Outlier Detection in Crowdsourced LiDAR Data Using Deep Learning-Based Semantic Segmentation

Crowdsourced drone LiDAR data, collected from diverse sensors and environments, often suffer from inconsistent data quality.Outliers in such datasets can distort the characteristics of terrain and structures, leading to inaccurate decision-making.This study aims to validate the quality of crowdsourced LiDAR data by developing a precise outlier detection method based on semantic segmentation.Using the open-source Semantic Terrain Points Labeling Synthetic 3D (STPLS3D) dataset, noise-augmented training data were generated through simulations.Subsequently, a Kernel Point Convolution (KPConv) model was trained.The trained model was applied to real-world crowdsourced LiDAR data and compared with existing methods, including Statistical Outlier Removal (SOR) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN).Experimental results demonstrated that the KPConv model outperformed SOR and DBSCAN in terms of accuracy and reliability, effectively identifying inherent noise in the original data.These findings highlight the utility of deep learning-based outlier detection methods for ensuring the quality of crowdsourced LiDAR data.