Self-Distillation Attention for Efficient and Accurate Motion Prediction in Autonomous Driving

The accuracy and stability of motion prediction are crucial for the safe planning of autonomous driving systems. The widely used attention mechanisms effectively improve prediction accuracy. However, their computational cost grows quadratically with sequence length, presenting challenges for handling complex, large-scale scenarios. The attention patterns of motion prediction tasks exhibit significant data-related sparsity, indicating that not all scene elements are worth interaction. Therefore, this paper proposes a novel motion prediction framework that enhances the efficiency and stability of interaction fusion while achieving promising prediction accuracy. Firstly, an adaptive self-distillation attention module based on sparse interaction graph is designed. This module adaptively filters high-value sequences for each target to achieve sparse attention calculation and balance scene scale. Due to the lack of interaction labels in the original dataset, a self-distillation strategy is employed for model training. Secondly, a multi-stage dynamic anchor decoder is introduced that leverages the information filtering and aggregation capabilities of the sparse attention mechanism to improve prediction accuracy. The decoder dynamically updates the target anchor representations as the prediction progresses, ensuring consistency between interaction states and the prediction process in long-term forecasting. This approach effectively focuses attention calculation on the most relevant scene context fusion and trajectory decoding at each prediction stage. Validation results on Argoverse 1 and Argoverse 2 demonstrate that the proposed method achieves competitive accuracy while effectively reducing computational resource consumption. The proposed method can also serve as a plug-in that can be seamlessly incorporated into standard motion prediction pipelines to optimize scene interaction, making it more friendly to low-cost devices.