Stochastic Robust Trajectory Prediction

Personnel: Chen Zhou, Ghassan Alregib, Armin Parchami, Kunjan Singh

Goal: To model two major stimuli: individual stochastic goals, and social compliance in human trajectory prediction.

Challenges:  Humans have high-level internal plans to achieve their desired destinations. Internal plans vary between individuals thereby goals can be under-deterministic, especially with adversarial intention. Hence, stochasticity exists in human behaviors. In addition to under-deterministic individual goals, huamns comply with social conventions by interacting with others. Social interactions can also influence the decision-making. Therefore, developing a framework that learns stochastic individual goals and social interactions to forecast human trajectories is challenging and desirable.

Our Work: In this work, we introduce a trajectory prediction framework that encompasses the modeling of two essential stimuli of human behavior, external social interactions, and individual goals. We propose a robust relation learning paradigm via region-wise temporal dynamics to model the social compliance between agents. The robustness experiments indicate that the region-based relation representations are less vulnerable to spatial noise perturbation compared to approaches with edge- based relation learning. In addition, we present an empirical study of goal estimation to uncover the discrepancy between training and test behavior data, motivating the non- deterministic goal estimation to account for human stochasticity. Specifically, a CVAE is adopted to estimate multiple plausible goals. Finally, we integrate the proposed region- based relation learning module and the multi-goal estimation into our full framework to model external social compliance and individual goals. The comparative experiments demonstrate that our framework achieves superior or comparative performance against state-of-the-art methods on the ETH-UCY human trajectory dataset.

References: 

  1. J. Lee, M. Prabhushankar, and G. AlRegib, "Gradient-Based Adversarial and Out-of-Distribution Detection," in International Conference on Machine Learning (ICML) Workshop on New Frontiers in Adversarial Machine Learning, Baltimore, MD, Jul. 2022. [PDF]