Abstract

Accurate camera calibration is a cornerstone of aerial imaging, essential for precise 3D reconstruction, mapping, and motion estimation. Traditional calibration methods often depend on predefined objects and periodic recalibration, which are impractical in dynamic aerial environments. This study investigates the potential of AI-based calibration methods, specifically GeoCalib, CTRL-C, and DeepCalib, to address these challenges. Using the ISPRS Vaihingen dataset, evaluate these methods against the conventional approach. The research focuses on intrinsic parameter estimation and its impact on 3D reconstruction accuracy. Our findings reveal that CTRL-C achieved the highest precision, with a mean reconstruction error of 1.59e10^-5, significantly outperforming GeoCalib (1.0549) and DeepCalib (0.2110). Additionally, DeepCalib demonstrated strong performance in minimizing Chamfer Distance (0.4220) and Hausdorff Distance (0.2502), while GeoCalib exhibited broader error distributions. These results underscore the superior capability of AI-based techniques in delivering accurate and reliable calibration for aerial imaging systems.

Statistics