Physics–Environment Interaction Network for dense crowd behavior recognition

Jiaqi Yu, Yanshan Zhou, Renjie Pan, Pingrui Lai, Hua Yang

Abstract: Highlights•Based on fluid-crowd similarity, we propose PEIN to capture motion patterns for dense crowd behavior recognition.•By linking N-S equations with crowd social properties, we build a physics-informed stream to model crowd behavior dynamics.•A dual cross-attention mechanism enables two-stream interaction for comprehensive dense crowd behavior understanding.•We introduce MCBD, a dense crowd dataset with structured and unstructured scenes, closer to real-world conditions than DCFD.

External IDs:doi:10.1016/j.patcog.2025.112123