Customizable multimodal trajectory prediction via nodes of interest selection for autonomous vehicles

To safely navigate through complex traffic scenarios, autonomous vehicles (AVs) must accurately predict the future trajectories of surrounding agents. Therefore, there has been a surge of interest in the problem of trajectory prediction for AVs. Building upon existing studies, we aim to push the boundaries of state-of-the-art research by tackling the following challenges: (1) the interaction between agents is heavily dependent on road geometry and topology; (2) certain modalities of the surrounding agent are non-informative for the AV and can be disregarded; and (3) the diversity of multimodal prediction is limited by the maximum number of modalities. In this study, we propose Customizable Multimodal Transformer (CMT), a deep learning model which facilitates customizable multimodal trajectory prediction. First, inspired by the dependency between agent interaction and road geometry and topology, we propose that map information can be utilized to better understand agent interaction. Furthermore, we propose the concept of nodes of interest (NOI), which represents the area of interest of the AV. By manipulating the nodes in the NOI, CMT can generate customized prediction results where irrelevant modalities can be disregarded without compromising the safety of the AV, leading to reduced computational costs. Finally, we propose to enhance the diversity of multimodal prediction results through Gaussian mixture reduction via clustering (GMRC). Extensive experiments on nuScenes and Argoverse datasets demonstrate that CMT not only outperforms previous state-of-the-art models, but also exhibits great potential for reducing computational costs and improving inference speed for trajectory prediction of AVs. Code is available at https://github.com/Promisery/CMT.