RescueVision AI — Intelligent Disaster Rescue System

AI system that analyzes disaster footage to detect survivors, distress gestures & damage zones — giving rescue teams a real-time priority map to save lives faster.

Comment

Inspiration

In the 2023 Turkey earthquake, rescue teams searched manually — covering only 0.3 km² per hour. Survivors died waiting. In India, the same problem repeats every year — Kerala floods, Uttarakhand landslides, Mumbai building collapses. We asked: what if AI could find survivors in 2 seconds?

What it does

RescueVision AI analyzes disaster footage — drone video, CCTV, or field images — and outputs a rescue priority map. It detects survivors, recognizes distress gestures like raised arms and SOS signals, maps damage zones, and tells rescue teams exactly which zone to respond to first.

How we built it

We built a full-stack AI system:

  • YOLOv8n for real-time person and damage detection
  • MediaPipe Pose for gesture recognition using 33 body landmarks
  • OpenCV for video frame extraction and processing
  • FastAPI backend with 6 REST endpoints
  • React dashboard with live detection viewer and priority heatmap

The system processes one frame in under 2 seconds. We tested it on real disaster images and got live detections — Person P-001, confidence 72%, Zone C CRITICAL.

Challenges we ran into

Integrating MediaPipe with Python 3.14 on Windows was the biggest challenge — the new API completely changed from older versions. We also had to build the priority scoring engine from scratch since no existing tool converts raw detections into a ranked dispatch order.

Accomplishments we're proud of

A fully working prototype with real AI detections. Not a concept — a system that actually runs, analyzes real disaster photos, and outputs actionable rescue priorities.

What we learned

How to combine multiple AI models into a single pipeline, how to build a production FastAPI backend, and how to design AI outputs that are actually useful for non-technical rescue workers.

What's next

Phase 2: Train on AIDER and FloodNet disaster datasets, add WhatsApp alerts for rescue teams, integrate GPS extraction from drone footage. Phase 3: Pilot with NDRF/SDRF teams and deploy on Raspberry Pi for offline field use.

Built With

Share this project:

Updates