Abstract
Predicting the multimodal future motions of neighboring agents is essential for an autonomous vehicle to navigate complex scenarios. It is challenging as the motion of an agent is affected by the complex interaction among itself, other agents, and the local roads. Unlike most existing works, which predict a fixed number of possible future motions of an agent, we propose a map-adaptive predictor that can predict a variable number of future trajectories of an agent according to the number of lanes with candidate centerlines (CCLs). The predictor predicts not only future motions guided by single CCLs but also a scene-reasoning prediction and a motion-maintaining prediction. These three kinds of predictions are produced integrally via a single graph operation. We represent the driving scene with a heterogeneous hierarchical graph containing nodes of two types. An agent node contains its dynamics feature encoded from its historical states, and a CCL node contains the CCL's sequential feature. We propose a hierarchical graph operator (HGO) with an edge-masking technology to regulate the information flow in graph operations and obtain the encoded scene feature for the trajectory decoder. Experiments on two large-scale real-world driving datasets show that our method realizes map-adaptive prediction and outperforms strong baselines.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have