PSS-LIO A Computationally Efficient LiDAR-Inertial Odometry Framework for UAV Building Inspection via Geometric Predictability and Structure Prior Learning

Unmanned aerial vehicle (UAV)-based building inspection demands accurate and computationally efficient state estimation in GNSS-denied environments. Existing LiDAR-inertial odometry (LIO) systems suffer from suboptimal observation utilization: frame-based and point-by-point methods process observations without exploiting geometric predictability, scalar information metrics risk observability degradation by conflating directional constraints, and recent plane merging methods improve map quality but do not feed learned structure back into state estimation. This paper presents PSS-LIO, a building-inspection-oriented LIO framework integrating three complementary mechanisms. First, a prediction-based coarse filtering mechanism bypasses observations conforming to range predictions from mapped surfaces. Second, an information-driven fine selection strategy decomposes Fisher Information along state subspaces, enabling multi-resolution updates with directional coverage guarantees. Third, a structure-prior guided update mechanism learns dominant surface orientations online as adaptive soft constraints, directly enhancing state estimation rather than merely improving map representation. Experiments on public datasets and real-world UAV deployments demonstrate that PSS-LIO achieves consistent accuracy improvements while substantially reducing computational requirements.