Strip and Spatial Social Pooling for Trajectory Prediction

Abstract

AbstractInterdependencies among all moving vehicles in an autonomous driving scenario are critical to precise trajectory prediction. However, to effectively extract the interdependencies remain to be a challenge. In this paper, we propose a new method, which uses an LSTM encoder–decoder and a strip and spatial social pooling module, to cope with this issue. Our social pooling module extracts the interdependencies from two aspects, namely the complexity of the driving scenario and the degree of importance of the surrounding vehicles. Specifically, the proposed social pooling module consists of two sub-modules: a strip pooling sub-module (SPM) and a spatial attention sub-module (SAM). The strip pooling sub-module captures the interaction information among a target vehicle and its horizontal surrounding vehicles and vertical surrounding vehicles. And the spatial attention sub-module emphasizes the degree of importance of the surrounding vehicles. Extensive evaluations on a naturalistic driving dataset: next-generation simulation demonstrates that our method has competitive performance compared with several state-of-the-art methods.KeywordsTrajectory predictionVehicle interdependenciesStrip poolingSpatial attention