Continuous Behavioral Prediction in Lane-Change for Autonomous Driving Cars in Dynamic Environments

Abstract

It is essential for autonomous driving cars to understand and predict other surrounding cars' behaviors, especially in urban environments, due to the high traffic volumes and complex interactions. Modeling the interaction among cars and their behaviors is challenging. The behavior estimation of a surrounding car serves as prior knowledge which helps the trajectory planner generate a path to perform properly with the other vehicles. It closes the gap between the high-level decision making and path planning. A new data-driven method is proposed to extend our previous behavior estimation. The new method predicts the continuous lane-change trajectory of a target car by modeling the interaction of all its surrounding vehicles' trajectories, without over-the-air communication between vehicles. The advantages of this approach are: 1. Learning the interactive model from real data; 2. Giving long-horizon estimation of the continuous trajectory of a target vehicle. The method is trained and evaluated on a public dataset. The experimental results show that the proposed method successfully predicts trajectories considering mutual interactions among cars, with low error based on the ground-truth.