Estimating Physical Camera Parameters for 3dav Video Coding

While video images are currently considered as two-dimensional (2D) entities, a recent trend is to extend video towards a 3D representation. In current video-coding algorithms like MPEG-4, global camera-motion is described as a set of abstract parameters, which have no direct relation to the physical camera movement. To bring physical meaning to these parameters, we present a new algorithm for estimating the extrinsic camera parameters and the varying focal length of a rotating camera from the MPEG-4 perspective motion-model parameters. Most previous camera calibration algorithms did not consider calibration from video data and give inaccurate results if rotation angles are small. Our algorithm applies a global non-linear optimization that is more robust on video sequences and achieves a very high estimation accuracy. The optimization is carried out using a four-stage algorithm, hereby reducing the total number of iterations in the gradientdescent search by a factor of 150.