In order to facilitate feature extraction video recordings were conducted under laboratory conditions, i.e. controlled environment with diffuse lighting and a unicolored blue background. The scene was illuminated frontally by six fluorescent lamps, each equipped with two tubes generating true natural daylight. Diffusion filters were mounted in front of the lamps for spreading the light beam and reducing shadows.
The signers wear dark clothes with long sleeves and perform from a standing position. Moreover each signer was instructed to move his hands from a resting position beside the hips to the signing location and after signing back to the same resting position. The hands are visible throughout the whole sequence, and their start and end positions are constant and identical which simplifies tracking.
For recording we used a camera which is commonly employed in machine vision tasks. This camera was connected via IEEE 1394 interface (also known as FireWire) with the computer, so that all videos could be recorded digitally without the need of any frame grabber. The main reason for choosing a machine vision camera instead of a common television camera was that we were able to program our own recording software. Our software allows to control the camera settings and ensures an almost full automatic capturing of the sign language corpus. Further post-processing work was thus reduced to a minimum.
All videos were recorded directly onto hard disk using an image resolution of 776x578 pixels at 30 fps. This high spatial resolution ensures reliable extraction of manual and facial features from the same input image. For quick random access to individual frames, each video clip was stored as a sequence of images. The following table summarizes the technical details about the camera, lens, and illumination. The figure on the right side illustrates the signer's position relative to the camera and lamps.
|
|