LiDAR Filtering Algorithms: Comparison between Conventional and Deep Learning Approaches

LiDAR enables high-precision 3D point cloud acquisition. Ground filtering is critical for DEM/DSM generation. Addressing challenges posed by multi-scenario point clouds and complex terrains, this study compares three traditional algorithms (PTD, IPTD, MPTPD) with deep learning models (PointNet++, RandLA-Net, KPConv). Key findings from OpenGF dataset experiments:Accuracy: IPTD (96.39%) trailed KPConv (97.79%) by 1.4% in mountainous areas; PTD (93.37%) lagged RandLA-Net (94.96%) by 1.59% in plains. Fault Tolerance: PTD demonstrated strongest noise resistance (0.24% accuracy drop) among traditional algorithms; KPConv was most stable (0.56% drop) in deep learning models. Efficiency: PTD outperformed MPTPD/IPTD by processing mountainous low-density data several times faster, achieving optimal overall performance. While deep learning models offer marginally higher accuracy, traditional algorithms (notably PTD) provide superior efficiency/stability for small-and medium-scale engineering applications, guiding DEM production and algorithm selection.