Asynchronous Visual-Inertial Motion Parsing via Sparse Differential Dynamics for Agile Mobile Platform
Accurate motion perception is crucial for agile mobile platforms navigating in dynamic environments. However, visual-inertial systems frequently suffer from dominant ego-motion interference and the inherent asynchronous sampling rates between visual sensors and inertial measurement units (IMUs). In this paper, we propose a novel framework for asynchronous visual-inertial motion parsing based on sparse differential dynamics. To alleviate computational bottlenecks, we first employ a sparse visual activation strategy that efficiently extracts high-value features and suppresses redundant background information. To address the irregular and asynchronous sensor sampling, we formulate the evolution of the latent motion field as a continuous-time Neural Controlled Differential Equation (NCDE). This continuous representation acts as a robust temporal prior, decoupling global ego-motion from independent object dynamics. Real-world unmanned aerial vehicle (UAV) flight experiments demonstrate that the proposed framework effectively handles frame loss and aggressive maneuvers. By strictly enforcing physical consistency, our method significantly reduces trajectory drift, offering a lightweight and robust solution for agile mobile platform.