An Improved Discriminator for GAN-Based Trajectory Prediction Models

Abstract

Pedestrian trajectory prediction is an important component in autonomous systems, such as self-driving cars and social robots. It aims to accurately predict or plan future paths for pedestrians according to their movement histories. Recent studies have shown promising progress and most of them use some advanced encoder-decoder structures with Generative Adversarial Networks (GANs) to generate a distribution of multiple plausible paths of an agent. However, GAN-based models suffer from hard-training problems and training Recurrent Neural Networks (RNNs) is especially difficult. In this paper, we propose a discriminator that shares its encoder with the generator to reduce the training difficulty. We incorporate this discriminator into two successful stochastic models designed for pedestrian trajectory prediction. Our experimental results demonstrate that the new discriminator outperforms the baseline structures in general on multiple datasets.