Robust visual pose measurement and uncertainty suppression for UAVs in dynamic landing environments

Abstract High-precision relative attitude measurement is a critical prerequisite for autonomous recovery of micro unmanned aerial vehicles (UAVs) in GNSS-restricted and dynamic environments. However, motion blur and drastic scale changes often lead to a surge in visual observation noise, causing significant drift in state estimation. To address this, this paper proposes a holistic visual metrology system that deeply fuses enhanced perception with uncertainty suppression. First, at the perception layer, an improved YOLO11n-Landing network serves as a high-fidelity instrument, utilizing multi-scale feature enhancement and anti-blur attention mechanisms to boost detection confidence in dynamic scenes. Second, a height-constrained Kalman filter was constructed at the fusion layer, explicitly managing uncertainty propagation through a physical covariance model. The validity of the optimal fusion interval was confirmed based on parameter sensitivity analysis. Crucially, at the system level, a perception-control coupled isolation mechanism has been introduced. Unlike traditional loosely coupled architectures, this mechanism uses the visual confidence factor as an information hub, synchronously mapping it to the filter covariance, switching gain K , and boundary layer thickness. This dynamically reconstructs the system’s response bandwidth, actively isolating invalid observations, and physically blocking the propagation of measurement outliers into the control loop. Experimental results demonstrate that compared to traditional benchmark methods, this integrated system significantly converges the root mean square error to 0.156 m, reduces the error variance by approximately 74.7%, and achieves a 99.0% landing success rate in Monte Carlo simulations. This work not only provides a highly system-integrity-preserving technical solution for robust autonomous landing of micro UAVs, but its core coupling paradigm also demonstrates broad prospects for generalization to high-risk tasks such as aerial docking and close-range inspection.