Enabling Multi-LiDAR Sensing in GNSS-Denied Environments: SLAM Dataset, Benchmark, and UAV Tracking with LiDAR-as-a-camera

Abstract

The rise of Light Detection and Ranging (LiDAR) sensors has profoundly impacted industries ranging from automotive to urban planning. As these sensors become increasingly affordable and compact, their applications are diversifying, driving precision, and innovation. This thesis delves into LiDAR's advancements in autonomous robotic systems, with a focus on its role in simultaneous localization and mapping (SLAM) methodologies and LiDAR as a camera-based tracking for Unmanned Aerial Vehicles (UAV). Our contributions span two primary domains: the Multi-Modal LiDAR SLAM Benchmark, and the LiDAR-as-a-camera UAV Tracking. In the former, we have expanded our previous multi-modal LiDAR dataset by adding more data sequences from various scenarios. In contrast to the previous dataset, we employ different ground truth-generating approaches. We propose a new multi-modal multi-lidar SLAM-assisted and ICP-based sensor fusion method for generating ground truth maps. Additionally, we also supplement our data with new open road sequences with GNSS-RTK. This enriched dataset, supported by high-resolution LiDAR, provides detailed insights through an evaluation of ten configurations, pairing diverse LiDAR sensors with state-of-the-art SLAM algorithms. In the latter contribution, we leverage a custom YOLOv5 model trained on panoramic low-resolution images from LiDAR reflectivity (LiDAR-as-a-camera) to detect UAVs, demonstrating the superiority of this approach over point cloud or image-only methods. Additionally, we evaluated the real-time performance of our approach on the Nvidia Jetson Nano, a popular mobile computing platform. Overall, our research underscores the transformative potential of integrating advanced LiDAR sensors with autonomous robotics. By bridging the gaps between different technological approaches, we pave the way for more versatile and efficient applications in the future.