Unsupervised Identification of Driving Styles from Naturalistic Driving Data Through a Context-Normalized Framework.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Automotive Engineering

Background:

• Personalizing driving assistance systems requires accurate driving style identification.
• Current methods rely on subjective feature engineering and labeled data, limiting use of unlabeled naturalistic driving data.

Purpose of the Study:

• To propose an unsupervised framework for driving style identification from naturalistic driving data.
• To extract objective and interpretable driving style representations without labeled data.

Main Methods:

• Developed a Constrained Convolutional Autoencoder (CCAE) with self-attention for context normalization against the Worldwide Harmonized Light Vehicles Test Cycle (WLTC).
• Extracted Driving Adaptability Characteristics (DACs) and refined them using frequency-domain analysis.
• Utilized kernel-mapped clustering for driving style partitioning and validated with accident records and CAN-bus data.

Main Results:

• Identified driving styles showed varying historical accident probabilities.
• The CCAE model demonstrated superior cluster differentiation compared to traditional methods.
• Ablation analysis confirmed the effectiveness of WLTC-based context normalization.

Conclusions:

• The proposed context-normalization framework successfully extracts interpretable and externally relevant driving style representations from unlabeled naturalistic driving data.
• This method enhances objectivity and enables the use of large unlabeled datasets for driving style analysis.