KnowMTP: A Knowledge-Guided Framework for Multi-Agent Trajectory Prediction in Autonomous Driving

Rufan Bai, Tianyi Xue, Tiantian Zhou, Weiwei Wu, Changle Li, Yuhuan Lu

Abstract:

Accurate multi-agent trajectory prediction (MTP) is essential for autonomous driving, enabling vehicles to anticipate the future behaviors of surrounding agents and plan safe, efficient maneuvers.
Although recent data-driven approaches based on the encoder-decoder architecture have achieved notable success, they often rely solely on attention mechanisms to process uniformly structured inputs, overlooking valuable domain knowledge and introducing unnecessary complexity.
This paper presents KnowMTP, a knowledge-guided framework for MTP that integrates driving safety and behavioral knowledge to intelligently reorganize input data.
KnowMTP achieves this through two complementary filtering mechanisms that combine a Motion-Similar Agent Selection filter identifying relevant neighbors based on motion similarity with a Safety-Critical Agent Selection filter prioritizing agents posing potential collision risks, thereby embedding cognitive awareness of safety constraints.
The framework is model-agnostic and integrates seamlessly into existing encoder-decoder architectures.
Experiments on benchmark datasets show that KnowMTP consistently improves the performance of state-of-the-art MTP baselines, achieving 18.8\% gains on WOMD with 22.1\% lower computational cost while enhancing the plausibility and safety awareness of predicted trajectories.