SalsaNext-based Point Cloud Segmentation Method for Estimating Drivable Area in Off-Road Environments

Abstract

Given that the majority of South Korea’s terrain is mountainous, the realization of autonomous industrial machines in off-road environments is particularly important domestically. In off-road environments, localizing the drivable area is more advantageous for generating driving paths than recognizing surrounding objects. In this study, we propose a method for estimating the drivable area among point cloud data acquired in off-road environments. Our method semantically segments the modified RELLIS-3D point cloud data into undrivable areas, drivable areas, and dynamic objects for drivable area estimation. Test results for SalsaNext trained on modified RELLIS-3D showed precision of 90.6%, recall of 91.0%, F1 score of 90.8%, and mIoU of 0.627, respectively.