Abstract

This thesis addresses the critical challenge of achieving real-time and precise calibration for multi-camera systems, particularly crucial for applications demanding high precision. Departing from conventional methodologies characterized by limited adaptability and incapacity for on-the-fly recalibrations, this research introduces an innovative neural network-based approach designed to facilitate dynamic real-time calibration. This work incorporates the camera pose synthesis pipeline to simulate real-world camera parameters, introducing various levels of perturbation to the camera parameters to enhance model robustness. Also, a differentiable projection model is utilized to establish a direct correlation between 3D geometries and their 2D image projections, thereby facilitating concurrent optimization of intrinsic and extrinsic camera parameters. This solution entails a multi-headed deep neural network regression model and is tailored for multi-camera systems equipped with onboard processing capabilities, leveraging direct 2D projections of 3D fiducials. A comprehensive series of experiments are conducted to demonstrate the superior efficacy of the proposed method over traditional calibration techniques. This research contributes significantly to the advancement of real-time multi-camera system calibration, offering promising implications for diverse domains reliant on precise temporal synchronization and spatial accuracy.