A deep learning approach for vehicle velocity prediction considering the influence factors of multiple lanes

Abstract

<abstract> <p>Predicting the future velocity of vehicles is essential for the safety of autonomous driving and the Intelligent Transport System. This study investigates how the surrounding vehicles influence a driving vehicle. Based on the HighD dataset, a scenario that considers the current lane and the neighboring lanes is selected while the drivers' visual angles and visual gap angles along with other parameters in the dataset are characterized as features. To predict the velocity of a driving vehicle and calibrate the influence of surrounding vehicles, a Transformer-based model integrating the features of multiple vehicles is proposed, and different features are added to the layers while constructing the model. Moreover, the information from previous timestamps of the vehicle state is integrated to estimate the duration of the influences, since the influence of an incident is not instantaneous. In our experiments, we find that the duration of the influence on the driving state perfectly fits the driver's reaction time when maneuvers occur in the surrounding vehicles. In addition, we further quantify the importance of the influence on the vehicle velocity prediction based on the Random Forest and obtain some practical conclusions, for instance, the velocity of a vehicle is more influenced by the front vehicle in the left lane than that in the right lane, but is still mainly influenced by the front vehicle in the current lane.</p> </abstract>