Pizza-LIO: Intensity-Enhanced LiDAR-Inertial Odometry With Adaptive Scan Slicing

LiDAR-inertial odometry (LIO) is a fundamental technique for robotics and autonomous systems. However, most existing LIO methods directly couple low-frequency LiDAR frames with high-frequency IMU measurements, and the instability of LiDAR intensity values further leads to unreliable feature extraction, both of which limit the system from achieving higher accuracy. This paper presents Pizza-LIO, an intensity-enhanced LIO framework conceptually inspired by the flexible slicing process of a pizza, which employs adaptive scan slicing to achieve a pseudo continuous-time formulation. The proposed method can adaptively determine the scan slicing size by considering the relative sampling rates of LiDAR and IMU sensors under different sensor configurations. To maintain stable pose estimation, sparse slices with insufficient features are adaptively merged. Moreover, since raw intensity is easily affected by environments, a local Gaussian-weighting and dual-mode switching strategy is proposed for robust intensity correction. During slice-to-map registration, only planar points with reliable intensity values are jointly constrained by geometric and intensity information, while non-planar points are optimized solely using geometric constraints. Extensive experiments on a private dataset and two public benchmarks (NTU and NCLT) demonstrate that Pizza-LIO achieves lower APE RMSE than state-of-the-art methods (1.317, 0.175, and 1.011 m, respectively), corresponding to improvements of 8.6%, 6.1%, and 14.5%, indicating superior accuracy and robustness across diverse scenarios.