Method and computing system for vehicle motion forecasting

Abstract

A method for vehicle motion forecasting includes the following steps. A lane graph structure is generated according to a raw map data. Multiple occupancy flow graphs which are homogeneous to data format of the lane graph structure are established according to trajectory data of a plurality of vehicles in multiple consecutive frames and the lane graph structure. Multiple temporal edges between the occupancy flow graphs are established according to the trajectory data of the vehicles in the consecutive frames to construct a temporal occupancy flow graph. Feature aggregation is performed on the temporal occupancy flow graph to generate multiple updated node features, and a motion forecasting of an ego-vehicle is generated according to the updated node features.