DGSG: A Efficient Goal Directed Sequence Generator for Pedestrian Trajectory Prediction

Abstract

Trajectory prediction is a crucial task in modern era as it can help ego robots work safely in crowded environments. It’s yet challenging for the stochasticity of human motion and the restriction of platform. Previous method ignore the problem of time and space complexity. Based on recently developed Variational Auto Encoder(VAE), we proposed a trajectory prediction model named goal directed sequence generator(DGSG). In this model, the prediction task is divided into two modules achieved by light neural network respectively. The goal estimation module is supported by a VAE based network with a reformed loss function to modify the relationship between destinations and observed trajectories. And the sequence generation module prediction future trajectory based on the destination. Our experiments have shown that our method has achieve a state-of-art performance in commonly used datasets. Furthermore, experiments prove that our method is easy to deploy for the outstanding time and space complexity.