Research On Semantic Segmentation Algorithm Of ALS Point Cloud Based On Geometric Attention

With the development and popularization of the light detection and ranging(Li DAR)systems,it becomes more and more accessible to acquire point clouds in various scenes,particularly the airborne laser scanning(ALS)point cloud,which can be widely used in urban management,powerline corridor surveying and forest monitoring,etc.,due to its high scanning platform and wide coverage.The above applications all need to extract the category information of the point cloud,so the semantic segmentation of the ALS point cloud has thus been an important research topic in photogrammetry and remote sensing community.In particular,compared with the indoor point cloud and terrestrial laser scanning point cloud,the characteristics of ALS point cloud are distinctive in three aspects,(1)numerous geometric instances,e.g.tracts of roofs and facades;(2)extreme scale variations between different categories,such as the size of cars vs.the scale of roof,make it intractable to well classify all categories at different scales simultaneously.;(3)discrepancy distribution along the elevation,which can effectively help distinguish different categories,such as roof and ground points.Although many efficient ALS point cloud semantic segmentation algorithms have been proposed in the past few decades,how to further improve ALS point cloud classification accuracy by focusing on its unique properties has not yet been resolved.To this end,we propose an end-to-end geometry-attentional network consisting of geometry-aware convolution,dense hierarchical architecture and elevation-attention module to embed the three characteristics effectively.Among them,the geometryaware convolution takes the low dimensional geometric feature vectors as priori to induce the high dimensional local pattern representation learning,so as to obtain the geometric discriminative point cloud features.The dense hierarchical architecture upsamples the point feature matrix of each level of the down sampling encoder back to the original size,and integrates the receptive fields information of each level through nested skip connection paths to enhance the multi-scale expression ability of the network.The elevation-attention module transforms the height coordinates of the input point cloud into the weight coefficient of the elevation-attention,weighting the original feature channel and emphasizing the role of point-wise height information,thereby assisting the classification end to end.We conducted a large number of ablation experiments and accuracy analysis to verify the effectiveness of each module for the performance improvement of the ALS point cloud semantic segmentation.Our final geometry-attentional network significantly improves the semantic segmentation accuracy of ALS point cloud.Evaluated on the ISPRS Vaihingen 3D Semantic Labeling benchmark,our method achieves the state-of-the-art performance in terms of average F1 score and overall accuracy.Additionally,we performed detailed space and time complexity analysis of the proposed network and the baseline network.Under similar parameters and running time,without retraining,our model trained on the above Vaihingen 3D dataset can also achieve a better result on the dataset of 2019 IEEE GRSS Data Fusion Contest 3D point cloud classification challenge,achieving above 10%improvement both in terms of OA and average F1 score over the baseline,which verifies more powerful generalization ability of our model.