英语论文网

ector. This would allow the cameras to sense visible light and perhaps eliminate the need for a separate illuminant. Later, we describe applications which benefit from visible-light based sensing. While for our initial implementation we have chosen to mount the display vertically such that the user may stand, it is also possible to mount the display surface horizontally to make a table. In this case a 'short throw' projector such as the NEC WT600 may be desirable. Finally, a microphone is rigidly attached to the display surface to enable the simple detection of 'knocking' on the display. Except for the microphone, there are no wires attached, making TouchLight more robust for public installation.

图像处理-IMAGE PROCESSING:

Introduction:

The goal of TouchLight image processing is to compute an image of the objects touching the surface of the display, such as the user's hand. Due to the transparency of the display, each camera view shows the objects on the display and objects beyond the surface of the display, including the background and the rest of the user. With two cameras, the system can determine if a given object is on the display surface or above it.

The touch image is produced by directly combining the output of the two video cameras. Depth information may be computed by relating binocular disparity, the change in image position an object undergoes from one view to another view, to the depth of the object in world coordinates. In computer vision there is a long history of exploiting binocular disparity to compute the depth of every point in a scene. Such depth from stereo algorithms is typically computationally intensive, difficult to make robust, and constrain the physical arrangement of the cameras. Often such general stereo algorithms are applied in scenarios that in the end do not require general depth maps. Here we are interested in the related but easier problem of determining what is located on a particular plane in three dimensions (the display surface) rather than the depth of everything in the scene.The algorithm detailed here runs in real time (30Hz) on a Pentium 4, operating on 640x480 images.

Image Rectification:

The TouchLight image processing algorithm proceeds by transforming the image from the left camera left I and the image from the right camera right I such that in the transformed images points ) , ( y x Ileft and ) ,( y x Iright refer to the same physical point on the display surface. Secondly, this transform is such that the point ) , ( y x may be trivially mapped to real world dimensions (i.e., inches) on the display surface. For both criteria, it suffices to find the homography from each camera to the display surface, which we obtain during a manual calibration phase. In the case of using wide angle lenses to make a compact setup, it is important to remove the effects of lens distortion imparted by wide angle lenses. Given the lens distortion parameters, we undistort the input image by bilinear interpolation. Sample images are shown in Figure 3b. During a manual calibration phase, the 4 corners of the display are manually located in each view. This specifies a projective transform bringing pixels in the lens distortion corrected image to display surface coordinates. Together with the lens distortion correction, the projective transform completes the homography from camera view to display coordinates. Sample resulting image