Multimodal Motion Prediction Incorporating the Constraints of Basic Driving Behaviors

Due to the inherent uncertainty within agents, multimodal motion prediction is required in real-world scenarios. Mainstream algorithms employ anchors or prior candidates to achieve multimodal predictions, and the primary challenge is to enhance the rationality and diversity of the modality distribution while ensuring prediction accuracy. This article presents a progressive modality construction and refinement strategy incorporating the constraints of basic behaviors for multimodal motion prediction. First, different generalized modality regions are partitioned based on high-level control commands, incorporating the constraints from basic driving behaviors. It can effectively aggregate training samples with similar driving behaviors, mitigating modal collapse and enhancing data utilization. Second, a hybrid-proposal refinement decoder is designed to integrate adaptive anchor-free queries with region-anchor corresponding to generalized modality region, further refining the diversity of agent behaviors. To fully leverage the advantages of the modality partition, region-aware attention and lane filtering strategies are proposed to balance prediction accuracy and efficiency effectively. Finally, validation on the Argoverse 1 and Argoverse 2 motion forecasting datasets demonstrates that the proposed model can achieve high prediction accuracy while efficiently utilizing training data to cover multimodal motion behaviors reasonably.