Raw point clouds are extensively used in 3D object detection due to their detailed spatial positioning information, which surpasses that of voxelized point clouds. However, point-based 3D object detection networks often prioritize local information over the scene’s overall context when selecting sampling locations, mainly focusing on the cloud’s spatial distribution. This is particularly problematic for small objects like cyclists and pedestrians, where point cloud data are naturally sparse, negatively impacting the accuracy of regression and classification. To tackle these challenges, we introduce a novel single-stage 3D detection algorithm named AF-SSD. Our approach begins with a sampling module driven by a self-attention mechanism that merges spatial and feature distances, facilitating a comprehensive understanding of the point cloud’s global information while retaining more points of foreground objects. Additionally, we apply fuzzy classification theory in the bounding box regression of the detection head to refine the Non-Maximum Suppression (NMS), further mitigating missed detections due to the spatial clustering of multiple small objects. Our experiments on the KITTI dataset demonstrate that our algorithm achieves competitive regression precision and performance for pedestrian detection, in line with current state-of-the-art methods.