This paper proposes a multi-scale and center-aligned cylindrical voxel feature construction method that not only preserves the geometric structure of point clouds, but also adapts to the sparsity, disorganization, and uneven density of large-scale point clouds. Firstly, the center-aligned cylindrical voxelization, compared to traditional voxelization methods such as regular voxelization and cylindrical voxelization, focuses on preserving geometric constraints, offering a more accurate spatial representation for neural networks. Secondly, voxel features are constructed at multiple resolution to capture both local details and global structure. This enhances the model’s adaptability to objects of different sizes and complex scenes. Our model is validated on the large-scale outdoor SemanticKITTI dataset and the indoor S3DIS dataset. Experimental results demonstrate that the proposed model is sufficiently lightweight, with only 2 M parameters, and achieves a mean intersection over union of 70.2% on the SemanticKITTI dataset, establishing a leading benchmark at this level of parameter size.