Method, device and storage media for multi-agent motion prediction

Abstract

A multi-agent motion prediction method is performed by a system. The system may take each of the agents in a traffic scenario as a central agent respectively, and divide the traffic scenario into different areas according to the central agent. After that, a local eigenvector is obtained for each of the central agents in the area, and the coordinate system of local eigenvectors between all of the central agents is corrected. As a result, the motion for each of the central agents in accordance with local eigenvectors of each of the central agents and long-range dependencies is predicted by obtaining long-range dependencies between each of the central agents.

Dependencies are inferred from the claim text.