POINT CLOUD PROCESSING ON 3D LIDAR DATA FOR UNMANNED SURFACE VEHICLES NAVIGATION IN MARINE ENVIRONMENTS

Abstract

Unmanned Surface Vehicles (USVs) require reliable autonomous navigation systems in dynamic maritime environments. LiDAR sensors, while effective for obstacle detection, are susceptible to noise and distortion from wave oscillations and water surface absorption, generating false positives and negatives in point cloud data. Most existing preprocessing-clustering evaluations are confined to terrestrial or simulated settings, with limited benchmarking under real maritime conditions marked by wave-induced motion, surface reflections, and sparse returns. This study addresses that gap through one of the first systematic evaluations of preprocessing-clustering combinations using real harbour LiDAR data collected under genuine environmental disturbance. A Velodyne HDL-32E LiDAR sensor was deployed at Klang Harbour, Malaysia, to collect real-world point cloud datasets. Three preprocessing techniques — Voxel Grid Downsampling, Statistical Outlier Removal (SOR), and Radius Outlier Removal (ROR) — were evaluated with two clustering algorithms chosen for their contrasting noise-handling strategies: Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Euclidean Cluster Extraction (ECE). Performance was assessed using precision and accuracy against manually annotated ground truth. DBSCAN combined with SOR achieved the highest performance, with peak accuracy of 57% (ε = 1.0 m) and peak precision of 62% (ε = 2.0 m), as SOR removes wave-induced outliers while preserving obstacle geometry, enabling DBSCAN's density criterion to isolate genuine clusters more reliably. ROR consistently degraded performance by over-removing genuine sparse returns, while ECE underperformed across all configurations due to its lack of native noise handling. The overall pipeline reached approximately 60% accuracy, a modest ceiling reflecting persistent wave motion, surface reflections, sparse long-range returns, and moving vessels that geometry-based clustering alone cannot resolve. This study contributes a methodological benchmark for preprocessing-clustering pipeline selection in maritime LiDAR processing, and practical guidance for selecting efficient, noise-robust configurations for real-time USV obstacle detection. Future work should incorporate model-based object detection and multi-sensor fusion to improve classification capability and system reliability.