Julian Gaal
Feature-Based Mobile Mapping with 3D Lidars on Embedded GPUs

Abstract
This thesis presents two solutions to the Simultaneous Localization and Mapping (SLAM) problem that share a common core. Featsense uses lidar point cloud features for odometry estimation, while Warpsense presents a GPU-accelerated Point-to-TSDF scan matching algorithm that performs localization in a high resolution, continuous Truncated Signed Distance Field (TSDF) representation of the environment. Both methods share the same mapping backend, a highly GPU-optimized TSDF generation module that allows the generation of efficient triangle meshes in post-processing.
WhileWarpsense is capable of mapping indoor environments at high resolution on an embedded system with an integrated GPU-accelerator, Featsense is stable, accurate and fast in mapping truly large environments in various indoor, outdoor and recording scenarios on both low power and desktop computing devices. Both approaches are developed with real-time requirements for the processing of dense, high-resolution lidar data, are extensively evaluated and present exciting opportunities for future improvements.
The main contribution of this thesis is an TSDF map generation speedup by a factor of 100 when compared to its reference implementation and a Point-to-TSDF registration speedup by a factor of 8.6 at the same power budget, made possible on the NVIDIA AGX Xavier embedded system. Additionally, Featsense presents a highly optimized lidar odometry and mapping algorithm with a sparse feature extractor for the high-resolution Ouster OS1-128 lidar sensor and a GPU-accelerated TSDF mapping backend. These improvements enable the recording and continuous mobile mapping of large-scale environments with high-resolution lidar sensors.