A deep learning solution for compressed semantic segmentation of LiDAR point cloud maps

Abstract

Navigating Unmanned Aerial Vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environments often relies on pre-built Light Detection and Ranging (LiDAR) maps. However, the large memory footprint and high computational cost of these point cloud maps pose significant challenges for resource-constrained UAVs. This paper proposes a deep learning solution using a lightweight, modified RandLA-Net architecture to efficiently compress and semantically segment these maps. Our results demonstrate a significant reduction in model size and memory usage while maintaining competitive segmentation accuracy, presenting a viable solution for real-time, on-board processing on embedded systems.