A two-dimensional lateral interaction crash risk evaluation model considering imbalanced data

Abstract

There are more than 539,000 crashes caused by vehicle-vehicle lateral interactions annually. To address this issue, proactively capturing lane-changing (LC) intentions and evaluating interaction crash potentials are a promising adopted approach. However, existing analyses were mainly conducted using balanced data, which is contrary to the fact that LC is a small probability event under natural driving conditions. Besides, previous crash risk evaluation methods mainly focused on the longitudinal conflicts, which have ignored the prevailing horizontal conflicts. To address the previous gaps, a two-dimensional lateral interaction crash risk evaluation model was proposed. This model considers both horizontal and longitudinal crash risks during LC interaction process, and modified focal loss function was introduced to deal with the imbalanced data for LC intention identification. The empirical analyses were conducted using Highway Drone Dataset (highD). Results showed that compared to the traditional loss function, the recall of the proposed model has been improved from 79% to 93%.

Keywords:

• crash risk evaluation
• lateral interaction
• imbalanced data
• long short-term memory model
• focal loss function

Authors’ contributions

Rongjie Yu: conceptualization, methodology, writing - review & editing. Ning Xie: data analysis, visualization, writing. Hui Zhang: conceptualization & review editing.

Additional information

Funding

This study was Sponsored by Shanghai Rising-Star Program (Nos. 52172349 and 71771174).