Static-dynamic global graph representation for pedestrian trajectory prediction

Abstract

Effectively understanding social interactions among pedestrians plays a significant role in accurate pedestrian trajectory prediction. Previous works used either distance-based or data-driven methods to model interactions. However, the distance-based method has difficulty modeling complex interactions and ignores interactive pedestrians that are beyond a certain distance. The data-driven method models interactions among all pedestrians in a scene and introduces noninteractive pedestrians into the model due to the lack of proper supervision. To overcome these limitations, we first propose a novel global graph representation, which considers the spatial distance (from near to far) and the motion state (from static to dynamic), to explicitly model the social interactions among pedestrians. The global graph representation consists of two subgraphs: the static and the dynamic graph representations, where the static graph considers only the nearby pedestrians within a certain distance threshold, and the dynamic graph considers the interactive pedestrians that will likely collide soon. The proposed graph representation explicitly models the interaction by incorporating both the static (location) and dynamic states (velocity) in a distance-based manner. Then, based on the global graph representation, a novel data driven graph encoding network is proposed to extract the interaction features. It adopts two independent LSTMs and an attention module to encode the interaction feature from the perspective of the ego-pedestrian. Finally, the proposed prediction method is evaluated on two benchmark pedestrian trajectory prediction datasets, and comparisons are made with the state-of-the-arts. Experimental results demonstrate the effectiveness of the proposed method.