STI-TP: A Spatio-temporal interleaved model for multi-modal trajectory prediction of heterogeneous traffic agents

Abstract

Trajectory prediction for heterogeneous traffic agents in autonomous driving is a challenging and crucial task. A large amount of research has laid a solid foundation for this field. However, achieving accurate trajectory prediction remains a great challenge. In this paper, we propose a spatio-temporal interleaved model for multi-modal trajectory prediction of heterogeneous agents. The model consists of the following novel components: (1) a M̲ulti-S̲cale H̲eterogeneous A̲gent I̲nteraction E̲ncoder (MS-HAIE), which adopts different temporal receptive fields for heterogeneous agents to match their different traveling speed and enhance the model’s expressive capability. (2) a C̲ross A̲ttention-guided S̲patio-T̲emporal I̲nterleaving M̲odule (CA-STIM), which combines the trajectory interaction information and independent temporal information of agents, so as to improve the spatial-dependent modeling capability of Transformer. (3) a M̲ulti-modal T̲rajectory D̲ecoder (MTD) to capture the multi-modality of traffic agents’ trajectories and strengthen the network’s comprehension and response capabilities. Our proposed method is evaluated on the Apolloscape and the Argoverse dataset, demonstrating superior performance over other state-of-the-art (SOTA) methods with a reduction of 12.48% WSADE and 26.04% WSFDE, respectively, compared to our baseline on the Apolloscape dataset.