In the field of autonomous driving, profound scene understanding is crucial, and semantic segmentation of LiDAR point clouds plays a key role in this context. A prevalent issue in point cloud datasets is the imbalance in class distribution. To address this, we introduce the InstanceAug data augmentation pipeline, which balances the class distribution by duplicating instances within scenes. This approach significantly enhances the robustness of our model. Deep learning models for point cloud processing often use sparse convolution for efficiency, but this limits feature transmission and the receptive field. Building on the strengthened dataset, we present KA-Seg, an innovative attention-based framework. KA-Seg refines sparse voxel features to further enhance robustness. Its core feature is an attention mechanism with super voxel partitioning and key point subsampling, which greatly improves the model’s ability to identify complex spatial patterns and focus on important voxel regions. Inspired by Transformer architecture, KA-Seg utilizes learnable key point sampling for global feature querying, expanding the model’s spatial understanding. This method augments spatial information processing across the point cloud and achieves a 1.3% higher mean intersection over union (mIoU) on the test set compared to the baseline model. Our code is publicly available at https://github.com/cvkdnk/kaseg .