAPI + ONNX Runtime
Offline	✅ Zero external API

---

✅ Constraint Track A

#	Constraint	Requirement	Status
C-A1	Ukuran Model	≤ 50 MB	✅ 11.70 MB
C-A2	Platform	CPU-only	✅ CPUExecutionProvider
C-A3	Kecepatan	≤ 3.000 ms	✅ 38.1 ms max
C-A4	Framework	PyTorch / ONNX	✅ Ultralytics + ONNX Runtime
C-A5	Offline	Zero API calls	✅ Fully offline

---

🎬 Demo

🐳 Docker compose up → deteksi

🔌 API OpenAPI → inject → GPS

⚡ Benchmark 8 frame → avg 157 ms

🎛️ Setup RPi scripts/setup_pi.sh

 

📍 Frontend  •  API  •  Swagger UI

---

🚀 Quick Start

# Docker (recommended)
cp model/best.onnx model.onnx
docker compose up --build     # → localhost:3000

# Development (2 terminal)
cd backend && pip install -r requirements.txt && uvicorn app.main:app --reload --port 8000
cd frontend && npm install && npm run dev -- --port 5173

# Raspberry Pi 4
sudo ./scripts/setup_pi.sh   # auto-install semua dependensi

---

📡 API

Method	Endpoint	Fungsi
GET	/health	Status sistem + model
POST	/detect	Deteksi 1 gambar
POST	/detect/batch	Batch (max 100)
POST	/inject	Inject config dinamis
GET	/export/csv	Export CSV
GET	/export/json	Export JSON

curl -X POST "http://localhost:8000/detect?manual_lat=-7.34&manual_lon=110.45" \
  -F "file=@drone_test_frames/frame_0008.jpg"

---

📊 Benchmark

Metrik	YOLOv5n	YOLOv8n
[email protected]	0.4684	0.5280
ONNX size	7.49 MB	11.70 MB
CPU latency (max)	19.8 ms	38.1 ms

Detail: docs/architecture_comparison.txt · docs/demo-benchmark.gif

---

🗂️ Struktur

RescueVision/
├── scripts/        # Utilitas: dataset, benchmark, setup RPi
├── backend/app/    # FastAPI: routes, inference, GPS
├── frontend/src/   # React: components, hooks, API client
├── docs/           # Demo GIF/Cast, proposal, laporan
├── notebooks/      # Training + inference notebook
└── docker-compose.yml

---

👥 Tim

Nama	Role
Farrel Ghozy Affifudin	DevOps & Frontend
Fatih Jawwad Al Mumtaz	AI/ML & Backend

---

📄 Lisensi

Dataset VisDrone-DET 2019 — riset non-komersial. Model YOLOv8n — AGPL-3.0.

---