Research On Laser Point Cloud Classification For Attention Mechanism

With the development and popularization of lidar sensors,the cost of acquiring 3D lidar point cloud data has been greatly reduced,which has greatly promoted the development of lidar point cloud data in academic research and industry applications.As one of the important applications of lidar point cloud data,point cloud classification is widely used in power line inspection,3D reconstruction,forestry detection and so on.But the distribution of 3D laser point cloud data is scattered,and there is no strict adjacency relationship between point pairs.How to organize point clouds and form their spatial relationships are foundational for point cloud classification.In addition,compared with remote sensing images,density of point cloud often presents uneven distribution,which seriously affects the accuracy of point cloud classification.Considering the characteristics of 3D laser point cloud data,this paper conducts research on point cloud local feature extraction and network structure,and proposes the following two solutions to improve the accuracy of point cloud classification:(1)Local spatial position attention and multi-scale feature net for ALS point cloud classification(AMMSF-Net)was proposed.The uneven spatial distribution and scale variations between different categories bring challenges to the fine classification of point cloud data.In this section,an attention mechanism and multi-scale feature fusion network(AMMSF-Net)for ALS point cloud classification was proposed.In the network,a local spatial position attention layer was used to learn local contextual features;and an attention skip connection was added to dynamic fusion the corresponding features among the encoder and decoder,which can retain detail features an d contextual information.The multi-scale feature in the decoder fusion module obtains the final semantic probability map by concatenating the features at different scales into MLP(Multilayer Perceptron)and CML(Conditional Markov Layer),which achieves the correlation of the feature maps between different scales and different levels,and enhances the expression of targets at different scales.Experimental results in Vaihingen3D and CSPC show that AMMSF-Net can distinguish ground objects in point cloud effectively compared with other methods.(2)A cross attention and pyramid decoding feature adaptive fusion for laser point cloud classification(CAPDAF-Net)was proposed.Considering the problem that only taking the local neighborhood geometric information as the classification feature can not effectively improve the classification accuracy of point cloud.CAPDAF-Net enhances local neighborhood features through cross-attention,mines the context information of local neighborhoods,and fuses multiscale feature of decoder pyramid structure through the adaptive.Results in Toronto-3D and CSPC showed that CAPDAF-Net can help improve the accuracy of point cloud classification.