DNGF-LS: A Dynamic Non-Ground Points Filtering LiDAR SLAM Based on Sparse 4DCNN

Achieving precise localization and high-quality static map construction in unfamiliar environments is fundamental for mobile robots. However, in environments like urban roads, dynamic objects such as pedestrians and vehicles can cause LiDAR SLAM odometry drift, leaving ghostings in the constructed map, which can lead to the failure of subsequent tasks such as path planning based on prior maps. Therefore, performing dynamic points filtering on point clouds is crucial for mobile robots, this paper introduces a novel Dynamic Non-Ground points Filtering LiDAR SLAM(DNGF-LS). Firstly, employing a Coarse-to-Fine Concentric Zone Model(C-F-CZM) to represent point clouds and perform ground segmentation addresses the sparsity issue caused by distant points in ground plane fitting, yielding accurate non-ground point clouds. Subsequently, converting the current frame's previous ten frames of point clouds to 4D point clouds with the same viewpoint, utilizing sparse voxel grids to represent 4D point clouds, and inputting them into a sparse 4D Convolutional Neural Network(CNN) for dynamic points filtering. Finally, retaining static points for pose estimation and mapping using LiDAR Odometry And Mapping(LOAM). Experimental results on highly dynamic datasets demonstrate that DNGF-LS can construct high-quality static maps and achieve superior localization accuracy compared to LIO-SEGMOT and SUMA++.