ThermalVoxCalib: automatic extrinsic calibration between a non-repetitive scanning 3D LiDAR and a thermal camera

A low-cost non-repetitive scanning 3D LiDAR (Light Detection and Ranging) is receiving more attention (the Livox LiDAR). The low cost promotes the large-scale mass-production of various robot systems. Fusing this LiDAR with a thermal camera will improve the perception ability in low illumination...

Full description

Saved in:
Bibliographic Details
Main Author: Liu, Yiyao
Other Authors: Wang Dan Wei
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2022
Subjects:
Online Access:https://hdl.handle.net/10356/159538
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:A low-cost non-repetitive scanning 3D LiDAR (Light Detection and Ranging) is receiving more attention (the Livox LiDAR). The low cost promotes the large-scale mass-production of various robot systems. Fusing this LiDAR with a thermal camera will improve the perception ability in low illumination environments. However, the requirement of large-scale calibration tasks and the non-uniform scanning pattern poses new challenges. The main difficulty lies in the automatic detection of the calibration target: i) Most target-based calibration methods rely on human intervention to detect the target from the point cloud, but it is inefficient for mass production. ii) Some recently proposed methods rely on intensity difference to detect the target, but it is not general for different targets. To solve the problems, ThermalVoxCalib is proposed. The novelties are: 1) The first research to achieve the calibration between a non-repetitive scanning 3D LiDAR and a thermal camera. 2) Propose a robust method to automatically detect a planar target from the point cloud. It can be generalized to other ranging sensors and other planar targets. 3) We propose to estimate the extrinsic parameter by minimizing 2D re-projection error, and compare the accuracy versus minimizing 3D matching error. Quantitative and qualitative experiments are conducted in both simulation and real environment. Experiments demonstrate the proposed target detection algorithm is robust and accurate. The rotation and translation error can reach 0.172° and 0.01m. The 2D re-projection error can reach 0.58pixels.