HD-Net: Cognitive-Theory-Powered Trajectory Prediction for Autonomous Driving

Existing trajectory prediction models often rely on attention mechanisms to aggregate interaction features among scene elements, while simply stacking layers tends to compromise their feasibility. Recent researches have attempted to boost accuracy by introducing driving emotion or style into consideration. Even so, these approaches usually depend on data augmentation and rarely leverage cognitive theories for modeling. To advance research on cognitive-based methods, a heuristic-driven method called HD-Net is proposed in this work, which leverages cognitive theory without any expanded data. The core insight of the method lies in first defining an implicit formula for the emulation of information retrieval process, which is then used to efficiently model a graph attention framework for continuous predictions. Extensive experiments are performed on the ETH/UCY and Argoverse datasets, which demonstrate that HDNet outperforms baselines for both pedestrian and vehicle prediction tasks. Moreover, our work show that the continuous adjustment process of information can significantly enhances contextual feature and improves the prediction accuracy.