Verfahren und System zur Vorhersage der Flugbahn eines sich in einer Realen Szene Bewegenden Agenten

Details

Veröffentlichungs-Nr.

EP4528590

Aktenzeichen

EP23198287

Anmeldetag

19. September 2023

Veröffentlichung

26. März 2025

Rechtsraum

EP

IPC

G06N3/0455G06N3/0499B60W60/00

Abstract

This method predicts a future trajectory of at least one agent in motion in a real scene comprising static elements and elements with a variable state over time respectively represented as a first class and as a second class of polylines. It comprises:- representing (E20) N_AC past trajectories of agents in motion in said scene as a third class of polylines, polylines i of said classes (MP, TL, AG) being represented by a global pose p_i and a local attribute u_i;- modelling (E30) interactions between said polylines (MP, TL, AG) using an attention based neural network transformer comprising modules configured to update local attributes u_i of the polylines of one source class with attention to polylines of at least one target class, said source class and said at least one target class being one of said first, second or third classes of polylines; and- determining (E40) a future trajectory of at least one said agent from outputs of said attention based neural network transformer.

Anmelder

Firmen

TOYOTA JIDOSHA KABUSHIKI KAISHA
ETH Zurich

Land

🇯🇵 Japan

Fachgebiete