Abstract: Light detection and ranging (LiDAR)-acquired point-cloud data are extensive and characterized by their non-uniform density, with points typically denser near the sensor and sparser at greater distances. Efficient sampling of point clouds is crucial for reducing computational complexity while preserving critical environmental information. However, classical sampling methods, such as farthest point sampling (FPS) and random sampling, fail to adequately address the challenges posed by the imbalanced distribution of foreground and background points. Oversampling of background points or insufficient coverage of foreground regions can result in the loss of essential target information, particularly for small or distant objects, thus ultimately degrading the performance of three-dimensional (3D) object-detection networks. Although FPS has been widely adopted in many point-based detection frameworks, its sequential nature limits its efficiency and effectiveness in complex scenarios. Hence, we propose a novel graph feature augmentation sampling (GFAS) method, which leverages graph convolutional networks and supervised learning to enhance sampling efficiency and detection performance. The proposed method introduces a graph-feature-generation module that aggregates local and global features of point clouds using multilayer graph convolutions, thus enabling the extraction of rich geometric and spatial information. Additionally, it incorporates a parallel sampling mechanism that selects foreground points based on their feature scores, thereby significantly improving sampling efficiency. By utilizing foreground–background classification labels as supervision signals, GFAS ensures a higher proportion of foreground points in the sampling process, which is particularly beneficial for detecting objects. Extensive experiments are conducted on two large-scale autonomous driving datasets, i.e., KITTI and nuScenes, to validate the effectiveness of GFAS. On the KITTI dataset, GFAS achieves significant improvements in terms of average precision for car detection, with gains of 6.2%, 6.89%, and 8.58% under easy, moderate, and hard levels, respectively. Similar improvements are observed for pedestrian and cyclist detection, thus demonstrating the robustness of the proposed method across different object categories. On the nuScenes dataset, the proposed method improves car- and pedestrian-detection performance significantly, with higher precision levels by 4.2% and 8.3%, respectively, compared with the baseline model. These results highlight the strong generalizability of GFAS in large-scale and complex driving scenarios. Ablation studies reveal that GFAS significantly increases the proportion of foreground points in the sampling process, with the ratio approaching 99% in the final layers. Visualization results show that GFAS effectively concentrates sampling points on foreground objects, thus avoiding the uniform distribution issue of classical FPS methods. Additional experiments on other 3D object-detection models, such as 3D single stage object detector (3DSSD) and PointVoxel-RCNN (PV-RCNN), further validate the flexibility and scalability of the proposed method. In conclusion, this paper proposes an efficient and parallel point-cloud-sampling method. By integrating graph-feature extraction and supervised learning, GFAS not only improves sampling efficiency but also enhances detection performance, particularly for challenging scenarios. The proposed method can be easily integrated into existing point-cloud-based detection frameworks. Its ability to retain a high proportion of foreground points while maintaining computational efficiency highlights its practicality.