Optimisation of cross-modal behaviour representation for complex interaction scenarios

Complex interactions among agents, objects, and environments require robust behaviour understanding from heterogeneous multimodal cues. However, existing methods struggle with temporal asynchrony, relation-dependent semantic shifts, and modality unreliability. This paper proposes an interaction-aware cross-modal behaviour representation optimisation framework integrating modality-specific feature extraction, temporal semantic alignment, relation-conditioned contextual modelling, and adaptive cross-modal fusion. The framework learns unified, discriminative representations while reducing contextual misinterpretations. Experiments demonstrate superior performance over conventional fusion methods and representative baselines, while ablation studies validate the effectiveness of individual components.