The network processes diverse input modalities, including bounding boxes, pedestrian pose, and ego-vehicle motion information. We enhance prediction performance by employing specialized encoders to ...