A Novel Feature Extraction Framework with Direction-Aware KNN for ALS Point Cloud Semantic Segmentation

In the field of remote sensing, 3D point clouds provide precise environmental perception information. However, their non-uniformity and irregularity present significant processing challenges. Most current point cloud processing algorithms use the K-Nearest Neighbor (KNN) method to select neighboring points before feature extraction. However, the neighboring points selected by KNN are fixed, meaning the receptive field is a regular spherical shape. This rigidity prevents flexible adaptation to variations in object shapes and can lead to information confusion during subsequent feature propagation. To address this issue, we propose a novel Direction-Aware KNN (DAKNN) algorithm, which manipulates the shape of the receptive field during the neighbor selection process to adaptively focus on the directions where object features are most prominent. Based on the DAKNN algorithm, we designed a feature extraction framework that uses a unified scaling factor to adjust the shape of the receptive field. This approach introduces more diverse directional focus features with fewer points, enhancing the discriminability of features for different objects. It is worth noting that the proposed direction-aware feature extraction framework is general and can be integrated into most existing encoder-decoder network structures. Experimental results on the STPLS3D and S3DIS dataset demonstrate that adopting the proposed framework significantly improves the performance of existing methods.