Georgia Institute of Technology
An Evaluation of Tightly-coupled Lidar-inertial Odometry Paradigms and Their Impact on Downstream Volumetric Reconstruction
Abstract
dc:description.abstractAccurate 3D representation in complex environments increasingly relies on a two-stage pipeline: tightly-coupled LiDAR-Inertial Odometry (LIO) followed by surface reconstruction. However, systematic evaluations reveal significant literature gaps. First, Direct LiDAR-Inertial Odometry (DLIO), a recent, highly performant method, is frequently omitted from standard benchmarks. Second, existing evaluations default to simple, road-like trajectories where tracking errors are minimal, failing to expose algorithmic limitations under complex motion. Third, benchmarks focus almost exclusively on trajectory accuracy, neglecting system-level resource consumption and processing constraints. Concurrently, the subsequent volumetric surface reconstruction stage is rarely challenged with tracking noise and pose errors propagated from the LIO backend, remaining limited to small-scale static setups evaluated under ideal conditions. To address these limitations, This thesis benchmarks three state-of-the-art LIO algorithms: DLIO, FAST-LIO2, and LIO-SAM, across demanding urban and unstructured environments, assessing tracking precision alongside CPU usage, memory, and latency. We then execute a two-phase evaluation of the VDB-GPDF reconstruction framework, establishing an intrinsic baseline against 3D ground-truth meshes before challenging it with actual, noisy LIO outputs. Our analysis reveals distinct trade-offs: DLIO ensures tracking stability in vegetation despite high memory usage, FAST-LIO2 excels in urban settings with low latency but high CPU saturation, and LIO-SAM violates real-time constraints. Crucially, tracking errors remain minimal; downstream mapping limitations are driven by external factors like occlusions and dynamic objects rather than localization failures.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, MS
- Department thesis:degree_department
- Computer Science
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Revel, Antoine
- Advisor dc:contributor.advisor
-
- Pradalier, Cédric
- Committee members dc:contributor.committeemember
-
- Tsai, James
- Yang, Zhongyu
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82238