PPLIO: Plane-to-Plane LiDAR-Inertial Odometry With Multi-View Constraint in Real-Time

Recent feature-based LiDAR-inertial odometry (LIO) algorithms mainly adopt a point-to-plane model, which requires complex association between numerous feature points and planes. Since these methods only use distance as a parameter to associate feature points, they are prone to a mismatch of features in the presence of initial errors. In this study, we propose a LIO system based on a plane-to-plane model with a sliding window, named PPLIO. To extract high-quality plane features, the geometric and statistical properties are utilized and robust data association is performed with two parameters from the plane basis: distance and normal vector. In the state estimator part, we design two plane-to-plane measurement models using the plane features including the proposed plane RANSAC for outlier rejection. Each measurement model is formulated in the multi-state constraint Kalman filter framework for accurate estimation. For detailed verification, we conduct the Monte-Carlo simulation to evaluate the feature accuracy and the data association. In addition, we perform comparative evaluation against state-of-the-art LIO/LIO-mapping algorithms on both indoor and outdoor experiments, and present results that PPLIO outperforms other real-time algorithms in localization performance with detailed analysis.