AI Traffic Accident Severity Prediction System

📋 Project Overview

AI Traffic Accident Severity Prediction is an intelligent system designed to predict the severity level of traffic accidents in real-time based on multiple environmental, vehicular, and driver-related factors. The system leverages machine learning algorithms to assess accident risk and provide actionable safety recommendations.

🎯 Objective

Develop a production-ready ML-powered application that:

• Predicts accident severity (Minor, Serious, Fatal)
• Analyzes distance-based hotspot risk zones
• Provides real-time travel recommendations
• Integrates spatial analysis for accident-prone locations across Asia

---

🏗️ System Architecture

┌─────────────────────────────────────────────────────────────┐
│              AI Traffic Accident Prediction                  │
├─────────────────────────────────────────────────────────────┤
│                                                               │
│  Frontend (React.js)          Backend (Flask)                │
│  ├─ PredictionForm            ├─ ML Model (joblib)           │
│  ├─ ResultPage                ├─ Feature Encoding            │
│  └─ ResultCard                ├─ Hotspot Analysis            │
│                                └─ Risk Scoring (1-5 scale)   │
│                                                               │
│  Database: CSV-based Hotspot Data (Asia)                     │
│  Model: RandomForest / XGBoost Classification                │
└─────────────────────────────────────────────────────────────┘

---

💾 Dataset & Preprocessing

Data Source

• Real-world traffic accident records from India
• ~10,000 accident records with 25+ attributes
• Geospatial data for 25+ accident hotspots across Asia

Key Features

Category	Features
Temporal	Hour, Day of Week, Rush Hour Flag, Weekend Flag
Environmental	Weather, Road Type, Road Condition
Vehicle	Vehicle Type, Vehicle Condition
Driver	Age, License Status, Experience (years), Speed Habit
Safety	Lighting Conditions, Traffic Control, Alcohol Involvement

Data Preprocessing Steps

1. Cleaning: Removed duplicates, handled missing values
2. Feature Engineering: Created risk scores (1-5 scale) for subjective features
3. Encoding: Categorical variables mapped to numerical codes
4. Validation: Removed outcome variables (Speed Limit, Casualties, Fatalities) to prevent data leakage

Output: Severity Classification

• 0 = Minor → Low injury accidents
• 1 = Serious → Moderate to serious injuries
• 2 = Fatal → Fatal accidents

---

🧪 Experiments & Model Training

Feature Scoring System

Experience Score (1-5 Scale)

• 5: < 6 months → Extreme Risk
• 4: 6-12 months → Very High Risk
• 3: 1-2 years → High Risk
• 2: 2-5 years → Medium Risk
• 1: 5+ years → Low Risk

Weather Risk Score (1-5 Scale)

• 5: Storm, Tornado, Hurricane → Extreme Risk
• 4: Heavy Rain, Snow, Hail → Very High Risk
• 3: Rain, Fog, Mist → High Risk
• 2: Cloudy, Drizzle → Medium Risk
• 1: Clear, Sunny → Low Risk

Driver Age Risk Score (1-5 Scale)

• 5: < 20 years → Very High Risk
• 4: 20-25 or > 70 → High Risk
• 3: 60-70 → Medium Risk
• 1: 25-60 → Low Risk

Speed Habit Score (1-5 Scale)

• 1: < 60 km/h → Safe/Moderate
• 2: 60-80 km/h → Slightly Risky
• 3: 80-100 km/h → Risky
• 4: > 100 km/h → Very Risky

Vehicle Condition Score (1-5 Scale)

• 5: Poor/Bad/Terrible → Extreme Risk
• 3: Average/Medium/OK → Medium Risk
• 1: Good/Excellent → Low Risk

Model Selection & Comparison

Model	Type	Advantages
RandomForest	Ensemble	Fast inference, handles non-linear patterns
XGBoost	Gradient Boosting	Superior performance on imbalanced data
Logistic Regression	Linear	Baseline for comparison

Selected Model: RandomForest/XGBoost (Best performance on validation set)

Hyperparameter Tuning

• GridSearchCV for parameter optimization
• Cross-validation (5-fold) for robust evaluation
• Class weight balancing for imbalanced dataset

---

🔧 Technical Stack

Backend

• Framework: Flask (Python)
• ML Libraries: scikit-learn, XGBoost, joblib
• Data Processing: pandas, NumPy
• Geospatial: DBSCAN Clustering,Distance-based hotspot analysis
• API: RESTful with CORS support

Frontend

• Framework: React.js
• Styling: Bootstrap 5 + Custom CSS
• State Management: React Hooks (useState)
• Routing: React Router v6
• HTTP Client: Axios

Deployment

• Backend: Flask development server (Production-ready with gunicorn)
• Frontend: Node.js + npm
• Database: CSV-based (Scalable to PostgreSQL)

---

📊 Key Features

1. ML-Based Severity Prediction

• Real-time accident risk classification
• Probability scoring (0-100%)
• Risk level mapping (Low/Medium/High)

2. Distance-Based Hotspot Analysis

• Proximity-based risk assessment
• 5-zone classification system:
  • 0-10 km: 🔴🔴🔴 High Critical Red Zone
  • 10-50 km: 🔴🔴 Critical Zone
  • 50-150 km: 🟡 Medium Risk Zone
  • 150-300 km: 🟢 Low Risk Zone
  • 300+ km: ✅ No Risk

3. Combined Recommendations

• ML prediction + Hotspot risk integration
• Travel permission system (ALLOWED/NOT ALLOWED)
• Real-time safety alerts

4. Interactive Hotspot Map

• Geospatial visualization of accident zones
• Dynamic map generation with Folium
• Risk zone color-coding

---

🚀 API Endpoints

Prediction Endpoint

POST /predict

Input: Accident parameters (time, location, weather, driver info, etc.)
Output: Severity prediction + probability + hotspot analysis

Hotspot Analysis

POST /hotspot_analysis

Input: Latitude, Longitude, Radius (km)
Output: Distance-based risk classification + nearby hotspots

Map Generation

POST /generate_hotspot_map

Output: Interactive HTML map with accident zones

---

📈 Model Evaluation Metrics

• Precision: Class-wise precision for each severity level
• Recall: Sensitivity for identifying high-risk scenarios
• F1-Score: Balanced performance metric
• ROC-AUC: Classification performance across thresholds
• Confusion Matrix: True/False positives and negatives

---

🔐 Data Quality & NaN Prevention

Robust Probability Handling

• Safe extraction from predict_proba() output
• NaN detection and fallback to 0.0
• Validation of array shape before indexing

Frontend Safety Checks

• Probability formatter handles undefined/null values
• Graceful fallback to "N/A" for invalid data
• Type conversion validation

---

📁 Project Structure

ai-traffic-prediction/
├── backend/
│   ├── models/
│   │   ├── best_model.joblib
│   │   ├── severity_mapping.joblib
│   │   └── feature_names.joblib
│   ├── data/
│   │   └── processed/
│   │       └── asia_accident_hotspots_enhanced.csv
│   ├── prediction_api.py
│   └── spatial_analysis.py
├── frontend/
│   ├── src/
│   │   ├── components/
│   │   │   ├── PredictionForm.js
│   │   │   ├── ResultPage.js
│   │   │   └── ResultCard.js
│   │   └── pages/
│   └── package.json
└── README.md

---

🎯 Usage

1. Start Backend

cd backend
python prediction_api.py
# Server runs on http://127.0.0.1:5000

2. Start Frontend

cd frontend
npm install
npm start
# App runs on http://localhost:3000

3. Make Predictions

1. Fill in accident details in the form
2. Submit for real-time prediction
3. View severity level + probability + recommendations
4. Analyze hotspot risks on interactive map

---

🔍 Example Prediction Input

{
  "datetime": "2025-11-10T13:00",
  "State Name": "Maharashtra",
  "City Name": "Mumbai",
  "Driver Age": 35,
  "Driver License Status": "Valid",
  "driver_experience": "8 years",
  "driver_speed_habit": 45,
  "vehicle_condition": "good",
  "Weather Conditions": "Clear",
  "Road Type": "Urban Road",
  "Road Condition": "Dry",
  "Lighting Conditions": "Bright",
  "Traffic Control Presence": "Lights",
  "Vehicle Type Involved": "Car",
  "alcohol_flag": 0
}

Output:

{
  "prediction_summary": {
    "severity": "Minor",
    "severity_code": 0,
    "ml_probability": 78.45,
    "ml_risk_level": "LOW"
  },
  "combined_risk": {
    "travel_safe": "YES",
    "combined_recommendation": "✅ Location safe - travel allowed"
  }
}

---

🛠️ Future Enhancements

• Multi-language support
• Mobile app development (Flutter/React Native)
• Real-time traffic data integration
• Police/Hospital location proximity analysis
• Historical accident trend analysis
• User feedback loop for model retraining
• Cloud deployment (AWS/GCP/Azure)
• Advanced geospatial clustering

---

📝 License

This project is open-source and available under the MIT License.

---

👥 Contributors

Project Leader: Biswajit Sow
Team: AI/ML Development Team
**Institution:**UEM