Integrating fuzzy trajectory data and artificial intelligence methods for multi-style lane-changing behavior prediction

TL;DR

This paper introduces a new AI framework that combines fuzzy trajectory data and machine learning to improve multi-style lane-changing behavior prediction using detailed vehicle trajectory data.

Contribution

It presents a novel integration of fuzzy trajectory data with supervised learning methods, accounting for different driving styles in lane-changing prediction.

Findings

The integrated approach outperforms traditional prediction methods.

Speed-related features are more influential after fuzzy processing.

Driving styles are better reflected in lateral movement than in lane-changing duration.

Abstract

Artificial intelligence algorithms have been extensively applied in the field of intelligent transportation, especially for driving behavior analysis and prediction. This study proposes a novel framework by integrating fuzzy trajectory data, unsupervised learning and supervised learning methods to predict lane-changing behaviors taking multi driving styles into account. The microscopic trajectory data from the Highway Drone Dataset (HighD) are employed to construct two types of datasets, including precise trajectory datasets and fuzzy trajectory datasets for lane-changing prediction models. The fuzzy trajectory data are developed based on different driving styles, which are clustered by the K-means algorithm. Two typical supervised learning methods, including random forest and long-short-term memory combined with convolutional neural network, are further applied for lane-changing behavior prediction. Results indicate that (1) the proposed integration approach performs better than the conventional lane-changing prediction; (2) the relative speed-related features have a greater contribution to the lane-changing prediction after being processed by fuzzy rules based on driving styles; and (3) the difference among driving styles is more reflected from the state of lateral movement rather than the lane-changing duration.