.png){kind=icon}
Return
MPDiff: Multiscale Physics-Informed Diffusion Model for Crowd Simulation
.png){kind=icon}
PRE
AI
DOI：10.1109/tcsvt.2026.3749134.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Crowd dynamics involve complex microscopic interactions and macroscopic flow patterns. Accurately capturing both levels is crucial for realistic crowd simulations. Integrating multi-scale physical priors within a unified framework while enhancing the realism and physical consistency of simulations remains a key challenge in crowd simulation. We propose MPDiff, a Multiscale Physics-Informed Diffusion Model for crowd simulation. By incorporating macro-scale flow priors from fluid dynamics and micro-scale behavioral priors from the Social Force Model (SFM) into a unified diffusion framework, MPDiff generates accurate and physically consistent crowd trajectories. The model comprises two core modules. The first is the Physics-Aware Crowd Representation Model (PA-CRM), which constructs three macroscopic flow field features, namely crowd inertia, crowd consistency, and escape tendency, from the Navier-Stokes (NS) equations, extracts three microscopic behavioral features, namely goal orientation, individual interaction, and motion continuity, from the SFM, and achieves adaptive cross-structure multiscale feature fusion via cross-attention. The second is the Multiscale Physical Feature-driven Generative Module (MPF-GM). As a conditional diffusion model guided by fused multiscale physical features, it incorporates physics-aware conditional priors into the denoising process via adaptive layer modulation and destination guidance, thereby enhancing the physical consistency and behavioral realism of trajectories. Experimental results on four real-world datasets (UCY, GC, Lane, and DCFD) demonstrate that MPDiff not only outperforms state-of-the-art approaches on key metrics such as displacement error, distribution alignment, and collision count in most scenarios but also effectively mitigates error accumulation in long-term rollout simulations, exhibiting superior stability and improved physical consistency.
Keywords:
Crowd simulation
physics-informed deep learning
multiscale feature fusion
conditional diffusion model
Journal
IF:
11.1
Papers:
1.8K
Citations:
3.1W
Organization
View more organizations
Cited Papers
No cited papers available