Modeling Multimodal Dynamic Spatiotemporal Graphs
share
Spatiotemporal graphs (STGs) are a powerful tool for modeling multi-agent interaction scenarios, used commonly in human trajectory prediction and proactive planning and decision making for safe human-robot interaction. However, many current STG-backed methods rely on a static graph assumption, i.e. that the underlying graphical structure maintains the same nodes and edges throughout the scenario. This assumption is frequently broken in real-world applications, especially in highly-dynamic problems such as human trajectory prediction in crowds. To remove the reliance on this assumption, we present a methodology for modeling and predicting agent behavior in both highly dynamic and multimodal scenarios (i.e. where the scene's graphical structure is time-varying and there are many possible highly-distinct futures for each agent). Our approach to model dynamic STGs augments prior multimodal, multi-agent modeling methods with a gating function on edge models that smoothly adds and removes edge influence from a node. We demonstrate the performance of our approach on the ETH multi-human trajectory dataset and on NBA basketball player trajectories. Both are highly-dynamic, multimodal, and multi-agent interaction scenarios which serve as proxies for many robotic applications.
READ FULL TEXT