A Framework for Combining Lateral and Longitudinal Acceleration to Assess Driving Styles Using Unsupervised Approach

• Published in: IEEE transactions on intelligent transportation systems 2024-01, Vol.25 (1), p.1-19
• Main Authors: Zhu, Yuanfang, Jiang, Meilan, Yamamoto, Toshiyuki, Ding, Naikan, Shinkai, Hiroko, Aoki, Hirofumi, Shimazaki, Kan
• Format: Article
• Language: English
• Subjects: Accidents; Algorithms; Behavioral sciences; Cluster analysis; Clustering; Driver behavior; Driving style assessment; Feedback; Fuels; G-G diagram; Intelligent transportation systems; large-scale driving data; Machine learning; Machine learning algorithms; Maneuvers; Probabilistic models; Risk assessment; risk profile; Road safety; Statistical analysis; Statistical methods; unsupervised approach; Vehicle safety; Vehicles
• ISSN: 1524-9050; 1558-0016
• DOI: 10.1109/TITS.2023.3310213

Driving style assessment plays an important role in intelligent transportation system (ITS) applications, such as driving feedback provision and usage-based insurance. Many previous studies used supervised algorithms to profile drivers. However, this cannot be applied to large-scale unlabeled driving data, which are increasingly prevalent in the ITS context. This paper proposes a framework that combines lateral and longitudinal accelerations to assess a driver's driving style using an unsupervised approach. The framework first detects risky acceleration maneuvers using a statistical method based on the G-G diagram that shows combinations of lateral and longitudinal accelerations. Hierarchical clustering was used to classify the average driving behavior of drivers into high-, medium-, and low-risk groups. Further, a unique Gaussian mixture model is trained for each driver to score their driving style and decompose risky acceleration maneuvers into several risk components. Finally, the spatio-temporal characteristics are extracted to provide implicit factors on the risky behavior of drivers. The proposed method was applied to a large-scale dataset obtained by in-vehicle data recorders. The results demonstrate the necessity to combine the two axes of acceleration for driver behavior assessment. The proposed method can model individual driving styles effectively from the driver's G-G diagram, enabling applicability to large-scale unlabeled data for driving style assessment. The extraction of spatio-temporal characteristics can improve the interpretability of the results obtained using machine learning algorithms. All these results can be used to create a driver's risk profile and provide tailored feedback for improving driving safety.