Performance Evaluation of ICP Variants for Lidar Data Registration: A Comparative Study on KITTI Visual Benchmark Suite
Open Access
- Author:
- Spackman, Everett
- Graduate Program:
- Spatial Data Science
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 28, 2023
- Committee Members:
- Anthony Robinson, Program Head/Chair
S. Ilgin Guler, Thesis Advisor/Co-Advisor
Karen Schuckman, Committee Member - Keywords:
- ICP Registration
Iterative Closest Point
lidar
LIDAR
Point to point ICP
Point to plane ICP
Generalized ICP
KISS ICP
Fast GICP
GICP
Iterative Closest Point Registration
Autonomous Vehicles
AV
Lidar Mapping - Abstract:
- Light Detection and Ranging (Lidar) is a remote sensing technology that has been widely adopted by many industries for applications ranging from post-processing to real-time calculations. As lidar technology continues to advance, data processing methods must also keep pace. Iterative Closest Point (ICP) registration is a crucial step in processing data into a final product by estimating correspondences between individual point clouds and transforming them into a unified dataset. This thesis evaluates several ICP methods, including classic techniques like point-to-point, point-to-plane, and generalized ICP, as well as newer methods such as Keep It Small and Simple (KISS) ICP, Fast Generalized ICP, and Voxelized Generalized ICP. The goal is to identify the fastest and most accurate variants of ICP. The testing was conducted on sequences 0 through 11 of the KITTI Visual Benchmark Suite, a widely used dataset for evaluating point cloud registration accuracy, measured by Translation Error, Rotation Error, Absolute Translation Error, Absolute Rotation Error, Relative Pose Translation Error, and Relative Pose Rotation Error and processing speed using Average Frequency. The registration processing and evaluation was performed using modified versions of the KISS ICP, Fast GICP, and Kitti-Eval libraries with processing done locally on a workstation with high compute power, including an AMD Threadripper Pro, Nvidia RTX A6000, and 32 gigabytes of Random Access Memory (RAM). The results show that KISS ICP processed the registration most accurately, while Fast GICP was the quickest to process all sequences. Overall, KISS ICP performed the best, providing the highest accuracy at nearly the fastest speeds. Future studies should focus on improving the source code used for registration, expanding the datasets used for evaluation, and testing more variations of ICP.