Abstract: To enhance the precise navigation capability of robot in complex environments with uneven terrain such as stairs and hills, and overcome issues in existing LiDAR-inertial odometry systems including significant cumulative error drift along Z-axis and low-quality map, a tightly coupled LiDAR-inertial odometry based on an adaptive voxel feature map is proposed. Firstly, the environment is adaptively partitioned into grids using an octree and Hash indexing, enabling efficient management and indexing of point cloud feature data through the construction of a voxel feature map. Then, a direct voxel-based feature indexing method is employed for feature extraction from the current frame based on the voxel feature map, avoiding redundant feature fitting, which improves the robustness and efficiency of feature extraction and fitting in the pose estimation module. Finally, the concept of co-visible voxel grids is introduced. When the same voxel grid is observed jointly by multiple frames in a sliding window, a sliding window-based bundle adjustment (BA) optimization problem is formulated, to adjust all poses within the window and update the states maintained by the extended Kalman filter based state estimation. Experiments on public datasets and in real-world scenarios demonstrate that the proposed method effectively mitigates cumulative errors along the Z-axis, achieving better global consistency for both trajectories and maps compared to traditional algorithms.