Researchers have developed a tightly-coupled LiDAR-inertial odometry algorithm that uses an adaptive voxel feature map to reduce cumulative Z-axis drift and improve map quality in complex environments such as stairs and hills. The method addresses limitations of existing systems that suffer from sparse vertical point cloud density and gravity-related error accumulation, which often lead to significant vertical drift and degraded mapping.
Key takeaways
- Adaptive voxel grid partitioning via octree and hash indexing enables efficient point cloud feature management and indexing.
- A two-stage feature residual computation method avoids redundant feature fitting, improving robustness and efficiency of pose estimation.
- Introduces co-visible voxel grids and sliding-window bundle adjustment (BA) to jointly optimize poses and update extended Kalman filter states, reducing cumulative Z-axis error.
- Experiments on public datasets and real-world scenarios show improved global consistency of trajectories and maps compared to traditional algorithms.
The algorithm integrates IMU forward/backward propagation for state estimation, constructs a voxel feature map with incremental KD-tree management, and uses a dynamic sliding window for multi-frame optimization. This approach does not rely on loop closure, making it suitable for long-duration navigation in satellite-denied environments. The work provides a practical solution for mobile robots operating on uneven terrain where accurate vertical positioning is critical.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-05-12 · “基于自适应体素地图的紧耦合激光惯性里程计”