DAG-Net: Double Attentive Graph Neural Network for Trajectory\n Forecasting

Abstract

Understanding human motion behaviour is a critical task for several possible\napplications like self-driving cars or social robots, and in general for all\nthose settings where an autonomous agent has to navigate inside a human-centric\nenvironment. This is non-trivial because human motion is inherently\nmulti-modal: given a history of human motion paths, there are many plausible\nways by which people could move in the future. Additionally, people activities\nare often driven by goals, e.g. reaching particular locations or interacting\nwith the environment. We address the aforementioned aspects by proposing a new\nrecurrent generative model that considers both single agents' future goals and\ninteractions between different agents. The model exploits a double\nattention-based graph neural network to collect information about the mutual\ninfluences among different agents and to integrate it with data about agents'\npossible future objectives. Our proposal is general enough to be applied to\ndifferent scenarios: the model achieves state-of-the-art results in both urban\nenvironments and also in sports applications.\n