Abstract:
We present a thorough evaluation and analysis of state-of-the-art (SOTA) human trajectory forecasting methods with respect to metrics for safe and socially-aware prediction, e.g., collision rate, in addition to traditional displacement metrics, e.g., average displacement error. First, we introduce a system for trajectory classification which is used to evaluate the strengths and weaknesses of various methods. We design a trajectory simulator that generates trajectories from these different categories to fully characterize performance. Second, we demonstrate that SOTA deep imitation learning (IL)-based trajectory prediction algorithms perform significantly poorer in certain trajectory categories. Surprisingly, we show that SOTA methods fail to demonstrate basic collision avoidance behavior when compared to traditional trajectory forecasting methods (e.g., social force models). We provide evidence that this failure may result from (1) a lack of diverse training data in real datasets, particularly a lack of trajectories involved in collision avoidance, as well as (2), a lack of explicit mechanisms for modeling collision-avoidance in deep SOTA architectures. Finally, we propose a trajectory forecasting solution that unifies techniques from the traditional energy-based Social Force model and the current SOTA deep IL method (Agentformer), which outperforms pure deep IL methods on collision avoidance while maintaining SOTA-level displacement error.

Committee:
Kris Kitani
Deva Ramanan
Michael Erdmann
Ye Yuan