Camera model calibration establishes an accurate mapping between the 3D physical world and 2D images, enhancing the precision and reliability of vision measurements. Conventional calibration methods typically require multiple images to estimate initial values for nonlinear optimization, otherwise resulting in convergence to local optimal solutions. In addition, it also suffers from parameter coupling issues, limiting measurement accuracy and efficiency. To address these challenges, this study introduces a novel calibration method based on a physics-informed neural network (PINN). The proposed method reformulates the camera model as a set of constraint equations integrated into the neural network loss function. Additionally, a multilevel neural network architecture is designed to model the hierarchical transformations across coordinate systems in the camera’s imaging process. This design enables the neural network to capture and preserve physical relationships, ensuring accurate propagation of spatial information through the imaging pipeline. The method reduces dependency on initial values in nonlinear optimization and enhances calibration accuracy, even under physics-informed few-shot learning scenarios, while maintaining physical interpretability. A custom vision experiment system was developed to evaluate the method. Experimental results demonstrate that the proposed approach improves calibration accuracy by 29.82% compared to traditional methods and by 30.89% compared to the BPNN-based approach. These results confirm the effectiveness and precision of the calibration strategy.
原文链接
分享
引用
