DynPurge: Dynamic Point Removal via Spatiotemporal Distribution Range in Global-Scale LiDAR Maps

Abstract

Accurate dynamic object removal from LiDAR-based maps is critical for reliable localization and navigation in urban environments, where moving targets often leave persistent point cloud trajectories during sequential scan accumulation, thereby degrading map fidelity. This study proposes a temporal distribution analysis framework called DynPurge that addresses this challenge by exploiting inherent spatiotemporal disparities between static and dynamic objects through differential patterns in point cloud timestamp distributions. Unlike conventional geometric or visibility-based approaches, the methodology achieves robust static-dynamic classification without relying on prior geometric assumptions or computationally intensive ray-tracing operations. Evaluated on SemanticKITTI, MCD, Argoverse 2 dataset sequences, the framework demonstrates competitive classification accuracy while maintaining computational efficiency comparable to baseline methods, effectively balancing the latency-accuracy trade-off for dynamic point cloud removal.