README.md

🚦 TRAFIQ – AI Traffic Accident Detection System

Overview:

TRAFIQ is an AI-powered traffic monitoring system designed to automatically detect vehicle collisions from video footage.

The system uses computer vision and deep learning to:

Detect vehicles

Identify collisions

Capture accident snapshots

Generate structured accident reports (JSON)

✅ Features

• Vehicle detection using YOLO
• Collision detection using bounding-box overlap (IoU)
• Accident classification with trained crash model
• Automatic accident snapshot capture
• JSON accident report generation

🧠 AI Models

• Vehicle Detection Model

Model: vehicle-model.pt

Detects:

- Cars

- Trucks

- Buses

- Motorcycles

• Crash Detection Model

Model: crash-model.pt

Classes:

['0', '1', '2']

Only "2" class is considered a valid accident.

⚙️ Technologies Used

• Python

• OpenCV

• YOLOv8 (Ultralytics)

• NumPy

• JSON

• VSCode

📁 Project Structure

backend/
└── ai-engine/
    ├── detect_video.py
    ├── incidents_log.json
    └── models/
        ├── best_vehicle.pt
        └── best_crash.pt

▶️ How to Run

1️⃣ Install Dependencies pip install ultralytics opencv-python numpy

2️⃣ Run Detection (backend/ai-engine/) python detect_video.py 🎥 Input Video

Place video file inside:

backend/ai-engine/

Example:

accident0.mp4

Recommended resolution:

1280x720 📸 Output Accident Snapshot

When an accident is detected:

snapshots/accident_YYYYMMDD_HHMMSS.jpg

Example:

snapshots/accident_20260225_184233.jpg JSON Report

File created automatically:

accident_log.json

Example:

  {
    "timestamp": "2026-02-25 18:42:33",
    "snapshot": "snapshots/accident_20260225_184233.jpg",
    "confidence": 0.91
  }

🧪 Detection Logic

An accident is confirmed when:

1️⃣ Two vehicles overlap (IoU ≥ threshold)

AND

2️⃣ Crash model detects class:

2

AND

3️⃣ Detection is stable across multiple frames

This reduces false detections.

Sprint 1 :

✔ Vehicle detection

✔ Collision detection

✔ Crash classification

✔ Snapshot capture

✔ JSON logging

✔ Video processing