Electronic Thesis and Dissertation Repository

Thesis Format

Integrated Article

Degree

Master of Science

Program

Computer Science

Supervisor

El-Sakka, Mahmoud R.

Abstract

This thesis introduces two novel methods for the extrinsic calibration of a thermal camera and a 3D LiDAR sensor, which are crucial for seamless data integration. The first method employs a distinctive calibration target, leveraging lines and plane equations correspondence in both modalities for a single pose, and incorporating more poses by matching the target's edges. It achieves reliable results, even with just one pose yielding 10.82% translation and 0.51-degree rotation errors. This outperforms alternative methods, which require eight pairs for similar results. The second method eliminates the need for a dedicated target. Instead, by collecting data during the sensor setup movement in environment and using a novel evolutionary algorithm optimizes a loss that measures alignment of humans in both modalities. This approach results in a 4.43% loss improvement compared to extrinsic parameters obtained by target-based methods. These methods save calibration time, reduce costs, and make sensor integration more accessible.

Summary for Lay Audience

This thesis introduces two calibration methods for seamlessly integrating a thermal camera and a 3D LiDAR dataset, focusing on aligning their coordinate systems via a rotation matrix and a translation vector.

The first method utilizes a distinctive calibration target visible in both sensors. For a single pair of thermal image-point cloud data, the algorithm establishes correspondences between the target's lines and plane equations in both modalities, determining extrinsic parameters. Further enhancement involves incorporating more pairs by matching the target's edges in both modalities. This method demonstrates reliability even with just one pair and exhibits notable performance with sparse LiDARs. In testing, it achieves 10.82% translation and 0.009 radian rotation errors with a single pose, surpassing methods requiring 8 data pairs. Beyond accuracy, this approach offers practical advantages. It notably reduces time expenditure by adopting a single-pose calibration strategy, which is particularly beneficial in scenarios like automobile sensor setups, where challenges in target positioning and thermal stability are prominent.

The second method introduces an extrinsic calibration approach that eliminates the need for a dedicated calibration target. Instead, it leverages data collected during sensor setup movements in environments with human presence, such as streets or farm fields. The algorithm optimizes extrinsic parameters based on a designed loss function measuring the alignment of humans in both modalities. To minimize this loss, a novel evolutionary algorithm is employed. This method exhibits a 4.43% improvement in loss compared to target-based calibration parameters in one dataset. Its efficacy extends to challenging real-world environments and stands out for not requiring an initial solution. Beyond accuracy improvements, this method presents a range of practical benefits, including cost reduction in creating thermal camera-visible calibration targets, time-saving in diverse pose collection, mitigating the tedium of repetitive calibration in scenarios with sensor drift, and enhanced accessibility by eliminating the need for a specialized target during calibration.

Download

Share

COinS