A Filtering Method for LiDAR Point Cloud Based on Multi-Scale CNN with Attention Mechanism

Point cloud filtering is an important prerequisite for three-dimensional surface modeling with high precision based on LiDAR data. To cope the issues of low accuracy or excessive model complexity in traditional algorithms, this paper proposes a method point multi-scale convolutional neural network incorporated attention mechanism. Firstly, regular image patch centering each constructed elevation information clouds. As thus, problem transformed into classification problem. Then, considering ability convolution to extract features at different scales and potential mechanism capture key images, framework constructed, coordinate kernel channel spatial modules. After this, feature maps clouds can be acquired scales. For these maps, weights layer regions further tuned adaptively, which makes training more targeted, thereby improving performance eventually separating ground points non-ground preferably. Finally, proposed compared cloth simulation (CSF), deep (DNN), k-nearest neighbor (KNN), (DCNN) scale-irrelevant terrain-adaptive (SITA) standard ISPRS dataset filter Qinghai. The experimental results show that obtain lower errors, proves superiority filtering.