Continuous-Time Event-Driven Temporal–Causal Attention for Intention-Driven Anticipation in Badminton

High-speed badminton requires defenders to anticipate opponents’ intentions from early kinematic cues under severe temporal and sensory constraints. Existing models rely on discrete-time assumptions and statistical correlations, failing under irregular sampling or missing frames. We propose a Continuous-Time Event-Driven Temporal–Causal Attention (CT-ACI) framework that models the intention state as a neural ordinary differential equation (Neural ODE), embeds a physical causal mask to prevent backward leakage, and introduces an adaptive integral triggering mechanism for event-driven predictions. This framework is designed as a robust anticipation algorithm for irregularly sampled sports visual data, providing early candidate signals for downstream decision support. A curriculum training strategy stabilizes optimization. When evaluated on the VideoBadminton dataset, the framework maintains stable prediction under severe missing frame conditions where discrete baselines degrade substantially. In addition, it reduces the long-term anticipation error while providing a binary intention signal under strict temporal–causal constraints. The trigger time T∗ is consistently treated as an operational proxy derived from observable defender motion, not as a validated cognitive decision boundary. The dataset contains only six singles matches and no player identifiers, while two matches are shared between training and test; thus, the reported results characterize generalization to unseen rallies and clips rather than to unseen matches or players.