Multi-Granular Transformer for Motion Prediction with LiDAR

Motion prediction has been an essential compo- nent of autonomous driving systems since it handles highly uncertain and complex scenarios involving moving agents of different types. In this paper, we propose a Multi-Granular TRansformer (MGTR) framework, an encoder-decoder net- work that exploits context features in different granularities for different kinds of traffic agents. To further enhance MGTR’s capabilities, we leverage LiDAR point cloud data by incorpo- rating LiDAR semantic features from an off-the-shelf LiDAR feature extractor. We evaluate MGTR on Waymo Open Dataset motion prediction benchmark and show that the proposed method achieved state-of-the-art performance, ranking 1st on its leaderboard 1.