Regional attention network with data-driven modal representation for multimodal trajectory prediction

Accurate and reliable prediction of vehicle trajectory is one of the core functions for automated vehicles, and it is a keystone for high-quality local motion planning and decision-making. However, due to the complex intervehicle interaction and the diversity of maneuvering classes, predicting future trajectory accurately is a challenging task. In this paper, we propose a novel regional attention network with data-driven modal representation for multi-modal trajectory prediction (RD-Net). In the encoding phase, we construct the intention-based regional attention mechanism to model complex inter-vehicle interaction more effectively. In the traffic scene, this mechanism can assign attention to different regions guided by the intention of the target vehicle and perform weighted aggregation of inter-vehicle interaction in different regions, which makes it possible to model non-local interaction at a larger range without introducing noise caused by unrelated surrounding vehicles. In the decoding phase, a data-driven method is proposed to learn the individualized future modal representation for each trajectory. This method comprehensively considers the commonality and individuality of the trajectory modes, which can guide the proposed model to predict multi-modal trajectory based on different lateral maneuver classes more accurately. Experimental results on two real-world datasets show that the average performance improvement of RD-Net is greater than 21.00%, with a maximum of 52.48%, compared with the existing state-ofart methods, which quantitatively illustrates the significant advantages of RD-Net. Additionally, various ablation experiments are conducted to evaluate the effectiveness of our proposed network components.