The core message of this work is to propose a robust trajectory prediction framework that models two major stimuli of human behavior: external social interactions and individual stochastic goals. The framework learns region-based social relations to capture the dynamics of crowd density changes, which is more robust to spatial noise perturbations compared to edge-based relation learning approaches. It also estimates multiple plausible goals to account for the stochasticity in human behavior.
This paper proposes a novel approach to synthetic dataset generation based on composite probabilistic Bézier curves, which is capable of generating ground truth data in terms of probability distributions over full trajectories, enabling the use of more expressive error metrics like the Wasserstein distance for model evaluation.
The FlexiLength Network (FLN) framework effectively addresses the Observation Length Shift issue in trajectory prediction, enabling robust performance across a range of observation lengths without substantial modifications to existing models.
Enhancing trajectory prediction through test-time training with a masked autoencoder and actor-specific token memory.