LIDAR SEMANTIC SEGMENTATION WITH A MULTI-RESERVOIR ECHO STATE NETWORK FOR OFF-ROAD TERRAIN PERCEPTION

Autonomous vehicle perception has been widely explored using camera images but is limited with respect to LiDAR point cloud processing. Furthermore, focus is primarily on well-regulated environments, obviating a need for an algorithm that can contextualize dynamic and complex conditions through 3D point cloud representation. In this report, an Echo State Network for LiDAR signal processing is introduced and evaluated for its ability to perform semantic segmentation on unregulated terrains, using the RELLIS-3D open-source dataset. The L-ESN contains 16 parallel reservoirs with point cloud processing time of 1.9 seconds and 83.1% classification rate of 4 classes defining terrain trafficability, with no prior feature extraction or normalization, and a training time of 31 minutes. A 2D cost map is generated from the segmented point cloud for integration as a perception node plug-in to system-level navigation architectures.

Citation: S. Gardner, M. R. Haider, P. Fiorini, S. Misko, J. Smereka, P. Jayakumar, D. Gorsich, L. Moradi, and V. Vantsevich, “Lidar Semantic Segmentation with a Multi-reservoir Echo State Network for Off-road Terrain Perception”, In Proceedings of the Ground Vehicle Systems Engineering and Technology Symposium (GVSETS), NDIA, Novi, MI, Aug. 16-18, 2022.