Sport Field Calibration with NeRF-guided Camera Optimization from a Single Image

Existing camera calibration methods struggle in dynamic sports environments due to transient scenes, poor lighting, and occlusions. Many state-of-the-art approaches depend on large-scale synthetic or multi-view datasets, which limit their real-world applicability. To overcome these challenges, we propose a novel monocular camera calibration framework for single-frame football field image, leveraging semantic attention mechanisms and neural radiance fields (NeRFs) to enhance camera pose estimation. Our approach extracts key structural features from the sports field using a DeepLabv3 + model to calculate initial object coordinates and camera calibration parameters. We refine spatial feature representation by integrating scene semantics into Pixel-NeRF via an attention mechanism for improved positional encoding. Finally, we globally optimize camera calibration parameters using semantic and geometric features to achieve more accurate 3D reconstruction and camera pose estimation. Our method achieves state-of-the-art results on the SoccerNetV3 and World Cup 2014 datasets, reducing the reprojection error by 0.5\% and reaching 92.1\% IoU.