BiTraP: Bi-directional Pedestrian Trajectory Prediction with Multi-modal Goal Estimation

• DOI: 10.48550/arxiv.2007.14558
• Publication Date: 2020-01-01
• Authors: Yu Yao; Ella Atkins; Matthew Johnson‐R...; Ram Vasudevan; Xiaoxiao Du
• Keywords: Autoencoder

Pedestrian trajectory prediction is an essential task in robotic applications such as autonomous driving and robot navigation. State-of-the-art predictors use a conditional variational autoencoder (CVAE) with recurrent neural networks (RNNs) to encode observed trajectories decode multi-modal future trajectories. This process can suffer from accumulated errors over long horizons (>=2 seconds). paper presents BiTraP, goal-conditioned bi-directional method based on the CVAE. BiTraP estimates goal (end-point) of introduces novel decoder improve longer-term accuracy. Extensive experiments show that generalizes both first-person view (FPV) bird's-eye (BEV) scenarios outperforms state-of-the-art results by ~10-50%. We also different choices non-parametric versus parametric target models CVAE directly influence predicted distributions. These provide guidance predictor design for collision avoidance navigation systems.