Pedestrian Trajectory Prediction Driven by Bidirectional Intention-Interaction

In complex environments, humans estimate the surrounding interactions to plan efficient trajectories. Accurately assessing the relationships and intensity of these interactions is crucial to optimize trajectory prediction. Existing methods implicitly integrate interaction features into trajectory representations, making it difficult to disentangle their influence and leading to sensitivity to indirect factors such as distance. We note that interactions and intentions are strongly correlated, as dynamic interactions among pedestrians are guided by their respective intentions and, in turn, influence these intentions. Therefore, we propose a novel intention-interaction bidirectionally driven pedestrian trajectory prediction network (I2I-Net). The network decomposes future trajectories into short-term intention-guided segments and iteratively refines intentions based on interaction estimation at each time step. Our proposed method achieves state-of-the-art performance in experiments on the ETH & UCY and SDD datasets, validating the method’s ability to precisely model the influence of human interactions on pedestrian trajectories.