Trajectory prediction method and system based on gated attention and probabilistic multi-modal

Abstract

The application discloses a kind of trajectory prediction method and system based on gated attention and probabilistic multi-modal, it is related to the field of automatic driving, the method includes: obtaining the historical trajectory data of multiple vehicles in target scene and coding;The encoded features are input into the interactive enhancement module, the multi-head self-attention weight is reweighted by introducing the gating calibration factor based on dynamic relationship, to model the interaction between vehicles in detail;Then, the enhanced features are input into the multi-modal prediction head, the prediction head is based on conditional variational autoencoder architecture, not only generates multiple future trajectories, but also outputs an independent explicit probability value for each trajectory through a parallel probability estimation branch.The application solves the problem that the existing technology is not fine enough for interaction modeling and the multi-modal prediction probability is not explicit by using the gating calibration attention mechanism and the multi-modal generation structure that can output explicit probability, significantly improves the accuracy, diversity and explainability of trajectory prediction.

Images