Abstract: Point cloud semantic segmentation in coal mine roadways is a critical technology for scene understanding in underground coal mine environments. However, the coexistence of multi-scale objects and imbalanced class samples in mine point clouds lead to poor segmentation performance for small-scale targets. To address the issue, a direction-guided grouping method for semantic segmentation of point clouds in coal mine roadways is introduced. Leveraging the directional distribution of typical small-scale objects (e.g., pipes and cables), a Direction-Guided Grouping (DIG) strategy is proposed to optimize the shape of grouping envelopes in point cloud segmentation networks, thereby increasing the proportion of small-scale object points within groups and enhancing feature extraction. Based on this strategy, two directional guided grouping methods are proposed: one is Ellipsoid Query-based Direction-Guided Grouping (DIG-EQ), which preserves stronger local neighborhood relationships, and the other is Space-Filling Curve-based Direction-Guided Grouping (DIG-SFC), which offers higher computational efficiency. Both methods can be flexibly adapted to different network architectures and significantly improve the performance of small-scale object recognition.To mitigate class imbalance, a hybrid loss function is employed to improve sensitivity to underrepresented categories. A semantic segmentation dataset of coal mine roadway point clouds collected from different coal mine roadways is constructed for evaluation. The results show that the proposed method attains mean Intersection-over-Union (mIoU) of 61.84%, 69.49% and 76.63% on PointNet++, PointNeXt-L, and Point Transformer V3 backbones, respectively, representing improvements of 15.83%, 6.25% and 0.35% over baseline models. For small-scale categories, mIoU reach 22.28%, 28.61% and 47.30%, with relative gains of 29.99%, 42.34% and 5.33%. These results demonstrate the effectiveness of the proposed approach in coal mine roadway scenarios.