Memory-Efficient Mapping Based on Geometry-Aware Adaptive Gaussian Control for LiDAR Localization

Efficient map representations are crucial for LiDAR-based localization, as they directly influence localization accuracy and memory efficiency. Among existing map representations, probabilistic and Gaussian-based maps have attracted considerable attention owing to their compact representation and efficient localization capability. However, existing approaches have limited ability to adaptively capture local geometric complexity while maintaining compact map representations. To construct a compact and localizable map representation for LiDAR-based localization, this paper proposes a geometry-aware adaptive Gaussian mapping framework based on geometry-aware optimization and adaptive Gaussian control. The proposed framework constructs a compact LiDAR Gaussian Map by optimizing and adaptively controlling Gaussian primitives according to local geometric characteristics. Experimental results demonstrate superior map compression compared with existing Gaussian-based map representations while achieving localization performance comparable to or better than that of the conventional Normal Distributions Transform localization pipeline. The proposed framework reduces map storage requirements by up to 75.7% while achieving localization accuracy comparable to or better than conventional NDT, demonstrating that localization-relevant geometric constraints can be preserved with fewer Gaussian primitives.