A camera calibration method based on scene and human semantics

Camera calibration is an important preliminary for many applications such as 3D reconstruction, motion capture and human-computer interactions. Although great achievements have been made in the past several years, it still remains two main problems. The one is that conventional calibration methods always depend on specific calibration tools (e.g., checkerboard). The other is that the calibration procedure is complex when these methods are used in multi-view large scenarios. In this paper, we propose a camera calibration method which can be applied to large-scale sports scenarios without a special calibration object. To tackle these challenges, our key-idea is to use scene and human semantics detected from RGB images to calibrate the intrinsic and extrinsic parameters of multi-view cameras. We first detect the scene and human keypoints, then the initial camera parameters are calculated with the static scene keypoints and a predefined sports field model. Finally, we combine the scene and human semantics to refine the camera intrinsic and extrinsic parameters. Experimental results demonstrate that our method successfully obtains accurate camera intrinsic and extrinsic parameters for large-scale sports scenarios.