Understanding crowd behaviors is crucial in many vital areas e.g. public safety, urban planning, autonomous vehicles, etc. Although numerous excellent models have been proposed for the study of crowd behaviors, challenges persist due to the complexity of human behaviors. This thesis focuses on two primary challenges: crowd dynamics modeling in low-density crowds and physical interaction modeling in high-density crowds. To this end, we propose novel methods for prediction and uncertainty analysis of crowd dynamics in low-density scenarios while we introduce and solve a new research question about full-body motion to model physical interactions.
This thesis first models crowd dynamics to predict crowd movements in low-density crowds, which is also known as human trajectory prediction. Existing methods are typically divided into model-based and model-free methods. We design a new framework Neural Social Physics incorporating the advantages of both methodologies based on neural differential equation models. Then we propose a novel method under the framework by combining the social force model with neural networks. Our method trains neural networks to estimate the parameters of the social force model instead of hand-picking or fixing them. A deep generative model is employed to capture the stochasticity of human trajectories. Through exhaustive evaluation, our method outperforms existing methods in prediction accuracy by up to 70%. In addition, our method provides plausible explanations for pedestrian behaviors and shows strong generalizability.
This thesis further explores uncertainty modeling in human trajectory prediction to capture stochastic future trajectories. The uncertainty of human trajectories consists of the data and model uncertainty. However, existing methods either only consider the data uncertainty or mix the two uncertainties, which is coarse-grained. To overcome the challenge, we propose a new Bayesian stochastic social force model, which captures fine-grained uncertainty through a decoupling strategy. Specifically, our method captures the data and model uncertainty by using a new Bayesian neural stochastic differential equation model and a deep generative model, respectively. We designed the uncertainty-aware training scheme allowing the two models to catch corresponding uncertainties well. Extensive experiments demonstrate that our method has strong explainability and improves the state-of-the-art prediction accuracy by as much as 60.17%.
Lastly, this thesis proposes a new research question that predicts 3D full-body motions under unexpected physical perturbation on both individual and group levels to study physical interactions between individuals. However, incorporating physical interactions in motion prediction brings new challenges e.g. complex interactions and data scarcity. To this end, we propose a latent differentiable physics model based on differentiable physics and neural networks. Our model introduces a latent physics space to learn body physics. Motions in the latent physics space are estimated first and converted then back into 3D full-body motions through a deep generative model. Considering that there is no similar research, we carefully choose 11 baselines from relevant domains and adapt them to the new task. Extensive evaluation and comparison demonstrate that our method outperforms other baselines in prediction accuracy by up to 70% and has outstanding data efficiency, strong generalizability, and good explainability.
