Embodied intelligence integrated with action semantic graph for dynamic capture modeling of dance
Abstract
To address the issues of insufficient local semantic expression, long-term temporal dependency attenuation, insufficient utilization of joint topological association, and weak stability of pose reconstruction in the dynamic capture modeling of high-degree-of-freedom human motion sequences, a dynamic capture modeling method integrating embodied intelligence and action semantic graph is proposed. The model takes the multi-modal motion states such as skeleton position, joint angle, and velocity increment as inputs, and constructs a semantic graph that couples joint nodes, action semantic nodes, and temporal relationship edges. In the encoding stage, constraints such as bone length consistency, joint activity boundary constraints, and temporal smoothness constraints are introduced to form a embodied prior-driven graph-spatio-temporal joint representation mechanism. Based on semantic-guided feature aggregation, temporal dependency modeling, and pose decoding reconstruction, high-precision dynamic capture modeling of continuous motion states is accomplished. The entire paper focuses on action state encoding, semantic graph construction, embodied constraint fusion, capture decoding optimization, and experimental verification.