A Robust Real-Time LiDAR Odometry with Adaptive Layered Voxel Map

This paper presents a Light Detection and Ranging (LiDAR) odometry system comprising three key components: ground segmentation, adaptive layered voxel map storage, and residual iterative optimization. To overcome the limitations of existing LiDAR odometry systems, which lack robustness and fail to dynamically adjust the number of local map points based on environmental features, we design an adaptive downsampling mechanism that adjusts voxel sizes in real time to optimize storage and improve robustness. To address the challenge of excessive map storage, we propose a ground segmentation algorithm that employs distinct storage strategies for ground and non-ground points, utilizing a plane-based voxel model for ground points and a point-based voxel model for non-ground points. This approach not only reduces map size but also accelerates the iterative optimization process while maintaining high accuracy through point-to-plane and point-to-point residuals. The proposed system is evaluated on the KITTI and R3LIVE++ datasets, with experimental results demonstrating its support for both solid-state and mechanical LiDARs, trajectory accuracy comparable to state-of-the-art methods, and efficient real-time performance.