Existing insurance telematics systems function as passive data recorders, assessing driver risk based on historical driving behaviour to calculate insurance premiums. However, such systems lack mechanisms for active risk prevention.

In this work, I developed an enhanced insurance telematics system capable of classifying dangerous driving behaviours in real time while integrating active risk prevention through speech-based alerting.

Driving Behaviour Classification

To identify dangerous driving behaviours, CARLA was used to collect a dataset of hazardous driving events in a controlled environment. XGBoost was then employed to classify these events into harsh acceleration, harsh braking, and sharp cornering. For feature engineering, both the Enveloped Power Spectrum and the Fourier Transform were applied to extract informative frequency-domain representations of vehicle motion.

The Device

Detection of Traffic Signs and Vehicles in Front

To detect potential hazards, a forward-facing camera was used to identify traffic signs such as “bends left” and “roundabout”, as well as to determine whether the driver was following the vehicle ahead too closely. A text-to-speech module delivered spoken alerts when the driver approached bends, roundabouts, or give-way signs at excessive speed, or when tailgating behaviour was detected.

To do so, I collected and annotated a dataset of approximately 3,000 images and trained YOLOv11 with a genetic evolution hyperparameter optimisation strategy. Vehicle distances were calculated based on the average width of a vehicle and the focal length of the camera. To detect speed, a GPS module was used.

Dashboard

To analyse driving behaviour, I implemented a dashboard using React with an Express back-end to receive data from a PostgreSQL database. This dashboard is designed to provide detailed feedback to the user with breakdowns of areas to improve upon in their driving.