ÄûÃʵ¼º½

In this paper, we introduce an event-triggered Maps of Dynamics (ETMoD) framework for modeling spatial motion patterns in non-stationary environments. Traditional approaches often rely on fixed grid resolutions and assume gradual temporal changes, limiting their effectiveness in real-world scenarios where motion patterns exhibit abrupt variations. To address these limitations, we propose a novel framework that employs a grid-shifting mechanism to generate context-aware cells based on historical observations. Temporal patterns are modeled using Neural Stochastic Differential Equations, while a diffusion model is integrated to handle abrupt changes in motion patterns through an event-triggered mechanism. Experimental results demonstrate that our framework outperforms state-of-the-art methods, particularly in capturing abrupt changes during peak activity periods, while significantly reducing training time.

Author: Junyi Shi, Qingyun Guo, Tomasz Piotr Kucner