SPGCN: Ground filtering method based on superpoint graph convolution neural network for vehicle LiDAR

Light detection and ranging (LiDAR) point clouds are sparse, unstructured, and disordered; hence, traditional convolutional neural networks are unsuitable for direct application in point-cloud data processing. Graph convolution neural networks (GCNs) can be used to process point-cloud data having the aforementioned characteristics; however, they are inefficient when the adjacent relationship of the point cloud is uncertain and adjacency-matrix elements are abundant. To solve these problems, a ground filtering method based on a superpoint graph convolution neural network (SPGCN) for vehicle LiDAR is proposed. With this method, the point-cloud data are transformed into a superpoint graph, and a GCN is employed for ground filtering. First, the points are projected on the range view and divided into grids, and the points in each grid are pooled in an average to form superpoints. Subsequently, a two-layer GCN is used to extract the features of superpoints, which are then assigned to each point. Finally, two fully connected layers are used to classify the ground and off-ground points. The SPGCN is trained and tested using the SemanticKITTI dataset, and the results confirm its effectiveness.