Deep learning-based lane change trajectory prediction

Abstract

For fully automated driving vehicles, lane change is one of the most common and important driving scenarios, and there are many accidents caused by lane change. For a vehicle in motion, if it cannot intelligently respond to its surrounding vehicle’s lane change, risks or a poor passenger experience may arise. To address the potential risk of such scenario, autonomous driving systems(ADS) need to have a better understanding of the lane change maneuver. Specifically there are two aspects, one is lane change intention, which involves determining whether a vehicle will change lanes and in which direction, and the other is the lane change process, which is most directly represented by the trajectory. Although many approaches are proposed to predict lane change intentions, limited studies focus on lane change trajectory prediction. In this paper, we propose to bridge the gap between lane change scenario and trajectory prediction. We utilize the real-traffic datasets from Next Generation SIMulation (NGSIM) and adopt the deep learning models to predict lane change trajectories of the target vehicle. On top of that, we conduct concise data processing to precisely extract lane change examples and select features considering both the vehicle’s own driving status and its surrounding traffic conditions. Extensive experiments are carried out to demonstrate that our deep learning-based approach performs better than some existing prediction models in different feature and prediction horizon settings.