Autonomous Semantic Mapping Refinement via Dynamic Graph Pruning for High-Throughput LiDAR Localization in Urban Canyons

**Abstract:** This paper presents a novel approach to autonomous semantic mapping refinement specifically tailored for high-throughput LiDAR localization in challenging urban canyon environments. Existing simultaneous localization and mapping (SLAM) systems often struggle with the computational burden of dense semantic maps, particularly in areas with repetitive structures and limited visibility. Our approach, Dynamic Graph Pruning for Semantic Localization (DGPSL), introduces a dynamic pruning algorithm that selectively removes redundant and low-information semantic nodes from the map graph in real-time based on LiDAR scan quality and localization confidence. This reduces computation while maintaining high localization accuracy and map robustness. The system leverages established LiDAR-based SLAM methodologies combined with recent advances in semantic segmentation and graph optimization to achieve a 10x reduction in graph size with minimal degradation in localization performance, enabling high-throughput operations suitable for autonomous vehicle deployment.