Driving Behaviour Intelligence: Telematics Feature Engineering & Risk Modelling

Author: Augustine Osaigbevo

---

Table of Contents

• Project Description
• Repository Structure
• Instructions (How to Run)
• Technical Approach
• Results & Performance
• Challenges & Mitigations
• Next Steps
• Credits

---

Project Description

This project demonstrates the end-to-end development of telematics-based driving behaviour analytics and risk classification using real-world vehicle sensor data. Leveraging open Oxford RobotCar INS and GPS datasets, it delivers robust feature engineering, event detection, risk modelling with machine learning (LightGBM), and insightful visual analytics—providing a blueprint for modern, data-driven motor insurance and behaviour-based risk assessment.

Relevance:

• Behaviour-based insurance requires meaningful, explainable risk signals from real-world sensor data.
• Accurate detection of risky driving events (e.g., harsh braking, sharp turns) is key for fairer pricing, safer roads, and improved underwriting.
• The full, production-quality workflow is suitable for demonstration in a professional portfolio or integration into telematics products.

---

Repository Structure

data/
  ├─ raw/                             # Original Oxford dataset after download/unzipping
  ├─ processed/                       # Pre-processed, feature-engineered datasets (CSV)
output/
  ├─ driving_events_counts.png        # Bar chart of detected events
  ├─ feature_importance.png           # LightGBM feature importance plot
  ├─ feature_correlation_matrix.png   # Feature correlation heatmap
  ├─ risk_map.png                     # Route visualisation coloured by risk
  ├─ speed_distribution.png           # Speed distribution histogram
  ├─ acc_mag_distribution.png         # Acceleration magnitude histogram
src/
  ├─ data_downloader.py               # Data download & auto-traversal detection
  ├─ data_preprocessing.py            # Robust sensor loading, timestamp parsing, cleaning
  ├─ feature_engineering.py           # Telematics feature extraction from INS/GPS
  ├─ sensor_fusion.py                 # Heading, sharp turn, and fusion features
  ├─ model.py                         # LightGBM classifier, parameter tuning, saving
  ├─ utils.py                         # Plots: feature importances, route, etc.
notebooks/
  ├─ 01_exploration.ipynb             # Data download, inspection, loading
  ├─ 02_feature_engineering.ipynb     # Feature creation, EDA, event analysis
  ├─ 03_model_training.ipynb          # ML training, hyperparameter tuning
  ├─ 04_inference_reporting.ipynb     # Inference, visualisation, export
requirements.txt                      # Python dependencies

---

Instructions (How to Run)

1. Install dependencies

%pip install -r requirements.txt

2. Download and Prepare Data

Run the first notebook (01_exploration.ipynb) to:

• Automatically download and extract the Oxford RobotCar INS/GPS dataset.
• Detect traversal directory and load relevant sensor files (gps.csv, ins.csv).

3. Feature Engineering & Event Detection

Run 02_feature_engineering.ipynb:

• Computes telematics features (speed, acceleration, yaw rate, etc.).
• Detects events such as harsh braking, acceleration, and sharp turns.
• Produces exploratory visualisations (event counts, feature distributions).

4. Train the Risk Model

Run 03_model_training.ipynb:

• Trains a LightGBM classifier to predict “risky” segments based on engineered features.
• Hyperparameter tuning (GridSearchCV) for robust results.
• Saves model and feature importance plot.

5. Inference & Reporting

Run 04_inference_reporting.ipynb:

• Predicts risk on the full dataset.
• Visualises the driving route coloured by predicted risk.
• Outputs all key plots and analyses to the output/ directory.

---

Technical Approach

1. Sensor Data Loading

• Robust loaders for GPS and INS files (handles Oxford’s native formats).
• INS position, velocity, and orientation are the primary signal source; GPS is used for validation and mapping.

2. Feature Engineering (src/feature_engineering.py)

• Speed: Horizontal speed from velocity components.
• Acceleration: Derived by differentiating INS velocities.
• Yaw rate: Change in heading (from INS yaw) per unit time.
• Event detection: Threshold-based labelling for harsh braking, acceleration, and sharp turning.
• Risk label: “Risky” flag generated from event occurrence.

3. Sensor Fusion (src/sensor_fusion.py)

• Computes heading and heading change from yaw.
• Sharp turn detection based on heading dynamics.

4. Machine Learning (src/model.py)

• Model: LightGBM classifier with balanced class weights and grid hyperparameter search.
• Features: Only non-derived/raw features (no leakage from event labels into model features).
• Evaluation: Stratified train/test split; classification metrics reported.

5. Visual Analytics (src/utils.py)

• Event counts: Bar charts of driving events.
• Feature importance: LightGBM feature importances.
• Correlation matrix: Heatmap of key engineered features.
• Route mapping: Colour-coded risk overlays on the real route.

---

Results & Performance

Sample Outputs

Step	Example Output
Event counts
Feature importances
Correlation matrix
Route risk map
Speed distribution
Accel. distribution

Model Training

• Model hyperparameters selected via grid search: {'learning_rate': 0.05, 'max_depth': 4, 'n_estimators': 80, 'num_leaves': 7}
• Feature importance ranks acceleration, yaw rate, and speed as primary predictors.

Performance

• Classification metrics reflect model accuracy based on raw telematics features only.
• Visualisations provide insights into both feature relevance and real-world event distributions.
• Route visualisation highlights high-risk and low-risk segments for interpretability.

---

Challenges & Mitigations

Challenge	Approach/Mitigation
Label leakage risk	No event-derived columns (harsh_brake, etc.) used as ML features.
Non-standard sensor formats	Loader functions robust to Oxford’s unique CSV headers.
Small labelled sample	Used all available data; design allows easy extension to new data.
Class imbalance	Model uses class weighting; clear reporting of risk class support.
Urban/stop-start traffic	Speed/acceleration thresholds selected for robustness in UK driving.

---

Next Steps

• Incorporate additional traversals for larger training data and stronger generalisation.
• Integrate map-based features (e.g., road type, weather, speed limits).
• Deploy risk scoring as a real-time or batch analytics service (Databricks, cloud, etc.).
• Benchmark against other models (e.g., XGBoost, neural networks).
• Conduct feature ablation and SHAP analysis for deeper explainability.

---

Credits

• Oxford RobotCar Dataset
• LightGBM

---

Contact: augustine.osaigbevo@gmail.com