Lidar-based semantic segmentation is critical for autonomous driving, but existing methods struggle with real-time performance in unstructured environments like open-pit mining areas. This paper proposes a real-time semantic segmentation network leveraging an attention mechanism to improve accuracy and speed. Key innovations include a global contextual information extraction module, bilinear upsampling layers to reduce channel redundancy, and coordinate attention-based upsampling to alleviate structural redundancy. This paper tested the proposed method on an unstructured road open-pit mining dataset, achieving an average IoU of 85.6% and an inference speed of 5.90ms, which is 4.3% and 14.6% higher than the baseline (SalsaNext) in terms of accuracy and speed, respectively.