Deep Learning-Based Semantic Segmentation of LiDAR Point Clouds for Autonomous Driving in Unstructured Environments

Autonomous driving in unstructured environments is a challenging task for robotics due to the presence of natural obstacles such as rocks, trees, water, and vegetation, which can hinder navigation. In these scenarios, Unmanned Ground Vehicles rely on robust and reliable perception systems, primarily using cameras and LiDAR sensors for situational awareness. This paper presents an experimental and quantitative analysis of three deep learning-based perception modules, exploring their use for semantic segmentation of raw LiDAR point clouds. We propose a late-fusion approach that incorporates LiDAR intensity, proving its effectiveness in reducing class confusion and improving segmentation performance. All models were trained and tested using the GOOSE dataset, which provides labeled data from unstructured environments. In addition, we analyze the impact of point cloud density and noise on semantic segmentation performance.