Driving behavior recognition is an active research topic as it has many potential applications, such as fleet management, vehicle anti-theft, and planning of car insurance policies. Nowadays, the most successful approaches to driving behavior recognition are based on machine learning algorithms. Each machine learning algorithm has its pros and cons, and no single algorithm fits all problems. Therefore, how to determine an appropriate algorithm suitable for discovering driving patterns is a critical step in driving behavior recognition. This paper aims to conduct an empirical study for driving behavior recognition and evaluate the recognition performance of popular machine-learning algorithms. The experimental results showed that many sensor values gathered from the CAN bus are either highly correlated with one another or less important attributed to driving behavior identification. Among traditional machine learning approaches, ensemble tree-based algorithms, such as Random Forests and Decision Trees have better performance when compared with other approaches.
