Why they take the risk to perform a direct left turn at intersections: A data-driven framework for cyclist violation modeling.

Bicycle crashes at intersection areas are posed a worrying traffic safety issue, and one of the main reasons for bicycle crashes is failing to avoid conflicts with motor vehicles and other bicycles. Clearly, cyclists are more exposed to risk if they perform a direct left turn (DLT) being mixed with left-turning vehicle under a left-turn phase. Owing to the lack of exposure data, the detection of DLT event and the mechanism behind the risky riding behavior have yet to be discovered. To bridge these gaps, this study proposes a DLT detection framework based on bike sharing trajectories. Moreover, this study seeks to understand the contributing factors to DLT behavior using the random parameters logit model with heterogeneity in means and variances (RPLHMV) to account for unobserved heterogeneity in the DLT cases dataset. Statistical analysis shows that DLT is most likely to occur on weekdays during peak periods under large commuting demand. As to what caused the DLT violations, law-obeying cyclists are more susceptible to external events, while risk-taking cyclists are subtly undermined by their habits. In addition, the model of RPLHMV reveals several significant contributing factors to the propensity of DLT violations, such as event time, available passing time for left-turning bicycles, and average cycling speed, whereas the indicator variables of actual waiting time, available passing space for left-turning bicycles, and preference for DLT violation become the emerging influential variables. This study is expected to help better understand DLT occurrence and propose countermeasures more efficiently for reducing cyclists' DLT rate.