Genetic Algorithm-Based Camera Calibration in Digital Twin

Cyber-physical system (CPS) supports the analysis, simulation, and automation based on the exchange of information between real and digital twin (DT) world. Among various sensing modalities, cameras are widely employed, and camera calibration plays a crucial role in achieving geometric consistency between cyber and physical domains. Genetic algorithm (GA) based camera calibration has been introduced with its global search capability, compared to gradient descent based approach. However, several challenges such as the need for predefined lower and upper bounds of each parameter, limited fitness function and slow convergence lead to a lack of open-source implementations. To overcome these bottlenecks, this paper proposes a novel GA based camera calibration framework that leverages DT. Unlike conventional approaches relying solely on reprojection error, the proposed fitness function better reflects the parameter errors by leveraging the geometric advantages of DT. Utilizing GA enables the use of the proposed non-differentiable fitness function. Besides, a hyperparameter scheduling strategy is integrated to accelerate the convergence. Experiments with five distortion parameters demonstrate that the proposed method consistently converges to near-optimal solution, even when keypoints are nearly planar and widely distributed. The proposed fitness functions reduce the camera position error by approximately 2.1 meters, and hyperparameter scheduling reduces the convergence time by half compared to fixed-parameter GA. Furthermore, a comparative analysis with and without tangential distortion validates the robustness under increased non-linear complexity. The source code is available at https://github.com/Kwan-Ho-Kim/GADT.