🚗 SafeVision AI — Real-Time Road Hazard Detection & Driver Alert System

Built for i.Mobilothon 5.0 | Team DevSphere

Transforming every dashcam or smartphone into an intelligent road-safety sensor.
Using AI, geospatial analytics, and privacy-preserving design, SafeVision AI detects road hazards in real time, anonymizes video, and creates a self-healing live hazard map for safer mobility.

---

🧠 Problem Statement

Issue	Description
Unsafe Roads	Indian roads expose drivers to unmarked speed breakers, potholes, debris, and stalled vehicles.
Data Waste	Although dashcams capture valuable data, it rarely becomes actionable intelligence.
Objective	Transform camera feeds into actionable, verified hazard data to prevent accidents.

---

💡 Our Solution — RoadSentry Engine

Component	Description
AI Detection	Detects potholes, speed breakers, and obstacles using YOLOv8.
Privacy Layer	Automatically blurs faces & license plates using OpenCV + MediaPipe.
GeoTagging	Adds latitude-longitude metadata for every hazard detected.
Consensus Engine	Filters false positives using DBSCAN clustering and cross-verification.
Driver Alerts	Sends instant hazard notifications to nearby users.

---

🧩 Core Features

Feature	Description
🚧 Hazard Detection	Real-time detection of potholes, bumps, debris.
🔒 Privacy Preservation	On-device anonymization via face and plate blurring.
🌍 Geospatial Intelligence	Automatic location tagging via GPS metadata.
⚡ Low Latency	Optimized YOLOv8 models ensure <100ms per frame.
🧭 Crowdsourced Validation	Duplicates filtered using DBSCAN clustering.
🛰️ Cloud Sync	Centralized live hazard map for city-wide visibility.

---

⚙️ Tech Stack

Layer	Technology
AI Models	YOLOv8 (Hazard Detection), YOLOv8 (License Plate), MediaPipe (Face Detection)
Backend	FastAPI, Python, OpenCV, PostGIS, WebSockets
Frontend	React + Vite + TailwindCSS
Deployment	Docker, GitHub Actions (CI/CD), Google Colab for model training
Data Pipeline	Roboflow (dataset prep), DBSCAN (clustering), JSON REST API

---

🏗️ System Architecture

Step	Process
1️⃣	Camera captures road footage.
2️⃣	YOLOv8 detects potholes, speed breakers, debris.
3️⃣	Faces & plates blurred locally (MediaPipe + OpenCV).
4️⃣	Data packaged with GPS + timestamp into JSON.
5️⃣	FastAPI backend validates & stores in PostGIS.
6️⃣	Live map shows verified hazards for nearby drivers.

---

📂 Folder Structure

Folder	Description
Backend/	FastAPI backend, ML inference, database integration
Frontend/	React-based dashboard and driver alert interface
models/	YOLOv8-trained models (road_hazard.pt, plate.pt)
Demo/	Demo script and sample input/output visuals
assets/	Architecture diagram, screenshots, and videos

---

🚀 Quick Start

Step	Command
1️⃣ Clone Repository	git clone https://github.com/jeetgoyal80/SafeVision-AI.git
2️⃣ Install Backend Deps	pip install -r requirements.txt
3️⃣ Run Backend	uvicorn app.main:app --reload
4️⃣ Launch Frontend	cd Frontend && npm install && npm run dev
5️⃣ Test Demo	python Demo/demo.py

---

🧠 Model Training Reference

Model	Dataset	Framework
road_hazard.pt	Pothole & Speed Breaker Dataset (Roboflow)	YOLOv8
plate_detection.pt	Indian License Plate Detection Dataset (Roboflow)	YOLOv8
face detection	Google MediaPipe	Built-in

Training Platform: Google Colab (Free GPU)
Optimization: Trained models exported to .pt format for lightweight, real-time inference.

---

🔒 Privacy by Design

Step	Security Measure
1️⃣	Detect faces & plates locally.
2️⃣	Apply Gaussian blur before upload.
3️⃣	Only anonymized frames transmitted.
4️⃣	No raw video or PII stored on server.

---

📊 Key Innovations

Innovation	Impact
Consensus-based Verification	Ensures accuracy by cross-matching multiple reports.
Self-Healing Maps	Automatically removes outdated hazards.
Zero-Tap Operation	Fully automated — no manual input needed.
Edge-first Design	Reduces backend costs and latency.

---

🎥 Media & Demo

Type	Description
🖥️ Dashboard Preview
📡 Live Feed Page
🗺️ Map View (Alert Visualization)
💻 Input Feed Sample
✅ Detected Output Sample
🎥 Demo Video	▶ Watch Demo on YouTube

🎬 Recommended Demo Video Content

• 30–60 sec live prototype demo
• Road hazard detection in real time (potholes / speed breakers / obstacles)
• Split-screen showing Raw Feed vs Processed Feed
• License-plate & face blurring demo
• Real-time alerts reflected on frontend map dashboard

---

💰 Implementation Cost Overview

Stage	Cost	Description
Prototype	₹0	Built using open-source tools
Pilot (10k users)	< ₹15,000/month	Free-tier cloud + microservices
Scaling	Cost-per-user ↓	Efficient edge inference model

---

🧾 References

Paper / Source	Link
Enhanced YOLOv8 for Real-Time Pothole Detection	arXiv
Vision-Based Pothole Detection Review	MDPI
YOLO-Based License Plate Recognition	MDPI
MediaPipe Face Detection Docs	Google Docs

---

👥 Team DevSphere

Member	Role	Contribution
Jeet Goyal (Team Lead)	ML & Backend Engineer	Developed core AI pipeline, handled training, backend, and integration
Neelam Patidar	Research & Presentation Lead	Prepared project documentation, research synthesis, and designed the final presentation for submission

---

🏁 Status

Milestone	Status
MVP Completed	✅
Backend + ML Integrated	✅
Frontend Dashboard	✅
Privacy Pipeline	✅
Deployment Ready	🚀

---

🏷️ Badges

Tech	Badge
Python
FastAPI
YOLOv8
React
TailwindCSS
Google Colab

---

🧭 Future Scope

Feature	Description
🎧 Audio Alerts	Bluetooth or CarPlay integration for live hazard warnings
🌧️ Weather Adaptive Models	Adjust detection under rain/fog conditions
🏙️ Government API Integration	Sync with municipal dashboards for road repairs
📱 Mobile Companion App	Allow user-side feedback & road status updates

---

© 2025 Team DevSphere | Built for Volkswagen i.Mobilothon 5.0