UniPredictor: Unified Network for Vehicle Intention and Trajectory Joint Prediction

For autonomous vehicles, it is crucial to comprehend the interaction among agents in the scene, and accurately predict their intention and trajectory. In this paper, we present UniPredictor, a novel and unified network designed for both intention and trajectory prediction tasks. To effectively capture the complex and diverse information within a scenario, the Spatial-Temporal Anchor-Attention Encoder processes agents’ historical states and map features across spatial and temporal dimensions, respectively. Subsequently, the Proposal Fusion Intention Decoder combines scene context encoded by the encoder with proposal features to infer intentions. Eventually, the Autoregressive Trajectory Decoder leverages query to extract future motion features from scene context and predicts trajectories using an autoregressive approach. Experiments on the dataset demonstrate that UniPredictor provides precise predictions, outperforming previous methods.