A Novel Trajectory Generator Based on a Constrained GAN and a Latent Variables Predictor

Abstract

Forecasting pedestrian trajectory is critical for versatile applications, such as autonomous driving and social robot, when they work in human-centric environments. However, it is challenging to predict pedestrians' future trajectories due to the inherent human properties and pedestrians' social interactions. Recent works predict future trajectories by using a generative model, which captures social interactions with pooling- or graph-based strategies and generates multi-modal outputs with latent variables sampled from random Gaussian noise. Nevertheless, they introduce little human knowledge, which is beneficial for improved prediction performance. In this work, we propose to learn informative latent variables from pedestrians' future trajectories. Moreover, we present a distance-direction pooling module, which captures social interactions in a more intuitive manner. Besides, we introduce an additional constraint on generative adversarial network optimization to generate more realistic results. Two benchmarking datasets, ETH (Pellegrini et al., 2010) and UCY (Leal-Taixé et al., 2014), are used to evaluate the proposed method. Comparisons between our method and several state-of-the-art methods demonstrate the superiority of the proposed method in generating more accurate future trajectories.

Keywords

• Trajectory forecasting
• generative adversarial network
• latent variable predictor
• future uncertainty