Learning Interactive Knowledge Graph for Trajectory Prediction

In the process of pedestrian movement, the trajectory is not only related to their subjective intention, but also affected by the surrounding agents and the environment. How to more effectively model the interaction between agents plays a very significant role in trajectory prediction task, which is also the focus of researchers’ work. To solve this problem, this paper proposes a knowledge graph construction method based on trajectory clustering to extract the interactive features between adjacent pedestrians. Based on the fact that the behavior of pedestrians has the property of group psychology, we first do spectral clustering on the trajectory of pedestrians to find their inter class information. Then, through the analysis of the clustering results and the movement angle of pedestrians, the interactive knowledge graph structure of each frame is constructed. Finally, we fuse it with the relative distance graph of pedestrians to encode the interactively social relation in trajectory prediction. Through the evaluation on the public ETH and UCY datasets, the superiority of our method is demonstrated by exhaustive experiments.